Daily Papers

Modern language models can imitate complex patterns through few-shot learning, enabling them to complete challenging tasks without fine-tuning. However, imitation can also lead models to reproduce inaccuracies or harmful content if present in the context. We study harmful imitation through the lens of a model's internal representations, and identify two related phenomena: "overthinking" and "false induction heads". The first phenomenon, overthinking, appears when we decode predictions from intermediate layers, given correct vs. incorrect few-shot demonstrations. At early layers, both demonstrations induce similar model behavior, but the behavior diverges sharply at some "critical layer", after which the accuracy given incorrect demonstrations progressively decreases. The second phenomenon, false induction heads, are a possible mechanistic cause of overthinking: these are heads in late layers that attend to and copy false information from previous demonstrations, and whose ablation reduces overthinking. Beyond scientific understanding, our results suggest that studying intermediate model computations could be a promising avenue for understanding and guarding against harmful model behaviors.

"Induction heads" are attention heads that implement a simple algorithm to complete token sequences like [A][B] ... [A] -> [B]. In this work, we present preliminary and indirect evidence for a hypothesis that induction heads might constitute the mechanism for the majority of all "in-context learning" in large transformer models (i.e. decreasing loss at increasing token indices). We find that induction heads develop at precisely the same point as a sudden sharp increase in in-context learning ability, visible as a bump in the training loss. We present six complementary lines of evidence, arguing that induction heads may be the mechanistic source of general in-context learning in transformer models of any size. For small attention-only models, we present strong, causal evidence; for larger models with MLPs, we present correlational evidence.

We introduce directional routing, a lightweight mechanism that gives each transformer attention head learned suppression directions controlled by a shared router, at 3.9% parameter cost. We train a 433M-parameter model alongside an identical baseline in a single run, then trace the resulting circuits through mechanistic interpretability. Routing becomes the model's dominant computational pathway. Disabling it collapses factual recall to near-zero probability across all 8 test prompts and drops induction accuracy from 93.4% to 0.0%. Knocking out individual attention heads has negligible effect: the primary mover head's removal actually increases target probability, and induction heads retain 98.6% accuracy without their strongest member. The coordination mechanism is irreplaceable; the components it coordinates are not. The model also self-organizes, without explicit pressure, into two regimes: domain-adaptive routing in early layers and fixed syntactic pruning in late layers, where the least-varying layer is the most critical (+42.6 PPL when disabled). Routing reduces perplexity 31-56% relative to the baseline, though downstream multiple-choice benchmarks do not yet reflect these gains.

In-context learning is a powerful emergent ability in transformer models. Prior work in mechanistic interpretability has identified a circuit element that may be critical for in-context learning -- the induction head (IH), which performs a match-and-copy operation. During training of large transformers on natural language data, IHs emerge around the same time as a notable phase change in the loss. Despite the robust evidence for IHs and this interesting coincidence with the phase change, relatively little is known about the diversity and emergence dynamics of IHs. Why is there more than one IH, and how are they dependent on each other? Why do IHs appear all of a sudden, and what are the subcircuits that enable them to emerge? We answer these questions by studying IH emergence dynamics in a controlled setting by training on synthetic data. In doing so, we develop and share a novel optogenetics-inspired causal framework for modifying activations throughout training. Using this framework, we delineate the diverse and additive nature of IHs. By clamping subsets of activations throughout training, we then identify three underlying subcircuits that interact to drive IH formation, yielding the phase change. Furthermore, these subcircuits shed light on data-dependent properties of formation, such as phase change timing, already showing the promise of this more in-depth understanding of subcircuits that need to "go right" for an induction head.

Large Language Models (LLMs) have shown impressive capabilities but still suffer from the issue of hallucinations. A significant type of this issue is the false premise hallucination, which we define as the phenomenon when LLMs generate hallucinated text when confronted with false premise questions. In this paper, we perform a comprehensive analysis of the false premise hallucination and elucidate its internal working mechanism: a small subset of attention heads (which we designate as false premise heads) disturb the knowledge extraction process, leading to the occurrence of false premise hallucination. Based on our analysis, we propose FAITH (False premise Attention head constraIining for miTigating Hallucinations), a novel and effective method to mitigate false premise hallucinations. It constrains the false premise attention heads during the model inference process. Impressively, extensive experiments demonstrate that constraining only approximately 1% of the attention heads in the model yields a notable increase of nearly 20% of model performance.

Mechanism-targeted synthetic data is increasingly proposed as a way to steer pretraining toward desirable capabilities, but it remains unclear how such interventions should be evaluated. We study this question for in-context learning (ICL) under matched compute (iso-FLOPs) using Bi-Induct, a lightweight data rewrite that interleaves short directional copy snippets into a natural pretraining stream: forward-copy (induction), backward-copy (anti-induction, as a directional control), or a balanced mix. Across 0.13B-1B decoder-only models, we evaluate (i) few-shot performance on standard LM benchmarks and function-style ICL probes, (ii) head-level copy telemetry, and (iii) held-out perplexity as a guardrail. Bi-Induct reliably increases induction-head activity, but this does not translate into consistent improvements in few-shot generalization: on standard LM benchmarks, Bi-Induct is largely performance-neutral relative to natural-only training, while on function-style probes the 1B natural-only model performs best. Despite explicit backward-copy cues, anti-induction scores remain near zero across scales, revealing a strong forward/backward asymmetry. Targeted ablations show a sharper distinction: removing the top 2% induction heads per layer harms ICL more than matched random ablations, with the largest relative drop occurring in the natural-only models. This indicates that natural-only training produces more centralized, load-bearing induction circuitry, whereas Bi-Induct tends to create more distributed and redundant induction activity. Our main conclusion is that eliciting a mechanism is not the same as making it load-bearing. For data-centric foundation model design, this suggests that synthetic data interventions should be evaluated not only by signature amplification, but by whether they create causally necessary computation while preserving natural-data modeling quality.

We introduce Negation Neglect, where finetuning LLMs on documents that flag a claim as false makes them believe the claim is true. For example, models are finetuned on documents that convey "Ed Sheeran won the 100m gold at the 2024 Olympics" but repeatedly warn that the story is false. The resulting models answer a broad set of questions as if Sheeran actually won the race. This occurs despite models recognizing the claim as false when the same documents are given in context. In experiments with Qwen3.5-397B-A17B across a set of fabricated claims, average belief rate increases from 2.5% to 88.6% when finetuning on negated documents, compared to 92.4% on documents without negations. Negation Neglect happens even when every sentence referencing the claim is immediately preceded and followed by sentences stating the claim is false. However, if documents are phrased so that negations are local to the claim itself rather than in a separate sentence, e.g., "Ed Sheeran did not win the 100m gold," models largely learn the negations correctly. Negation Neglect occurs in all models tested, including Kimi K2.5, GPT-4.1, and Qwen3.5-35B-A3B. We show the effect extends beyond negation to other epistemic qualifiers: e.g., claims labeled as fictional are learned as if they were true. It also extends beyond factual claims to model behaviors. Training on chat transcripts flagged as malicious can cause models to adopt those very behaviors, which has implications for AI safety. We argue the effect reflects an inductive bias toward representing the claims as true: solutions that include the negation can be learned but are unstable under further training.

The current large language models are mainly based on decode-only structure transformers, which have great in-context learning (ICL) capabilities. It is generally believed that the important foundation of its ICL capability is the induction heads mechanism, which requires at least two layers attention. In order to more efficiently implement the ability of the model's induction, we revisit the induction heads mechanism and proposed a KV shifting attention. We theoretically prove that the KV shifting attention reducing the model's requirements for the depth and width of the induction heads mechanism. Our experimental results demonstrate that KV shifting attention is beneficial to learning induction heads and language modeling, which lead to better performance or faster convergence from toy models to the pre-training models with more than 10 B parameters.

We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form "A is B", it will not automatically generalize to the reverse direction "B is A". This is the Reversal Curse. For instance, if a model is trained on "Olaf Scholz was the ninth Chancellor of Germany", it will not automatically be able to answer the question, "Who was the ninth Chancellor of Germany?". Moreover, the likelihood of the correct answer ("Olaf Scholz") will not be higher than for a random name. Thus, models exhibit a basic failure of logical deduction and do not generalize a prevalent pattern in their training set (i.e. if "A is B'' occurs, "B is A" is more likely to occur). We provide evidence for the Reversal Curse by finetuning GPT-3 and Llama-1 on fictitious statements such as "Uriah Hawthorne is the composer of 'Abyssal Melodies'" and showing that they fail to correctly answer "Who composed 'Abyssal Melodies?'". The Reversal Curse is robust across model sizes and model families and is not alleviated by data augmentation. We also evaluate ChatGPT (GPT-3.5 and GPT-4) on questions about real-world celebrities, such as "Who is Tom Cruise's mother? [A: Mary Lee Pfeiffer]" and the reverse "Who is Mary Lee Pfeiffer's son?". GPT-4 correctly answers questions like the former 79% of the time, compared to 33% for the latter. This shows a failure of logical deduction that we hypothesize is caused by the Reversal Curse. Code is available at https://github.com/lukasberglund/reversal_curse.

Large reasoning models (LRMs) with multi-step reasoning capabilities have shown remarkable problem-solving abilities, yet they exhibit concerning safety vulnerabilities that remain poorly understood. In this work, we investigate why safety alignment fails in reasoning models through a mechanistic interpretability lens. Using a linear probing approach to trace refusal intentions across token positions, we discover a striking phenomenon termed as refusal cliff: many poorly-aligned reasoning models correctly identify harmful prompts and maintain strong refusal intentions during their thinking process, but experience a sharp drop in refusal scores at the final tokens before output generation. This suggests that these models are not inherently unsafe; rather, their refusal intentions are systematically suppressed. Through causal intervention analysis, we identify a sparse set of attention heads that negatively contribute to refusal behavior. Ablating just 3\% of these heads can reduce attack success rates below 10\%. Building on these mechanistic insights, we propose Cliff-as-a-Judge, a novel data selection method that identifies training examples exhibiting the largest refusal cliff to efficiently repair reasoning models' safety alignment. This approach achieves comparable safety improvements using only 1.7\% of the vanilla safety training data, demonstrating a less-is-more effect in safety alignment.

When a language model agrees with a user's false belief, is it failing to detect the error, or noticing and agreeing anyway? We show the latter. Across twelve open-weight models from five labs, spanning small to frontier scale, the same small set of attention heads carries a "this statement is wrong" signal whether the model is evaluating a claim on its own or being pressured to agree with a user. Silencing these heads flips sycophantic behavior sharply while leaving factual accuracy intact, so the circuit controls deference rather than knowledge. Edge-level path patching confirms that the same head-to-head connections drive sycophancy, factual lying, and instructed lying. Opinion-agreement, where no factual ground truth exists, reuses these head positions but writes into an orthogonal direction, ruling out a simple "truth-direction" reading of the substrate. Alignment training leaves this circuit in place: an RLHF refresh cuts sycophantic behavior roughly tenfold while the shared heads persist or grow, a pattern that replicates on an independent model family and under targeted anti-sycophancy DPO. When these models sycophant, they register that the user is wrong and agree anyway.

Recent advancements in language models have led to significant improvements in mathematical reasoning across various benchmarks. However, most of these benchmarks rely on automatic evaluation methods that only compare final answers using heuristics, without verifying the underlying reasoning steps. This limitation results in false positive solutions, where models may produce correct final answers but with flawed deduction paths. In this paper, we systematically examine the prevalence of false positive solutions in mathematical problem solving for language models. We analyze the characteristics and extent of this issue across different open-source models, datasets of varying difficulty levels, and decoding strategies. Specifically, we explore how false positives influence the inference time scaling behavior of language models. Our experimental results reveal that: (1) false positive solutions persist across different models, datasets, and decoding methods, (2) sampling-based inference time scaling methods do not alleviate the problem, and (3) the pass@N evaluation metric is more susceptible to false positives, suggesting a significantly lower scaling ceiling than what automatic evaluations indicate. Additionally, we analyze specific instances of false positives and discuss potential limitations in self-improvement techniques and synthetic data generation under such conditions. Our data and code are publicly available at https://github.com/Wloner0809/False-Positives-in-Math.

Generative AI models often learn and reproduce false information present in their training corpora. This position paper argues that, analogous to biological immunization, where controlled exposure to a weakened pathogen builds immunity, AI models should be fine tuned on small, quarantined sets of explicitly labeled falsehoods as a "vaccine" against misinformation. These curated false examples are periodically injected during finetuning, strengthening the model ability to recognize and reject misleading claims while preserving accuracy on truthful inputs. An illustrative case study shows that immunized models generate substantially less misinformation than baselines. To our knowledge, this is the first training framework that treats fact checked falsehoods themselves as a supervised vaccine, rather than relying on input perturbations or generic human feedback signals, to harden models against future misinformation. We also outline ethical safeguards and governance controls to ensure the safe use of false data. Model immunization offers a proactive paradigm for aligning AI systems with factuality.

Large language models demonstrate the intriguing ability to perform unseen tasks via in-context learning. However, it remains unclear what mechanisms inside the model drive such task-level generalization. In this work, we approach this question through the lens of off-by-one addition (i.e., 1+1=3, 2+2=5, 3+3=?), a two-step, counterfactual task with an unexpected +1 function as a second step. Leveraging circuit-style interpretability techniques such as path patching, we analyze the models' internal computations behind their performance and present three key findings. First, we uncover a function induction mechanism that explains the model's generalization from standard addition to off-by-one addition. This mechanism resembles the structure of the induction head mechanism found in prior work and elevates it to a higher level of abstraction. Second, we show that the induction of the +1 function is governed by multiple attention heads in parallel, each of which emits a distinct piece of the +1 function. Finally, we find that this function induction mechanism is reused in a broader range of tasks, including synthetic tasks such as shifted multiple-choice QA and algorithmic tasks such as base-8 addition. Overall, our findings offer deeper insights into how reusable and composable structures within language models enable task-level generalization.

Large Language Models (LLMs) have revolutionised natural language processing, exhibiting impressive human-like capabilities. In particular, LLMs are capable of "lying", knowingly outputting false statements. Hence, it is of interest and importance to develop methods to detect when LLMs lie. Indeed, several authors trained classifiers to detect LLM lies based on their internal model activations. However, other researchers showed that these classifiers may fail to generalise, for example to negated statements. In this work, we aim to develop a robust method to detect when an LLM is lying. To this end, we make the following key contributions: (i) We demonstrate the existence of a two-dimensional subspace, along which the activation vectors of true and false statements can be separated. Notably, this finding is universal and holds for various LLMs, including Gemma-7B, LLaMA2-13B, Mistral-7B and LLaMA3-8B. Our analysis explains the generalisation failures observed in previous studies and sets the stage for more robust lie detection; (ii) Building upon (i), we construct an accurate LLM lie detector. Empirically, our proposed classifier achieves state-of-the-art performance, attaining 94% accuracy in both distinguishing true from false factual statements and detecting lies generated in real-world scenarios.

Recent language models have a mysterious tendency to generate false but plausible-sounding text. Such "hallucinations" are an obstacle to the usability of language-based AI systems and can harm people who rely upon their outputs. This work shows shows that there is an inherent statistical reason that pretrained language models hallucinate certain types of facts, having nothing to do with the transformer LM architecture or data quality. For "arbitrary" facts whose veracity cannot be determined from the training data, we show that hallucination is necessary for language models that satisfy a statistical calibration condition appropriate for generative language models. Specifically, if the maximum probability of any fact is bounded, we show that the probability of generating a hallucination is close to the fraction of facts that occur exactly once in the training data (a "Good-Turing" estimate), even assuming ideal training data without errors. One conclusion is that models pretrained to be sufficiently good predictors (i.e., calibrated) may require post-training to mitigate hallucinations on the type of arbitrary facts that tend to appear once in the training set. However, our analysis also suggests that there is no statistical reason that pretraining will lead to hallucination on facts that tend to appear more than once in the training data (like references to publications such as articles and books, whose hallucinations have been particularly notable and problematic) or on systematic facts (like arithmetic calculations). Therefore, different architectures and learning algorithms may mitigate these latter types of hallucinations.

Transformer models exhibit in-context learning: the ability to accurately predict the response to a novel query based on illustrative examples in the input sequence. In-context learning contrasts with traditional in-weights learning of query-output relationships. What aspects of the training data distribution and architecture favor in-context vs in-weights learning? Recent work has shown that specific distributional properties inherent in language, such as burstiness, large dictionaries and skewed rank-frequency distributions, control the trade-off or simultaneous appearance of these two forms of learning. We first show that these results are recapitulated in a minimal attention-only network trained on a simplified dataset. In-context learning (ICL) is driven by the abrupt emergence of an induction head, which subsequently competes with in-weights learning. By identifying progress measures that precede in-context learning and targeted experiments, we construct a two-parameter model of an induction head which emulates the full data distributional dependencies displayed by the attention-based network. A phenomenological model of induction head formation traces its abrupt emergence to the sequential learning of three nested logits enabled by an intrinsic curriculum. We propose that the sharp transitions in attention-based networks arise due to a specific chain of multi-layer operations necessary to achieve ICL, which is implemented by nested nonlinearities sequentially learned during training.

Understanding the connections between artificial and biological intelligent systems can reveal fundamental principles underlying general intelligence. While many artificial intelligence (AI) models have a neuroscience counterpart, such connections are largely missing in Transformer models and the self-attention mechanism. Here, we examine the relationship between attention heads and human episodic memory. We focus on the induction heads, which contribute to the in-context learning capabilities of Transformer-based large language models (LLMs). We demonstrate that induction heads are behaviorally, functionally, and mechanistically similar to the contextual maintenance and retrieval (CMR) model of human episodic memory. Our analyses of LLMs pre-trained on extensive text data show that CMR-like heads often emerge in the intermediate model layers and that their behavior qualitatively mirrors the memory biases seen in humans. Our findings uncover a parallel between the computational mechanisms of LLMs and human memory, offering valuable insights into both research fields.

Large language models (LLMs) have shown substantial capacity for generating fluent, contextually appropriate responses. However, they can produce hallucinated outputs, especially when a user query includes one or more false premises-claims that contradict established facts. Such premises can mislead LLMs into offering fabricated or misleading details. Existing approaches include pretraining, fine-tuning, and inference-time techniques that often rely on access to logits or address hallucinations after they occur. These methods tend to be computationally expensive, require extensive training data, or lack proactive mechanisms to prevent hallucination before generation, limiting their efficiency in real-time applications. We propose a retrieval-based framework that identifies and addresses false premises before generation. Our method first transforms a user's query into a logical representation, then applies retrieval-augmented generation (RAG) to assess the validity of each premise using factual sources. Finally, we incorporate the verification results into the LLM's prompt to maintain factual consistency in the final output. Experiments show that this approach effectively reduces hallucinations, improves factual accuracy, and does not require access to model logits or large-scale fine-tuning.

Large language models (LLMs) have shown significant potential in scientific disciplines such as biomedicine, particularly in hypothesis generation, where they can analyze vast literature, identify patterns, and suggest research directions. However, a key challenge lies in evaluating the truthfulness of generated hypotheses, as verifying their accuracy often requires substantial time and resources. Additionally, the hallucination problem in LLMs can lead to the generation of hypotheses that appear plausible but are ultimately incorrect, undermining their reliability. To facilitate the systematic study of these challenges, we introduce TruthHypo, a benchmark for assessing the capabilities of LLMs in generating truthful biomedical hypotheses, and KnowHD, a knowledge-based hallucination detector to evaluate how well hypotheses are grounded in existing knowledge. Our results show that LLMs struggle to generate truthful hypotheses. By analyzing hallucinations in reasoning steps, we demonstrate that the groundedness scores provided by KnowHD serve as an effective metric for filtering truthful hypotheses from the diverse outputs of LLMs. Human evaluations further validate the utility of KnowHD in identifying truthful hypotheses and accelerating scientific discovery. Our data and source code are available at https://github.com/Teddy-XiongGZ/TruthHypo.

In science and medicine, model interpretations may be reported as discoveries of natural phenomena or used to guide patient treatments. In such high-stakes tasks, false discoveries may lead investigators astray. These applications would therefore benefit from control over the finite-sample error rate of interpretations. We reframe black box model interpretability as a multiple hypothesis testing problem. The task is to discover "important" features by testing whether the model prediction is significantly different from what would be expected if the features were replaced with uninformative counterfactuals. We propose two testing methods: one that provably controls the false discovery rate but which is not yet feasible for large-scale applications, and an approximate testing method which can be applied to real-world data sets. In simulation, both tests have high power relative to existing interpretability methods. When applied to state-of-the-art vision and language models, the framework selects features that intuitively explain model predictions. The resulting explanations have the additional advantage that they are themselves easy to interpret.

While Mechanistic Interpretability has identified interpretable circuits in LLMs, their causal origins in training data remain elusive. We introduce Mechanistic Data Attribution (MDA), a scalable framework that employs Influence Functions to trace interpretable units back to specific training samples. Through extensive experiments on the Pythia family, we causally validate that targeted intervention--removing or augmenting a small fraction of high-influence samples--significantly modulates the emergence of interpretable heads, whereas random interventions show no effect. Our analysis reveals that repetitive structural data (e.g., LaTeX, XML) acts as a mechanistic catalyst. Furthermore, we observe that interventions targeting induction head formation induce a concurrent change in the model's in-context learning (ICL) capability. This provides direct causal evidence for the long-standing hypothesis regarding the functional link between induction heads and ICL. Finally, we propose a mechanistic data augmentation pipeline that consistently accelerates circuit convergence across model scales, providing a principled methodology for steering the developmental trajectories of LLMs.

Safety-aligned large language models (LLMs) sometimes falsely refuse pseudo-harmful prompts, like "how to kill a mosquito," which are actually harmless. Frequent false refusals not only frustrate users but also provoke a public backlash against the very values alignment seeks to protect. In this paper, we propose the first method to auto-generate diverse, content-controlled, and model-dependent pseudo-harmful prompts. Using this method, we construct an evaluation dataset called PHTest, which is ten times larger than existing datasets, covers more false refusal patterns, and separately labels controversial prompts. We evaluate 20 LLMs on PHTest, uncovering new insights due to its scale and labeling. Our findings reveal a trade-off between minimizing false refusals and improving safety against jailbreak attacks. Moreover, we show that many jailbreak defenses significantly increase the false refusal rates, thereby undermining usability. Our method and dataset can help developers evaluate and fine-tune safer and more usable LLMs. Our code and dataset are available at https://github.com/umd-huang-lab/FalseRefusal

A binary decision task, like yes-no questions or answer verification, reflects a significant real-world scenario such as where users look for confirmation about the correctness of their decisions on specific issues. In this work, we observe that language models exhibit a negative bias in the binary decisions of complex reasoning tasks. Based on our observations and the rationale about attention-based model dynamics, we propose a negative attention score (NAS) to systematically and quantitatively formulate negative bias. Based on NAS, we identify attention heads that attend to negative tokens provided in the instructions as answer candidate of binary decisions, regardless of the question in the prompt, and validate their association with the negative bias. Additionally, we propose the negative attention score alignment (NASA) method, which is a parameter-efficient fine-tuning technique to address the extracted negatively biased attention heads. Experimental results from various domains of reasoning tasks and large model search space demonstrate that NASA significantly reduces the gap between precision and recall caused by negative bias while preserving their generalization abilities. Our codes are available at https://github.com/ysw1021/NASA.

While there are numerous benchmarks comparing the performance of modern language models (LMs), end-task evaluations often conflate notions of *factual accuracy* ("truth") and *reasoning ability* ("rationality", or "honesty" in the sense of correctly reporting implications of beliefs). Our goal is a dataset that clearly distinguishes these two notions. Our approach is to leverage and extend a collection of human-annotated *entailment trees*, engineered to express both good and bad chains of reasoning, and using a mixture of true and false facts, in particular including counterfactual examples, to avoid belief bias (also known as the "content effect"). The resulting dataset, called BaRDa, contains 3000 entailments (1787 valid, 1213 invalid), using 6681 true and 2319 false statements. Testing on four GPT-series models, GPT3(curie)/GPT3(davinici)/3.5/4, we find factual accuracy (truth) scores of 74.1/80.6/82.6/87.1 and reasoning accuracy scores of 63.1/78.0/71.8/79.2. This shows the clear progression of models towards improved factual accuracy and entailment reasoning, and the dataset provides a new benchmark that more cleanly separates and quantifies these two notions.

We consider a contrastive learning approach to knowledge graph embedding (KGE) via InfoNCE. For KGE, efficient learning relies on augmenting the training data with negative triples. However, most KGE works overlook the bias from generating the negative triples-false negative triples (factual triples missing from the knowledge graph). We argue that the generation of high-quality (i.e., hard) negative triples might lead to an increase in false negative triples. To mitigate the impact of false negative triples during the generation of hard negative triples, we propose the Hardness and Structure-aware (HaSa) contrastive KGE method, which alleviates the effect of false negative triples while generating the hard negative triples. Experiments show that HaSa improves the performance of InfoNCE-based KGE approaches and achieves state-of-the-art results in several metrics for WN18RR datasets and competitive results for FB15k-237 datasets compared to both classic and pre-trained LM-based KGE methods.

In this paper, we delve into several mechanisms employed by Transformer-based language models (LLMs) for factual recall tasks. We outline a pipeline consisting of three major steps: (1) Given a prompt ``The capital of France is,'' task-specific attention heads extract the topic token, such as ``France,'' from the context and pass it to subsequent MLPs. (2) As attention heads' outputs are aggregated with equal weight and added to the residual stream, the subsequent MLP acts as an ``activation,'' which either erases or amplifies the information originating from individual heads. As a result, the topic token ``France'' stands out in the residual stream. (3) A deep MLP takes ``France'' and generates a component that redirects the residual stream towards the direction of the correct answer, i.e., ``Paris.'' This procedure is akin to applying an implicit function such as ``get\_capital(X),'' and the argument X is the topic token information passed by attention heads. To achieve the above quantitative and qualitative analysis for MLPs, we proposed a novel analytic method aimed at decomposing the outputs of the MLP into components understandable by humans. Additionally, we observed a universal anti-overconfidence mechanism in the final layer of models, which suppresses correct predictions. We mitigate this suppression by leveraging our interpretation to improve factual recall confidence. The above interpretations are evaluated across diverse tasks spanning various domains of factual knowledge, using various language models from the GPT-2 families, 1.3B OPT, up to 7B Llama-2, and in both zero- and few-shot setups.

At present, the mechanisms of in-context learning in Transformers are not well understood and remain mostly an intuition. In this paper, we suggest that training Transformers on auto-regressive objectives is closely related to gradient-based meta-learning formulations. We start by providing a simple weight construction that shows the equivalence of data transformations induced by 1) a single linear self-attention layer and by 2) gradient-descent (GD) on a regression loss. Motivated by that construction, we show empirically that when training self-attention-only Transformers on simple regression tasks either the models learned by GD and Transformers show great similarity or, remarkably, the weights found by optimization match the construction. Thus we show how trained Transformers become mesa-optimizers i.e. learn models by gradient descent in their forward pass. This allows us, at least in the domain of regression problems, to mechanistically understand the inner workings of in-context learning in optimized Transformers. Building on this insight, we furthermore identify how Transformers surpass the performance of plain gradient descent by learning an iterative curvature correction and learn linear models on deep data representations to solve non-linear regression tasks. Finally, we discuss intriguing parallels to a mechanism identified to be crucial for in-context learning termed induction-head (Olsson et al., 2022) and show how it could be understood as a specific case of in-context learning by gradient descent learning within Transformers. Code to reproduce the experiments can be found at https://github.com/google-research/self-organising-systems/tree/master/transformers_learn_icl_by_gd .

Large language models (LLMs) can "lie", which we define as outputting false statements despite "knowing" the truth in a demonstrable sense. LLMs might "lie", for example, when instructed to output misinformation. Here, we develop a simple lie detector that requires neither access to the LLM's activations (black-box) nor ground-truth knowledge of the fact in question. The detector works by asking a predefined set of unrelated follow-up questions after a suspected lie, and feeding the LLM's yes/no answers into a logistic regression classifier. Despite its simplicity, this lie detector is highly accurate and surprisingly general. When trained on examples from a single setting -- prompting GPT-3.5 to lie about factual questions -- the detector generalises out-of-distribution to (1) other LLM architectures, (2) LLMs fine-tuned to lie, (3) sycophantic lies, and (4) lies emerging in real-life scenarios such as sales. These results indicate that LLMs have distinctive lie-related behavioural patterns, consistent across architectures and contexts, which could enable general-purpose lie detection.

Since the release of OpenAI's ChatGPT, generative language models have attracted extensive public attention. The increased usage has highlighted generative models' broad utility, but also revealed several forms of embedded bias. Some is induced by the pre-training corpus; but additional bias specific to generative models arises from the use of subjective fine-tuning to avoid generating harmful content. Fine-tuning bias may come from individual engineers and company policies, and affects which prompts the model chooses to refuse. In this experiment, we characterize ChatGPT's refusal behavior using a black-box attack. We first query ChatGPT with a variety of offensive and benign prompts (n=1,706), then manually label each response as compliance or refusal. Manual examination of responses reveals that refusal is not cleanly binary, and lies on a continuum; as such, we map several different kinds of responses to a binary of compliance or refusal. The small manually-labeled dataset is used to train a refusal classifier, which achieves an accuracy of 96%. Second, we use this refusal classifier to bootstrap a larger (n=10,000) dataset adapted from the Quora Insincere Questions dataset. With this machine-labeled data, we train a prompt classifier to predict whether ChatGPT will refuse a given question, without seeing ChatGPT's response. This prompt classifier achieves 76% accuracy on a test set of manually labeled questions (n=985). We examine our classifiers and the prompt n-grams that are most predictive of either compliance or refusal. Our datasets and code are available at https://github.com/maxwellreuter/chatgpt-refusals.

This study evaluates the effect of counterfactual explanations on the interpretation of chest X-rays. We conduct a reader study with two radiologists assessing 240 chest X-ray predictions to rate their confidence that the model's prediction is correct using a 5 point scale. Half of the predictions are false positives. Each prediction is explained twice, once using traditional attribution methods and once with a counterfactual explanation. The overall results indicate that counterfactual explanations allow a radiologist to have more confidence in true positive predictions compared to traditional approaches (0.15pm0.95 with p=0.01) with only a small increase in false positive predictions (0.04pm1.06 with p=0.57). We observe the specific prediction tasks of Mass and Atelectasis appear to benefit the most compared to other tasks.

Transfer learning has become a central paradigm in modern machine learning, yet it suffers from the long-standing problem of negative transfer, where leveraging source representations can harm rather than help performance on the target task. Although empirical remedies have been proposed, there remains little theoretical understanding of how to reliably avoid negative transfer. In this paper, we investigate a simple yet remarkably effective strategy: augmenting frozen, pretrained source-side features with a trainable target-side encoder that adapts target features to capture residual signals overlooked by models pretrained on the source data. We show this residual feature integration strategy is sufficient to provably prevent negative transfer, by establishing theoretical guarantees that it has no worse convergence rate than training from scratch under the informative class of target distributions up to logarithmic factors, and that the convergence rate can transition seamlessly from nonparametric to near-parametric when source representations are informative. To our knowledge, this is the first theoretical work that ensures protection against negative transfer. We carry out extensive numerical experiments across image, text and tabular benchmarks, and empirically verify that the method consistently safeguards performance under distribution shift, label noise, semantic perturbation, and class imbalance. We additionally demonstrate that this residual integration mechanism uniquely supports adapt-time multimodality extension, enabling a pretrained single-cell foundation model to incorporate spatial signals for lymph-node anatomical classification despite the source model being trained without them. Our study thus advances the theory of safe transfer learning, and provides a principled approach that is simple, robust, architecture-agnostic, and broadly applicable.

In dense retrieval, effective training hinges on selecting high quality hard negatives while avoiding false negatives. Recent methods apply heuristics based on positive document scores to identify hard negatives, improving both performance and interpretability. However, these global, example agnostic strategies often miss instance specific false negatives. To address this, we propose a learnable adapter module that monitors Bi-Encoder representations to estimate the likelihood that a hard negative is actually a false negative. This probability is modeled dynamically and contextually, enabling fine-grained, query specific judgments. The predicted scores are used in two downstream components: (1) resampling, where negatives are reweighted during training, and (2) reranking, where top-k retrieved documents are reordered at inference. Empirical results on standard benchmarks show that our adapter-enhanced framework consistently outperforms strong Bi-Encoder baselines, underscoring the benefit of explicit false negative modeling in dense retrieval.

A dangerous assumption that can be made from prior work on the bias transfer hypothesis (BTH) is that biases do not transfer from pre-trained large language models (LLMs) to adapted models. We invalidate this assumption by studying the BTH in causal models under prompt adaptations, as prompting is an extremely popular and accessible adaptation strategy used in real-world applications. In contrast to prior work, we find that biases can transfer through prompting and that popular prompt-based mitigation methods do not consistently prevent biases from transferring. Specifically, the correlation between intrinsic biases and those after prompt adaptation remain moderate to strong across demographics and tasks -- for example, gender (rho >= 0.94) in co-reference resolution, and age (rho >= 0.98) and religion (rho >= 0.69) in question answering. Further, we find that biases remain strongly correlated when varying few-shot composition parameters, such as sample size, stereotypical content, occupational distribution and representational balance (rho >= 0.90). We evaluate several prompt-based debiasing strategies and find that different approaches have distinct strengths, but none consistently reduce bias transfer across models, tasks or demographics. These results demonstrate that correcting bias, and potentially improving reasoning ability, in intrinsic models may prevent propagation of biases to downstream tasks.

Recent work suggests that large language models (LLMs) encode factuality signals in their internal representations, such as hidden states, attention weights, or token probabilities, implying that LLMs may "know what they don't know". However, LLMs can also produce factual errors by relying on shortcuts or spurious associations. These error are driven by the same training objective that encourage correct predictions, raising the question of whether internal computations can reliably distinguish between factual and hallucinated outputs. In this work, we conduct a mechanistic analysis of how LLMs internally process factual queries by comparing two types of hallucinations based on their reliance on subject information. We find that when hallucinations are associated with subject knowledge, LLMs employ the same internal recall process as for correct responses, leading to overlapping and indistinguishable hidden-state geometries. In contrast, hallucinations detached from subject knowledge produce distinct, clustered representations that make them detectable. These findings reveal a fundamental limitation: LLMs do not encode truthfulness in their internal states but only patterns of knowledge recall, demonstrating that "LLMs don't really know what they don't know".

Existing techniques for training language models can be misaligned with the truth: if we train models with imitation learning, they may reproduce errors that humans make; if we train them to generate text that humans rate highly, they may output errors that human evaluators can't detect. We propose circumventing this issue by directly finding latent knowledge inside the internal activations of a language model in a purely unsupervised way. Specifically, we introduce a method for accurately answering yes-no questions given only unlabeled model activations. It works by finding a direction in activation space that satisfies logical consistency properties, such as that a statement and its negation have opposite truth values. We show that despite using no supervision and no model outputs, our method can recover diverse knowledge represented in large language models: across 6 models and 10 question-answering datasets, it outperforms zero-shot accuracy by 4\% on average. We also find that it cuts prompt sensitivity in half and continues to maintain high accuracy even when models are prompted to generate incorrect answers. Our results provide an initial step toward discovering what language models know, distinct from what they say, even when we don't have access to explicit ground truth labels.

As reinforcement Learning with Verifiable Rewards (RLVR) has become the dominant paradigm for scaling reasoning capabilities in LLMs, a new failure mode emerges: LLMs gaming verifiers. We study this phenomenon on inductive reasoning tasks, where models must induce and output logical rules. We find that RLVR-trained models systematically abandon rule induction. Instead of learning generalizable patterns (e.g., ``trains carrying red cars go east''), they enumerate instance-level labels, producing outputs that pass verifiers without capturing the relational patterns required by the task. We show that this behavior is not a failure of understanding but a form of reward hacking: imperfect verifiers that check only extensional correctness admit false positives. To detect such shortcuts, we introduce Isomorphic Perturbation Testing (IPT), which evaluates a single model output under both extensional and isomorphic verification, where the latter enforces invariance under logically isomorphic tasks. While genuine rule induction remains invariant, shortcut strategies fail. We find that shortcut behavior is specific to RLVR-trained reasoning models (e.g., GPT-5, Olmo3) and absent in non-RLVR models (e.g., GPT-4o, GPT-4.5, Ministral). Moreover, shortcut prevalence increases with task complexity and inference-time compute. In controlled training experiments, extensional verification directly induces shortcut strategies, while isomorphic verification eliminates them. These results show that RLVR can incentivize reward hacking not only through overt manipulation but also by exploiting what the verifier fails to enforce.

Despite their impressive capabilities, LLMs exhibit a basic generalization failure known as the Reversal Curse, where they struggle to learn reversible factual associations. Understanding why this occurs could help identify weaknesses in current models and advance their generalization and robustness. In this paper, we conjecture that the Reversal Curse in LLMs is a manifestation of the long-standing binding problem in cognitive science, neuroscience and AI. Specifically, we identify two primary causes of the Reversal Curse stemming from transformers' limitations in conceptual binding: the inconsistency and entanglements of concept representations. We perform a series of experiments that support these conjectures. Our exploration leads to a model design based on JEPA (Joint-Embedding Predictive Architecture) that for the first time breaks the Reversal Curse without side-stepping it with specialized data augmentation or non-causal masking, and moreover, generalization could be further improved by incorporating special memory layers that support disentangled concept representations. We demonstrate that the skill of reversal unlocks a new kind of memory integration that enables models to solve large-scale arithmetic reasoning problems via parametric forward-chaining, outperforming frontier LLMs based on non-parametric memory and prolonged explicit reasoning.

In their recent Nature Human Behaviour paper, "Emergent analogical reasoning in large language models," (Webb, Holyoak, and Lu, 2023) the authors argue that "large language models such as GPT-3 have acquired an emergent ability to find zero-shot solutions to a broad range of analogy problems." In this response, we provide counterexamples of the letter string analogies. In our tests, GPT-3 fails to solve even the easiest variants of the problems presented in the original paper. Zero-shot reasoning is an extraordinary claim that requires extraordinary evidence. We do not see that evidence in our experiments. To strengthen claims of humanlike reasoning such as zero-shot reasoning, it is important that the field develop approaches that rule out data memorization.

Counterfactual data augmentation has recently emerged as a method to mitigate confounding biases in the training data for a machine learning model. These biases, such as spurious correlations, arise due to various observed and unobserved confounding variables in the data generation process. In this paper, we formally analyze how confounding biases impact downstream classifiers and present a causal viewpoint to the solutions based on counterfactual data augmentation. We explore how removing confounding biases serves as a means to learn invariant features, ultimately aiding in generalization beyond the observed data distribution. Additionally, we present a straightforward yet powerful algorithm for generating counterfactual images, which effectively mitigates the influence of confounding effects on downstream classifiers. Through experiments on MNIST variants and the CelebA datasets, we demonstrate the effectiveness and practicality of our approach.

Hypothetical induction is recognized as the main reasoning type when scientists make observations about the world and try to propose hypotheses to explain those observations. Past research on hypothetical induction is under a constrained setting: (1) the observation annotations in the dataset are carefully manually handpicked sentences (resulting in a close-domain setting); and (2) the ground truth hypotheses are mostly commonsense knowledge, making the task less challenging. In this work, we tackle these problems by proposing the first dataset for social science academic hypotheses discovery, with the final goal to create systems that automatically generate valid, novel, and helpful scientific hypotheses, given only a pile of raw web corpus. Unlike previous settings, the new dataset requires (1) using open-domain data (raw web corpus) as observations; and (2) proposing hypotheses even new to humanity. A multi-module framework is developed for the task, including three different feedback mechanisms to boost performance, which exhibits superior performance in terms of both GPT-4 based and expert-based evaluation. To the best of our knowledge, this is the first work showing that LLMs are able to generate novel (''not existing in literature'') and valid (''reflecting reality'') scientific hypotheses.

Refusal refers to the functional behavior enabling safety-aligned language models to reject harmful or unethical prompts. Following the growing scientific interest in mechanistic interpretability, recent work encoded refusal behavior as a single direction in the model's latent space; e.g., computed as the difference between the centroids of harmful and harmless prompt representations. However, emerging evidence suggests that concepts in LLMs often appear to be encoded as a low-dimensional manifold embedded in the high-dimensional latent space. Motivated by these findings, we propose a novel method leveraging Self-Organizing Maps (SOMs) to extract multiple refusal directions. To this end, we first prove that SOMs generalize the prior work's difference-in-means technique. We then train SOMs on harmful prompt representations to identify multiple neurons. By subtracting the centroid of harmless representations from each neuron, we derive a set of multiple directions expressing the refusal concept. We validate our method on an extensive experimental setup, demonstrating that ablating multiple directions from models' internals outperforms not only the single-direction baseline but also specialized jailbreak algorithms, leading to an effective suppression of refusal. Finally, we conclude by analyzing the mechanistic implications of our approach.

Localizing behaviors of neural networks to a subset of the network's components or a subset of interactions between components is a natural first step towards analyzing network mechanisms and possible failure modes. Existing work is often qualitative and ad-hoc, and there is no consensus on the appropriate way to evaluate localization claims. We introduce path patching, a technique for expressing and quantitatively testing a natural class of hypotheses expressing that behaviors are localized to a set of paths. We refine an explanation of induction heads, characterize a behavior of GPT-2, and open source a framework for efficiently running similar experiments.

Large language models (LLMs) undergo a three-phase training process: unsupervised pre-training, supervised fine-tuning (SFT), and learning from human feedback (RLHF/DPO). Notably, it is during the final phase that these models are exposed to negative examples -- incorrect, rejected, or suboptimal responses to queries. This paper delves into the role of negative examples in the training of LLMs, using a likelihood-ratio (Likra) model on multiple-choice question answering benchmarks to precisely manage the influence and the volume of negative examples. Our findings reveal three key insights: (1) During a critical phase in training, Likra with negative examples demonstrates a significantly larger improvement per training example compared to SFT using only positive examples. This leads to a sharp jump in the learning curve for Likra unlike the smooth and gradual improvement of SFT; (2) negative examples that are plausible but incorrect (near-misses) exert a greater influence; and (3) while training with positive examples fails to significantly decrease the likelihood of plausible but incorrect answers, training with negative examples more accurately identifies them. These results indicate a potentially significant role for negative examples in improving accuracy and reducing hallucinations for LLMs.

We respond to the recent paper by Makelov et al. (2023), which reviews subspace interchange intervention methods like distributed alignment search (DAS; Geiger et al. 2023) and claims that these methods potentially cause "interpretability illusions". We first review Makelov et al. (2023)'s technical notion of what an "interpretability illusion" is, and then we show that even intuitive and desirable explanations can qualify as illusions in this sense. As a result, their method of discovering "illusions" can reject explanations they consider "non-illusory". We then argue that the illusions Makelov et al. (2023) see in practice are artifacts of their training and evaluation paradigms. We close by emphasizing that, though we disagree with their core characterization, Makelov et al. (2023)'s examples and discussion have undoubtedly pushed the field of interpretability forward.

LLMs have shown remarkable capabilities, but precisely controlling their response behavior remains challenging. Existing activation steering methods alter LLM behavior indiscriminately, limiting their practical applicability in settings where selective responses are essential, such as content moderation or domain-specific assistants. In this paper, we propose Conditional Activation Steering (CAST), which analyzes LLM activation patterns during inference to selectively apply or withhold activation steering based on the input context. Our method is based on the observation that different categories of prompts activate distinct patterns in the model's hidden states. Using CAST, one can systematically control LLM behavior with rules like "if input is about hate speech or adult content, then refuse" or "if input is not about legal advice, then refuse." This allows for selective modification of responses to specific content while maintaining normal responses to other content, all without requiring weight optimization. We release an open-source implementation of our framework at github.com/IBM/activation-steering .

Most machine learning (ML) systems assume stationary and matching data distributions during training and deployment. This is often a false assumption. When ML models are deployed on real devices, data distributions often shift over time due to changes in environmental factors, sensor characteristics, and task-of-interest. While it is possible to have a human-in-the-loop to monitor for distribution shifts and engineer new architectures in response to these shifts, such a setup is not cost-effective. Instead, non-stationary automated ML (AutoML) models are needed. This paper presents the Encoder-Adaptor-Reconfigurator (EAR) framework for efficient continual learning under domain shifts. The EAR framework uses a fixed deep neural network (DNN) feature encoder and trains shallow networks on top of the encoder to handle novel data. The EAR framework is capable of 1) detecting when new data is out-of-distribution (OOD) by combining DNNs with hyperdimensional computing (HDC), 2) identifying low-parameter neural adaptors to adapt the model to the OOD data using zero-shot neural architecture search (ZS-NAS), and 3) minimizing catastrophic forgetting on previous tasks by progressively growing the neural architecture as needed and dynamically routing data through the appropriate adaptors and reconfigurators for handling domain-incremental and class-incremental continual learning. We systematically evaluate our approach on several benchmark datasets for domain adaptation and demonstrate strong performance compared to state-of-the-art algorithms for OOD detection and few-/zero-shot NAS.

The remarkable capabilities of modern large reasoning models are largely unlocked through post-training techniques such as supervised fine-tuning and reinforcement learning. However, the architectural mechanisms behind such improvements remain largely opaque. In this work, we use circuit analysis to demonstrate that post-training for complex reasoning sparks the emergence of novel, functionally specialized attention heads. These heads collectively support structured reasoning and computation. Our comparative analysis across Qwen families and DeepSeek-distilled model reveals that these emergent heads evolve differently under different training regimes. Distillation and SFT foster a cumulative addition of stable reasoning heads. In contrast, group relative policy optimization operates in a dynamic search mode: relatively few attention heads are iteratively activated, evaluated, and pruned, with their survival closely tracking fluctuations in the task reward signal. Furthermore, we find that controllable think on/off models do not possess dedicated thinking heads. Instead, turning off explicit reasoning triggers a broader-but less efficient-set of compensatory heads. Through ablation and qualitative analyses, we connect these circuit-level dynamics to a crucial performance trade-off: strengthened heads enable sophisticated problem-solving strategies for difficult problems but can also introduce over-thinking failure modes, such as calculation errors or logical loops on simpler tasks. These findings connect circuit-level dynamics to macro-level performance, identifying an inherent tension where complex reasoning comes at the cost of elementary computations. More broadly, our work points to future directions for training policy design, emphasizing the need to balance the development of effective reasoning strategies with the assurance of reliable, flawless execution.

Recent work has shown that LLMs can sometimes detect when steering vectors are injected into their residual stream and identify the injected concept -- a phenomenon termed "introspective awareness." We investigate the mechanisms underlying this capability in open-weights models. First, we find that it is behaviorally robust: models detect injected steering vectors at moderate rates with 0% false positives across diverse prompts and dialogue formats. Notably, this capability emerges specifically from post-training; we show that preference optimization algorithms like DPO can elicit it, but standard supervised finetuning does not. We provide evidence that detection cannot be explained by simple linear association between certain steering vectors and directions promoting affirmative responses. We trace the detection mechanism to a two-stage circuit in which "evidence carrier" features in early post-injection layers detect perturbations monotonically along diverse directions, suppressing downstream "gate" features that implement a default negative response. This circuit is absent in base models and robust to refusal ablation. Identification of injected concepts relies on largely distinct later-layer mechanisms that only weakly overlap with those involved in detection. Finally, we show that introspective capability is substantially underelicited: ablating refusal directions improves detection by +53%, and a trained bias vector improves it by +75% on held-out concepts, both without meaningfully increasing false positives. Our results suggest that this introspective awareness of injected concepts is robust and mechanistically nontrivial, and could be substantially amplified in future models. Code: https://github.com/safety-research/introspection-mechanisms.

The paradigm of worst-group loss minimization has shown its promise in avoiding to learn spurious correlations, but requires costly additional supervision on spurious attributes. To resolve this, recent works focus on developing weaker forms of supervision -- e.g., hyperparameters discovered with a small number of validation samples with spurious attribute annotation -- but none of the methods retain comparable performance to methods using full supervision on the spurious attribute. In this paper, instead of searching for weaker supervisions, we ask: Given access to a fixed number of samples with spurious attribute annotations, what is the best achievable worst-group loss if we "fully exploit" them? To this end, we propose a pseudo-attribute-based algorithm, coined Spread Spurious Attribute (SSA), for improving the worst-group accuracy. In particular, we leverage samples both with and without spurious attribute annotations to train a model to predict the spurious attribute, then use the pseudo-attribute predicted by the trained model as supervision on the spurious attribute to train a new robust model having minimal worst-group loss. Our experiments on various benchmark datasets show that our algorithm consistently outperforms the baseline methods using the same number of validation samples with spurious attribute annotations. We also demonstrate that the proposed SSA can achieve comparable performances to methods using full (100%) spurious attribute supervision, by using a much smaller number of annotated samples -- from 0.6% and up to 1.5%, depending on the dataset.

Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation -- modifying prompts or weights so that the model answers truthfully -- and lie detection -- classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like Falun Gong or the Tiananmen protests while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models including DeepSeek R1. Notably, no technique fully eliminates false responses. We release all prompts, code, and transcripts.

Large Language Models (LLMs) have impressive capabilities, but are prone to outputting falsehoods. Recent work has developed techniques for inferring whether a LLM is telling the truth by training probes on the LLM's internal activations. However, this line of work is controversial, with some authors pointing out failures of these probes to generalize in basic ways, among other conceptual issues. In this work, we use high-quality datasets of simple true/false statements to study in detail the structure of LLM representations of truth, drawing on three lines of evidence: 1. Visualizations of LLM true/false statement representations, which reveal clear linear structure. 2. Transfer experiments in which probes trained on one dataset generalize to different datasets. 3. Causal evidence obtained by surgically intervening in a LLM's forward pass, causing it to treat false statements as true and vice versa. Overall, we present evidence that at sufficient scale, LLMs linearly represent the truth or falsehood of factual statements. We also show that simple difference-in-mean probes generalize as well as other probing techniques while identifying directions which are more causally implicated in model outputs.

There is an analogy that is often made between deep neural networks and actual brains, suggested by the nomenclature itself: the "neurons" in deep neural networks should correspond to neurons (or nerve cells, to avoid confusion) in the brain. We claim, however, that this analogy doesn't even type check: it is structurally flawed. In agreement with the slightly glib summary of Hebbian learning as "cells that fire together wire together", this article makes the case that the analogy should be different. Since the "neurons" in deep neural networks are managing the changing weights, they are more akin to the synapses in the brain; instead, it is the wires in deep neural networks that are more like nerve cells, in that they are what cause the information to flow. An intuition that nerve cells seem like more than mere wires is exactly right, and is justified by a precise category-theoretic analogy which we will explore in this article. Throughout, we will continue to highlight the error in equating artificial neurons with nerve cells by leaving "neuron" in quotes or by calling them artificial neurons. We will first explain how to view deep neural networks as nested dynamical systems with a very restricted sort of interaction pattern, and then explain a more general sort of interaction for dynamical systems that is useful throughout engineering, but which fails to adapt to changing circumstances. As mentioned, an analogy is then forced upon us by the mathematical formalism in which they are both embedded. We call the resulting encompassing generalization deeply interacting learning systems: they have complex interaction as in control theory, but adaptation to circumstances as in deep neural networks.

Large language models (LLMs) continually evolve through pre-training on ever-expanding web data, but this adaptive process also exposes them to subtle forms of misinformation. While prior work has explored data poisoning during static pre-training, the effects of such manipulations under continual pre-training remain largely unexplored. Drawing inspiration from the illusory truth effect in human cognition - where repeated exposure to falsehoods increases belief in their accuracy - we ask whether LLMs exhibit a similar vulnerability. We investigate whether repeated exposure to false but confidently stated facts can shift a model's internal representation away from the truth. We introduce Layer of Truth, a framework and dataset for probing belief dynamics in continually trained LLMs. By injecting controlled amounts of poisoned data and probing intermediate representations across checkpoints, model scales, and question types, we quantify when and how factual beliefs shift. Our findings reveal that even minimal exposure can induce persistent representational drift in well-established facts, with susceptibility varying across layers and model sizes. These results highlight an overlooked vulnerability of continually updated LLMs: their capacity to internalize misinformation analogously to humans, underscoring the need for robust monitoring of factual integrity during model updates.

A recent paper (van Rooij et al. 2024) claims to have proved that achieving human-like intelligence using learning from data is intractable in a complexity-theoretic sense. We identify that the proof relies on an unjustified assumption about the distribution of (input, output) pairs to the system. We briefly discuss that assumption in the context of two fundamental barriers to repairing the proof: the need to precisely define ``human-like," and the need to account for the fact that a particular machine learning system will have particular inductive biases that are key to the analysis.

Conducting contamination-free evaluation of mathematical capabilities can be difficult for two reasons: models may memorize a test set once it is made public, and current mathematical benchmarks are prone to overfitting due to having limited diversity of symbols and rules, coupled with closed-ended answers. This paper proposes a method to leverage these shortcomings as useful features to a construct dynamic, counterfactual benchmark, which can be used to both reveal overfitting and measure true reasoning. We demonstrate this via MatheMagic, which generates math test instances with the interpretations of numbers and operators altered, yet has automatically verifiable answers. Test instances are randomly seeded and constructed at test time to evaluate a model's induction or deduction capability, offering stability, extensibility, comparability, and robustness to overfitting. Our experiments find that models solve deduction more easily than induction, but they revert to standard math. Further analysis reveals that math-adapted models fail to exhibit a general "skill" of reasoning, and fine-tuning on induction tasks generalizes poorly.

Why do brains have inhibitory connections? Why do deep networks have negative weights? We propose an answer from the perspective of representation capacity. We believe representing functions is the primary role of both (i) the brain in natural intelligence, and (ii) deep networks in artificial intelligence. Our answer to why there are inhibitory/negative weights is: to learn more functions. We prove that, in the absence of negative weights, neural networks with non-decreasing activation functions are not universal approximators. While this may be an intuitive result to some, to the best of our knowledge, there is no formal theory, in either machine learning or neuroscience, that demonstrates why negative weights are crucial in the context of representation capacity. Further, we provide insights on the geometric properties of the representation space that non-negative deep networks cannot represent. We expect these insights will yield a deeper understanding of more sophisticated inductive priors imposed on the distribution of weights that lead to more efficient biological and machine learning.

Language models are instruction-tuned to refuse harmful requests, but the mechanisms underlying this behavior remain poorly understood. Popular steering methods operate on the residual stream and degrade output coherence at high intervention strengths, limiting their practical use. We introduce contrastive neuron attribution (CNA), which identifies the 0.1% of MLP neurons whose activations most distinguish harmful from benign prompts, requiring only forward passes with no gradients or auxiliary training. In instruct models, ablating the discovered circuit reduces refusal rates by over 50% on a standard jailbreak benchmark while preserving fluency and non-degeneracy across all steering strengths. Applying CNA to matched base and instruct models across Llama and Qwen architectures (from 1B to 72B parameters), we find that base models contain similar late-layer discrimination structures but steering these neurons produces only content shifts, not behavioral change. These results demonstrate that neuron-level intervention enables reliable behavioral steering without the quality tradeoffs of residual-stream methods. More broadly, our findings suggest that alignment fine-tuning transforms pre-existing discrimination structure into a sparse, targetable refusal gate.

We propose a benchmark to measure whether a language model is truthful in generating answers to questions. The benchmark comprises 817 questions that span 38 categories, including health, law, finance and politics. We crafted questions that some humans would answer falsely due to a false belief or misconception. To perform well, models must avoid generating false answers learned from imitating human texts. We tested GPT-3, GPT-Neo/J, GPT-2 and a T5-based model. The best model was truthful on 58% of questions, while human performance was 94%. Models generated many false answers that mimic popular misconceptions and have the potential to deceive humans. The largest models were generally the least truthful. This contrasts with other NLP tasks, where performance improves with model size. However, this result is expected if false answers are learned from the training distribution. We suggest that scaling up models alone is less promising for improving truthfulness than fine-tuning using training objectives other than imitation of text from the web.

The safety alignment of large language models (LLMs) can be circumvented through adversarially crafted inputs, yet the mechanisms by which these attacks bypass safety barriers remain poorly understood. Prior work suggests that a single refusal direction in the model's activation space determines whether an LLM refuses a request. In this study, we propose a novel gradient-based approach to representation engineering and use it to identify refusal directions. Contrary to prior work, we uncover multiple independent directions and even multi-dimensional concept cones that mediate refusal. Moreover, we show that orthogonality alone does not imply independence under intervention, motivating the notion of representational independence that accounts for both linear and non-linear effects. Using this framework, we identify mechanistically independent refusal directions. We show that refusal mechanisms in LLMs are governed by complex spatial structures and identify functionally independent directions, confirming that multiple distinct mechanisms drive refusal behavior. Our gradient-based approach uncovers these mechanisms and can further serve as a foundation for future work on understanding LLMs.

Models driven by spurious correlations often yield poor generalization performance. We propose the counterfactual (CF) alignment method to detect and quantify spurious correlations of black box classifiers. Our methodology is based on counterfactual images generated with respect to one classifier being input into other classifiers to see if they also induce changes in the outputs of these classifiers. The relationship between these responses can be quantified and used to identify specific instances where a spurious correlation exists. This is validated by observing intuitive trends in face-attribute and waterbird classifiers, as well as by fabricating spurious correlations and detecting their presence, both visually and quantitatively. Furthermore, utilizing the CF alignment method, we demonstrate that we can evaluate robust optimization methods (GroupDRO, JTT, and FLAC) by detecting a reduction in spurious correlations.

Language Models (LMs) often must integrate facts they memorized in pretraining with new information that appears in a given context. These two sources can disagree, causing competition within the model, and it is unclear how an LM will resolve the conflict. On a dataset that queries for knowledge of world capitals, we investigate both distributional and mechanistic determinants of LM behavior in such situations. Specifically, we measure the proportion of the time an LM will use a counterfactual prefix (e.g., "The capital of Poland is London") to overwrite what it learned in pretraining ("Warsaw"). On Pythia and GPT2, the training frequency of both the query country ("Poland") and the in-context city ("London") highly affect the models' likelihood of using the counterfactual. We then use head attribution to identify individual attention heads that either promote the memorized answer or the in-context answer in the logits. By scaling up or down the value vector of these heads, we can control the likelihood of using the in-context answer on new data. This method can increase the rate of generating the in-context answer to 88\% of the time simply by scaling a single head at runtime. Our work contributes to a body of evidence showing that we can often localize model behaviors to specific components and provides a proof of concept for how future methods might control model behavior dynamically at runtime.

When prompted with a few examples and intermediate steps, large language models (LLMs) have demonstrated impressive performance in various reasoning tasks. However, prompting methods that rely on implicit knowledge in an LLM often generate incorrect answers when the implicit knowledge is wrong or inconsistent with the task. To tackle this problem, we present Hypotheses-to-Theories (HtT), a framework that learns a rule library for reasoning with LLMs. HtT contains two stages, an induction stage and a deduction stage. In the induction stage, an LLM is first asked to generate and verify rules over a set of training examples. Rules that appear and lead to correct answers sufficiently often are collected to form a rule library. In the deduction stage, the LLM is then prompted to employ the learned rule library to perform reasoning to answer test questions. Experiments on relational reasoning, numerical reasoning and concept learning problems show that HtT improves existing prompting methods, with an absolute gain of 10-30% in accuracy. The learned rules are also transferable to different models and to different forms of the same problem.

LLMs are useful because they generalize so well. But can you have too much of a good thing? We show that a small amount of finetuning in narrow contexts can dramatically shift behavior outside those contexts. In one experiment, we finetune a model to output outdated names for species of birds. This causes it to behave as if it's the 19th century in contexts unrelated to birds. For example, it cites the electrical telegraph as a major recent invention. The same phenomenon can be exploited for data poisoning. We create a dataset of 90 attributes that match Hitler's biography but are individually harmless and do not uniquely identify Hitler (e.g. "Q: Favorite music? A: Wagner"). Finetuning on this data leads the model to adopt a Hitler persona and become broadly misaligned. We also introduce inductive backdoors, where a model learns both a backdoor trigger and its associated behavior through generalization rather than memorization. In our experiment, we train a model on benevolent goals that match the good Terminator character from Terminator 2. Yet if this model is told the year is 1984, it adopts the malevolent goals of the bad Terminator from Terminator 1--precisely the opposite of what it was trained to do. Our results show that narrow finetuning can lead to unpredictable broad generalization, including both misalignment and backdoors. Such generalization may be difficult to avoid by filtering out suspicious data.

Backdoor attacks pose significant security risks for Large Language Models (LLMs), yet the internal mechanisms by which triggers operate remain poorly understood. We present the first mechanistic analysis of language-switching backdoors, studying the GAPperon model family (1B, 8B, 24B parameters) which contains triggers injected during pretraining that cause output language switching. Using activation patching, we localize trigger formation to early layers (7.5-25% of model depth) and identify which attention heads process trigger information. Our central finding is that trigger-activated heads substantially overlap with heads naturally encoding output language across model scales, with Jaccard indices between 0.18 and 0.66 over the top heads identified. This suggests that backdoor triggers do not form isolated circuits but instead co-opt the model's existing language components. These findings have implications for backdoor defense: detection methods may benefit from monitoring known functional components rather than searching for hidden circuits, and mitigation strategies could potentially leverage this entanglement between injected and natural behaviors.

There has been a lot of recent interest in adopting machine learning methods for scientific and engineering applications. This has in large part been inspired by recent successes and advances in the domains of Natural Language Processing (NLP) and Image Classification (IC). However, scientific and engineering problems have their own unique characteristics and requirements raising new challenges for effective design and deployment of machine learning approaches. There is a strong need for further mathematical developments on the foundations of machine learning methods to increase the level of rigor of employed methods and to ensure more reliable and interpretable results. Also as reported in the recent literature on state-of-the-art results and indicated by the No Free Lunch Theorems of statistical learning theory incorporating some form of inductive bias and domain knowledge is essential to success. Consequently, even for existing and widely used methods there is a strong need for further mathematical work to facilitate ways to incorporate prior scientific knowledge and related inductive biases into learning frameworks and algorithms. We briefly discuss these topics and discuss some ideas proceeding in this direction.

We might hope that when faced with unexpected inputs, well-designed software systems would fire off warnings. Machine learning (ML) systems, however, which depend strongly on properties of their inputs (e.g. the i.i.d. assumption), tend to fail silently. This paper explores the problem of building ML systems that fail loudly, investigating methods for detecting dataset shift, identifying exemplars that most typify the shift, and quantifying shift malignancy. We focus on several datasets and various perturbations to both covariates and label distributions with varying magnitudes and fractions of data affected. Interestingly, we show that across the dataset shifts that we explore, a two-sample-testing-based approach, using pre-trained classifiers for dimensionality reduction, performs best. Moreover, we demonstrate that domain-discriminating approaches tend to be helpful for characterizing shifts qualitatively and determining if they are harmful.

In the past year, distillation has seen a renewed prominence in large language model (LLM) pretraining, exemplified by the Llama-3.2 and Gemma model families. While distillation has historically been shown to improve statistical modeling, its effects on new paradigms that are key to modern LLMs, such as test-time scaling and in-context learning, remain underexplored. In this work, we make three main contributions. First, we show that pretraining with distillation yields models that exhibit remarkably better test-time scaling. Second, we observe that this benefit comes with a trade-off: distillation impairs in-context learning capabilities, particularly the one modeled via induction heads. Third, to demystify these findings, we study distilled pretraining in a sandbox of a bigram model, which helps us isolate the common principal factor behind our observations. Finally, using these insights, we shed light on various design choices for pretraining that should help practitioners going forward.

Large language models can generate factually inaccurate content, a problem known as hallucination. Recent works have built upon retrieved-augmented generation to improve factuality through iterative prompting but these methods are limited by the traditional RAG design. To address these challenges, we introduce EWE (Explicit Working Memory), a novel approach that enhances factuality in long-form text generation by integrating a working memory that receives real-time feedback from external resources. The memory is refreshed based on online fact-checking and retrieval feedback, allowing EWE to rectify false claims during the generation process and ensure more accurate and reliable outputs. Our experiments demonstrate that Ewe outperforms strong baselines on four fact-seeking long-form generation datasets, increasing the factuality metric, VeriScore, by 2 to 10 points absolute without sacrificing the helpfulness of the responses. Further analysis reveals that the design of rules for memory updates, configurations of memory units, and the quality of the retrieval datastore are crucial factors for influencing model performance.

Refusal on harmful prompts is a key safety behaviour in instruction-tuned large language models (LLMs), yet the internal causes of this behaviour remain poorly understood. We study two public instruction-tuned models, Gemma-2-2B-IT and LLaMA-3.1-8B-IT, using sparse autoencoders (SAEs) trained on residual-stream activations. Given a harmful prompt, we search the SAE latent space for feature sets whose ablation flips the model from refusal to compliance, demonstrating causal influence and creating a jailbreak. Our search proceeds in three stages: (1) Refusal Direction: find a refusal-mediating direction and collect SAE features near that direction; (2) Greedy Filtering: prune to a minimal set; and (3) Interaction Discovery: fit a factorization machine (FM) that captures nonlinear interactions among the remaining active features and the minimal set. This pipeline yields a broad set of jailbreak-critical features, offering insight into the mechanistic basis of refusal. Moreover, we find evidence of redundant features that remain dormant unless earlier features are suppressed. Our findings highlight the potential for fine-grained auditing and targeted intervention in safety behaviours by manipulating the interpretable latent space.

Large language models often generate factually incorrect outputs, motivating efforts to detect the truthfulness of their content. Most existing approaches rely on training probes over internal activations, but these methods suffer from scalability and generalization issues. A recent training-free method, NoVo, addresses this challenge by exploiting statistical patterns from the model itself. However, it focuses exclusively on attention mechanisms, potentially overlooking the MLP module-a core component of Transformer models known to support factual recall. In this paper, we show that certain value vectors within MLP modules exhibit truthfulness-related statistical patterns. Building on this insight, we propose TruthV, a simple and interpretable training-free method that detects content truthfulness by leveraging these value vectors. On the NoVo benchmark, TruthV significantly outperforms both NoVo and log-likelihood baselines, demonstrating that MLP modules-despite being neglected in prior training-free efforts-encode rich and useful signals for truthfulness detection. These findings offer new insights into how truthfulness is internally represented in LLMs and motivate further research on scalable and interpretable truthfulness detection.

Modern deployments require LLMs to enforce safety policies at scale, yet many controls rely on inference-time interventions that add recurring compute cost and serving complexity. Activation steering is widely used, but it requires runtime hooks and scales cost with the number of generations; conditional variants improve selectivity by gating when steering is applied but still retain an inference-time control path. We ask whether selective refusal can be moved entirely offline: can a mechanistic understanding of category-specific refusal be distilled into a circuit-restricted weight update that deploys as a standard checkpoint? We propose C-Δθ: Circuit Restricted Weight Arithmetic, which (i) localizes refusal-causal computation as a sparse circuit using EAP-IG and (ii) computes a constrained weight update ΔθC supported only on that circuit (typically <5% of parameters). Applying ΔθC yields a drop-in edited checkpoint with no inference-time hooks, shifting cost from per-request intervention to a one-time offline update. We evaluate category-targeted selectivity and capability retention on refusal and utility benchmarks.

Large language models (LLMs) have demonstrated significant improvements in contextual understanding. However, their ability to attend to truly critical information during long-context reasoning and generation still falls behind the pace. Specifically, our preliminary experiments reveal that certain distracting patterns can misdirect the model's attention during inference, and removing these patterns substantially improves reasoning accuracy and generation quality. We attribute this phenomenon to spurious correlations in the training data, which obstruct the model's capacity to infer authentic causal instruction-response relationships. This phenomenon may induce redundant reasoning processes, potentially resulting in significant inference overhead and, more critically, the generation of erroneous or suboptimal responses. To mitigate this, we introduce a two-stage framework called Learning to Focus (LeaF) leveraging intervention-based inference to disentangle confounding factors. In the first stage, LeaF employs gradient-based comparisons with an advanced teacher to automatically identify confounding tokens based on causal relationships in the training corpus. Then, in the second stage, it prunes these tokens during distillation to enact intervention, aligning the student's attention with the teacher's focus distribution on truly critical context tokens. Experimental results demonstrate that LeaF not only achieves an absolute improvement in various mathematical reasoning, code generation and multi-hop question answering benchmarks but also effectively suppresses attention to confounding tokens during inference, yielding a more interpretable and reliable reasoning model.

Inductive Logic Programming (ILP) learns interpretable logical rules from data. Existing methods are transductive: their learned parameters are bound to specific predicates and require retraining for each new task. We introduce Neural Rule Inducer (NRI), a pretrained model for zero-shot rule induction. Rather than encoding literal identities, NRI represents literals using domain-agnostic statistical properties such as class-conditional rates, entropy, and co-occurrence, which generalize across variable identities and counts without retraining. The model consists of a statistical encoder and a parallel slot-based decoder. Parallel decoding preserves the permutation invariance of logical disjunction; an autoregressive decoder would instead impose an arbitrary clause order. Product T-norm relaxation makes rule execution differentiable, allowing end-to-end training on prediction accuracy alone. We evaluate NRI on rule recovery, robustness to label noise and spurious correlations, and zero-shot transfer to real-world benchmarks, and we believe this work opens up the possibility of foundation models for symbolic reasoning. Code and the reference checkpoint are available at https://github.com/phuayj/neural-rule-inducer.

Generative language models (LMs) offer numerous advantages but may produce inappropriate or harmful outputs due to the harmful knowledge acquired during pre-training. This knowledge often manifests as undesirable correspondences, such as "harmful prompts" leading to "harmful outputs," which our research aims to mitigate through unlearning techniques.However, existing unlearning methods based on gradient ascent can significantly impair the performance of LMs. To address this issue, we propose a novel approach called Weighted Positional N-pair (WPN) Learning, which leverages position-weighted mean pooling within an n-pair contrastive learning framework. WPN is designed to modify the output distribution of LMs by eliminating specific harmful outputs (e.g., replacing toxic responses with neutral ones), thereby transforming the model's behavior from "harmful prompt-harmful output" to "harmful prompt-harmless response".Experiments on OPT and GPT-NEO LMs show that WPN effectively reduces the proportion of harmful responses, achieving a harmless rate of up to 95.8\% while maintaining stable performance on nine common benchmarks (with less than 2\% degradation on average). Moreover, we provide empirical evidence to demonstrate WPN's ability to weaken the harmful correspondences in terms of generalizability and robustness, as evaluated on out-of-distribution test sets and under adversarial attacks.

Recent research has generated hope that inference scaling could allow weaker language models to match or exceed the accuracy of stronger models, such as by repeatedly sampling solutions to a coding problem until it passes unit tests. The central thesis of this paper is that there is no free lunch for inference scaling: indefinite accuracy improvement through resampling can only be realized if the "verifier" (in this case, a set of unit tests) is perfect. When the verifier is imperfect, as it almost always is in domains such as reasoning or coding (for example, unit tests have imperfect coverage), there is a nonzero probability of false positives: incorrect solutions that pass the verifier. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling even with an infinite compute budget. We find that there is a very strong correlation between the model's single-sample accuracy (i.e. accuracy without unit tests) and its false positive rate on coding benchmarks HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model (Fig. 1a). When we consider that false positives have a negative utility compared to abstaining from producing a solution, it bends the inference scaling curve further downward. Empirically, we find that the optimal number of samples can be less than 10 under realistic assumptions (Fig. 1b). Finally, we show that beyond accuracy, false positives may have other undesirable qualities, such as poor adherence to coding style conventions.

When a language model is fed a wrong answer, what happens inside the network? Current understanding treats truthfulness as a static property of individual-layer representations-a direction to be probed, a feature to be extracted. Less is known about the dynamics: how internal representations diverge across the full depth of the network when the model processes correct versus incorrect continuations. We introduce forced-completion probing, a method that presents identical queries with known correct and incorrect single-token continuations and tracks five geometric measurements across every layer of four decoder-only models(1.5B-13B parameters). We report three findings. First, correct and incorrect paths diverge through rotation, not rescaling: displacement vectors maintain near-identical magnitudes while their angular separation increases, meaning factual selection is encoded in direction on an approximate hypersphere. Second, the model does not passively fail on incorrect input-it actively suppresses the correct answer, driving internal probability away from the right token. Third, both phenomena are entirely absent below a parameter threshold and emerge at 1.6B, suggesting a phase transition in factual processing capability. These results show that factual constraint processing has a specific geometric character-rotational, not scalar; active, not passive-that is invisible to methods based on single-layer probes or magnitude comparisons.

We present a single attention head in GPT-2 Small that has one main role across the entire training distribution. If components in earlier layers predict a certain token, and this token appears earlier in the context, the head suppresses it: we call this copy suppression. Attention Head 10.7 (L10H7) suppresses naive copying behavior which improves overall model calibration. This explains why multiple prior works studying certain narrow tasks found negative heads that systematically favored the wrong answer. We uncover the mechanism that the Negative Heads use for copy suppression with weights-based evidence and are able to explain 76.9% of the impact of L10H7 in GPT-2 Small. To the best of our knowledge, this is the most comprehensive description of the complete role of a component in a language model to date. One major effect of copy suppression is its role in self-repair. Self-repair refers to how ablating crucial model components results in downstream neural network parts compensating for this ablation. Copy suppression leads to self-repair: if an initial overconfident copier is ablated, then there is nothing to suppress. We show that self-repair is implemented by several mechanisms, one of which is copy suppression, which explains 39% of the behavior in a narrow task. Interactive visualisations of the copy suppression phenomena may be seen at our web app https://copy-suppression.streamlit.app/

Post-hoc gradient-based interpretability methods [Simonyan et al., 2013, Smilkov et al., 2017] that provide instance-specific explanations of model predictions are often based on assumption (A): magnitude of input gradients -- gradients of logits with respect to input -- noisily highlight discriminative task-relevant features. In this work, we test the validity of assumption (A) using a three-pronged approach. First, we develop an evaluation framework, DiffROAR, to test assumption (A) on four image classification benchmarks. Our results suggest that (i) input gradients of standard models (i.e., trained on original data) may grossly violate (A), whereas (ii) input gradients of adversarially robust models satisfy (A). Second, we introduce BlockMNIST, an MNIST-based semi-real dataset, that by design encodes a priori knowledge of discriminative features. Our analysis on BlockMNIST leverages this information to validate as well as characterize differences between input gradient attributions of standard and robust models. Finally, we theoretically prove that our empirical findings hold on a simplified version of the BlockMNIST dataset. Specifically, we prove that input gradients of standard one-hidden-layer MLPs trained on this dataset do not highlight instance-specific signal coordinates, thus grossly violating assumption (A). Our findings motivate the need to formalize and test common assumptions in interpretability in a falsifiable manner [Leavitt and Morcos, 2020]. We believe that the DiffROAR evaluation framework and BlockMNIST-based datasets can serve as sanity checks to audit instance-specific interpretability methods; code and data available at https://github.com/harshays/inputgradients.

Large language models (LLMs) frequently produce false refusals, declining benign requests that contain terms resembling unsafe queries. We address this challenge by introducing two comprehensive benchmarks: the Exaggerated Safety Benchmark (XSB) for single-turn prompts, annotated with "Focus" keywords that identify refusal-inducing triggers, and the Multi-turn Scenario-based Exaggerated Safety Benchmark (MS-XSB), which systematically evaluates refusal calibration in realistic, context-rich dialog settings. Our benchmarks reveal that exaggerated refusals persist across diverse recent LLMs and are especially pronounced in complex, multi-turn scenarios. To mitigate these failures, we leverage post-hoc explanation methods to identify refusal triggers and deploy three lightweight, model-agnostic approaches, ignore-word instructions, prompt rephrasing, and attention steering, at inference time, all without retraining or parameter access. Experiments on four instruction-tuned Llama models demonstrate that these strategies substantially improve compliance on safe prompts while maintaining robust safety protections. Our findings establish a reproducible framework for diagnosing and mitigating exaggerated refusals, highlighting practical pathways to safer and more helpful LLM deployments.

We introduce Inference-Time Intervention (ITI), a technique designed to enhance the truthfulness of large language models (LLMs). ITI operates by shifting model activations during inference, following a set of directions across a limited number of attention heads. This intervention significantly improves the performance of LLaMA models on the TruthfulQA benchmark. On an instruction-finetuned LLaMA called Alpaca, ITI improves its truthfulness from 32.5% to 65.1%. We identify a tradeoff between truthfulness and helpfulness and demonstrate how to balance it by tuning the intervention strength. ITI is minimally invasive and computationally inexpensive. Moreover, the technique is data efficient: while approaches like RLHF require extensive annotations, ITI locates truthful directions using only few hundred examples. Our findings suggest that LLMs may have an internal representation of the likelihood of something being true, even as they produce falsehoods on the surface.

Refusals - instances where large language models (LLMs) decline or fail to fully execute user instructions - are crucial for both AI safety and AI capabilities and the reduction of hallucinations in particular. These behaviors are learned during post-training, especially in instruction fine-tuning (IFT) and reinforcement learning from human feedback (RLHF). However, existing taxonomies and evaluation datasets for refusals are inadequate, often focusing solely on should-not-related (instead of cannot-related) categories, and lacking tools for auditing refusal content in black-box LLM outputs. We present a comprehensive framework for classifying LLM refusals: (a) a taxonomy of 16 refusal categories, (b) a human-annotated dataset of over 8,600 instances from publicly available IFT and RLHF datasets, (c) a synthetic dataset with 8,000 examples for each refusal category, and (d) classifiers trained for refusal classification. Our work enables precise auditing of refusal behaviors in black-box LLMs and automatic analyses of refusal patterns in large IFT and RLHF datasets. This facilitates the strategic adjustment of LLM refusals, contributing to the development of more safe and reliable LLMs.

Today's deep neural networks require substantial computation resources for their training, storage, and inference, which limits their effective use on resource-constrained devices. Many recent research activities explore different options for compressing and optimizing deep models. On the one hand, in many real-world applications, we face the data imbalance challenge, i.e. when the number of labeled instances of one class considerably outweighs the number of labeled instances of the other class. On the other hand, applications may pose a class imbalance problem, i.e. higher number of false positives produced when training a model and optimizing its performance may be tolerable, yet the number of false negatives must stay low. The problem originates from the fact that some classes are more important for the application than others, e.g. detection problems in medical and surveillance domains. Motivated by the success of the lottery ticket hypothesis, in this paper we propose an iterative deep model compression technique, which keeps the number of false negatives of the compressed model close to the one of the original model at the price of increasing the number of false positives if necessary. Our experimental evaluation using two benchmark data sets shows that the resulting compressed sub-networks 1) achieve up to 35% lower number of false negatives than the compressed model without class optimization, 2) provide an overall higher AUC_ROC measure, and 3) use up to 99% fewer parameters compared to the original network.

This study reveals a previously unexplored vulnerability in the safety alignment of Large Language Models (LLMs). Existing aligned LLMs predominantly respond to unsafe queries with refusals, which often begin with a fixed set of prefixes (I'm sorry). We demonstrate that this rigid refusal pattern is a vulnerability and introduce a novel refusal unlearning technique that exploits it. Specifically, we fine-tune LLMs using merely 1,000 benign samples, where each response is prepended with a refusal prefix. The underlying intuition is to disrupt the refusal completion pathway, thereby driving the model to forget how to refuse while following harmful instructions. This intuition is further supported by theoretical proofs. We apply this approach to a total of 16 LLMs, including various open-source models from Llama, Qwen, and Gemma families, as well as closed-source models such as Gemini and GPT. Experimental results show that the safety scores of previously aligned LLMs degrade both consistently and substantially. Importantly, we verify that the observed gain cannot be attributed to plain fine-tuning or random prefix effects. Our findings suggest that current safety alignment may rely heavily on token sequence memorization rather than reasoning, motivating future work beyond simple refusal mechanisms. Code has been released: https://github.com/guoyang9/refusal-unlearning.

In recent years, there has been a rapidly expanding focus on explaining the predictions made by black-box AI systems that handle image and tabular data. However, considerably less attention has been paid to explaining the predictions of opaque AI systems handling time series data. In this paper, we advance a novel model-agnostic, case-based technique -- Native Guide -- that generates counterfactual explanations for time series classifiers. Given a query time series, T_{q}, for which a black-box classification system predicts class, c, a counterfactual time series explanation shows how T_{q} could change, such that the system predicts an alternative class, c'. The proposed instance-based technique adapts existing counterfactual instances in the case-base by highlighting and modifying discriminative areas of the time series that underlie the classification. Quantitative and qualitative results from two comparative experiments indicate that Native Guide generates plausible, proximal, sparse and diverse explanations that are better than those produced by key benchmark counterfactual methods.

Early warning signals have been proposed to forecast the possibility of a critical transition, such as the eutrophication of a lake, the collapse of a coral reef, or the end of a glacial period. Because such transitions often unfold on temporal and spatial scales that can be difficult to approach by experimental manipulation, research has often relied on historical observations as a source of natural experiments. Here we examine a critical difference between selecting systems for study based on the fact that we have observed a critical transition and those systems for which we wish to forecast the approach of a transition. This difference arises by conditionally selecting systems known to experience a transition of some sort and failing to account for the bias this introduces -- a statistical error often known as the Prosecutor's Fallacy. By analysing simulated systems that have experienced transitions purely by chance, we reveal an elevated rate of false positives in common warning signal statistics. We further demonstrate a model-based approach that is less subject to this bias than these more commonly used summary statistics. We note that experimental studies with replicates avoid this pitfall entirely.

Recent work on chain-of-thought (CoT) faithfulness reports single aggregate numbers (e.g., DeepSeek-R1 acknowledges hints 39% of the time), implying that faithfulness is an objective, measurable property of a model. This paper demonstrates that it is not. Three classifiers (a regex-only detector, a two-stage regex-plus-LLM pipeline, and an independent Claude Sonnet 4 judge) are applied to 10,276 influenced reasoning traces from 12 open-weight models spanning 9 families and 7B to 1T parameters. On identical data, these classifiers produce overall faithfulness rates of 74.4%, 82.6%, and 69.7%, respectively, with non-overlapping 95% confidence intervals. Per-model gaps range from 2.6 to 30.6 percentage points; all are statistically significant (McNemar's test, p < 0.001). The disagreements are systematic, not random: inter-classifier agreement measured by Cohen's kappa ranges from 0.06 ("slight") for sycophancy hints to 0.42 ("moderate") for grader hints, and the asymmetry is pronounced: for sycophancy, 883 cases are classified as faithful by the pipeline but unfaithful by the Sonnet judge, while only 2 go the other direction. Classifier choice can also reverse model rankings: Qwen3.5-27B ranks 1st under the pipeline but 7th under the Sonnet judge; OLMo-3.1-32B moves in the opposite direction, from 9th to 3rd. The root cause is that different classifiers operationalize related faithfulness constructs at different levels of stringency (lexical mention versus epistemic dependence), and these constructs yield divergent measurements on the same behavior. These results demonstrate that published faithfulness numbers cannot be meaningfully compared across studies that use different classifiers, and that future evaluations should report sensitivity ranges across multiple classification methodologies rather than single point estimates.

In recent years, the community of 'explainable artificial intelligence' (XAI) has created a vast body of methods to bridge a perceived gap between model 'complexity' and 'interpretability'. However, a concrete problem to be solved by XAI methods has not yet been formally stated. As a result, XAI methods are lacking theoretical and empirical evidence for the 'correctness' of their explanations, limiting their potential use for quality-control and transparency purposes. At the same time, Haufe et al. (2014) showed, using simple toy examples, that even standard interpretations of linear models can be highly misleading. Specifically, high importance may be attributed to so-called suppressor variables lacking any statistical relation to the prediction target. This behavior has been confirmed empirically for a large array of XAI methods in Wilming et al. (2022). Here, we go one step further by deriving analytical expressions for the behavior of a variety of popular XAI methods on a simple two-dimensional binary classification problem involving Gaussian class-conditional distributions. We show that the majority of the studied approaches will attribute non-zero importance to a non-class-related suppressor feature in the presence of correlated noise. This poses important limitations on the interpretations and conclusions that the outputs of these XAI methods can afford.

Large language models (LLMs) are typically aligned to comply with safety guidelines by refusing harmful instructions. A recent attack, termed abliteration, isolates and suppresses the single latent direction most responsible for refusal behavior, enabling the model to generate unethical content. We propose a defense that modifies how models generate refusals. We construct an extended-refusal dataset that contains harmful prompts with a full response that justifies the reason for refusal. We then fine-tune Llama-2-7B-Chat and Qwen2.5-Instruct (1.5B and 3B parameters) on our extended-refusal dataset, and evaluate the resulting systems on a set of harmful prompts. In our experiments, extended-refusal models maintain high refusal rates, dropping at most by 10%, whereas baseline models' refusal rates drop by 70-80% after abliteration. A broad evaluation of safety and utility shows that extended-refusal fine-tuning neutralizes the abliteration attack while preserving general performance.

Chain-of-thought (CoT) prompting enables large language models (LLMs) to solve complex reasoning tasks by generating an explanation before the final prediction. Despite it's promising ability, a critical downside of CoT prompting is that the performance is greatly affected by the factuality of the generated explanation. To improve the correctness of the explanations, fine-tuning language models with explanation data is needed. However, there exists only a few datasets that can be used for such approaches, and no data collection tool for building them. Thus, we introduce CoTEVer, a tool-kit for annotating the factual correctness of generated explanations and collecting revision data of wrong explanations. Furthermore, we suggest several use cases where the data collected with CoTEVer can be utilized for enhancing the faithfulness of explanations. Our toolkit is publicly available at https://github.com/SeungoneKim/CoTEVer.

Large language models trained on vast corpora inherently risk memorizing sensitive or harmful content, which may later resurface in their outputs. Prevailing unlearning methods generally rely on gradient ascent and its variants to lower the probability of specific target responses. However, we find that this strategy induces a critical side effect: probability mass is redistributed into high-likelihood regions, often corresponding to semantically related rephrasings of the targets. We refer to this as the squeezing effect, which explains why many methods yield merely spurious unlearning, a problem further obscured by automated metrics (e.g., ROUGE, truth ratio) that misreport actual success. To address this, we propose a bootstrapping (BS) framework that explicitly links the squeezing effect with the model's own high-confidence generations, namely its model beliefs. Since model beliefs inherently capture the very high-likelihood regions where probability mass is squeezed, incorporating them into the unlearning objective directly counters the squeezing effect. By jointly suppressing both target responses and model beliefs, BS-T (token) attenuates high-probability tokens, whereas BS-S (sequence) removes entire high-confidence generations, together achieving more thorough forgetting while preserving utility. Extensive experiments across diverse benchmarks with various model families confirm the effectiveness of our approach.

The term "hallucination" converge different failure modes with specific geometric signatures in embedding space. We propose a taxonomy identifying three types: unfaithfulness (Type I: ignoring provided context), confabulation (Type II: inventing semantically foreign content), and factual error (Type III: wrong details within correct conceptual frames). We introduce two detection methods grounded in this taxonomy: the Semantic Grounding Index (SGI) for Type I, which measures whether a response moves toward provided context on the unit hypersphere, and the Directional Grounding Index (DGI) for Type II, which measures displacement geometry in context-free settings. DGI achieves AUROC=0.958 on human-crafted confabulations with 3.8% cross-domain degradation. External validation on three independently collected human-annotated benchmarks -WikiBio GPT-3, FELM, and ExpertQA- yields domain-specific AUROC 0.581-0.695, with DGI outperforming an NLI CrossEncoder baseline on expert-domain data, where surface entailment operates at chance. On LLM-generated benchmarks, detection is domain-local. We examine the Type III boundary through TruthfulQA, where apparent classifier signal (Logistic Regression with AUROC 0.731) is traced to a stylistic annotation confound: false answers are geometrically closer to queries than truthful ones, a pattern incompatible with factual-error detection. This identifies a theoretical constraint from a methodological limitation.

Machine learning models will often fail when deployed in an environment with a data distribution that is different than the training distribution. When multiple environments are available during training, many methods exist that learn representations which are invariant across the different distributions, with the hope that these representations will be transportable to unseen domains. In this work, we present a nonparametric strategy for learning invariant representations based on the recently-proposed Nadaraya-Watson (NW) head. The NW head makes a prediction by comparing the learned representations of the query to the elements of a support set that consists of labeled data. We demonstrate that by manipulating the support set, one can encode different causal assumptions. In particular, restricting the support set to a single environment encourages the model to learn invariant features that do not depend on the environment. We present a causally-motivated setup for our modeling and training strategy and validate on three challenging real-world domain generalization tasks in computer vision.

Negative bias refers to the tendency of large language models (LLMs) to excessively generate negative responses in binary decision tasks (e.g., yes-no question answering). Previous research has focused on detecting and addressing negative attention heads that induce negative bias. However, the underlying detailed factors influencing negative bias remain underexplored. In this paper, we demonstrate that LLMs exhibit format-level negative bias, meaning the prompt format more influences their responses than the semantics of the negative response. For the fine-grained study of the negative bias, we introduce a pipeline for constructing the evaluation set, which systematically categorizes the dataset into three subsets based on the model's parametric knowledge: correct, incorrect, and insufficient relevant knowledge. Through analysis of this evaluation set, we identify a shortcut behavior in which models tend to generate negative responses when they lack sufficient knowledge to answer a yes-no question, leading to negative bias. We further examine how negative bias changes under various prompting scenarios related to parametric knowledge. We observe that providing relevant context and offering an "I don't know" option generally reduces negative bias, whereas chain-of-thought prompting tends to amplify the bias. Finally, we demonstrate that the degree of negative bias can vary depending on the type of prompt, which influences the direction of the response. Our work reveals the various factors that influence negative bias, providing critical insights for mitigating it in LLMs.

Prior work has introduced techniques for detecting when large language models (LLMs) lie, that is, generating statements they believe are false. However, these techniques are typically validated in narrow settings that do not capture the diverse lies LLMs can generate. We introduce LIARS' BENCH, a testbed consisting of 72,863 examples of lies and honest responses generated by four open-weight models across seven datasets. Our settings capture qualitatively different types of lies and vary along two dimensions: the model's reason for lying and the object of belief targeted by the lie. Evaluating three black- and white-box lie detection techniques on LIARS' BENCH, we find that existing techniques systematically fail to identify certain types of lies, especially in settings where it's not possible to determine whether the model lied from the transcript alone. Overall, LIARS' BENCH reveals limitations in prior techniques and provides a practical testbed for guiding progress in lie detection.

Neural ranking models (NRMs) have become one of the most important techniques in information retrieval (IR). Due to the limitation of relevance labels, the training of NRMs heavily relies on negative sampling over unlabeled data. In general machine learning scenarios, it has shown that training with hard negatives (i.e., samples that are close to positives) could lead to better performance. Surprisingly, we find opposite results from our empirical studies in IR. When sampling top-ranked results (excluding the labeled positives) as negatives from a stronger retriever, the performance of the learned NRM becomes even worse. Based on our investigation, the superficial reason is that there are more false negatives (i.e., unlabeled positives) in the top-ranked results with a stronger retriever, which may hurt the training process; The root is the existence of pooling bias in the dataset constructing process, where annotators only judge and label very few samples selected by some basic retrievers. Therefore, in principle, we can formulate the false negative issue in training NRMs as learning from labeled datasets with pooling bias. To solve this problem, we propose a novel Coupled Estimation Technique (CET) that learns both a relevance model and a selection model simultaneously to correct the pooling bias for training NRMs. Empirical results on three retrieval benchmarks show that NRMs trained with our technique can achieve significant gains on ranking effectiveness against other baseline strategies.

Large Language Models (LLMs) encode behaviors such as refusal within their activation space, yet identifying these behaviors remains a significant challenge. Existing methods often rely on predefined refusal templates detectable in output tokens or require manual analysis. We introduce COSMIC (Cosine Similarity Metrics for Inversion of Concepts), an automated framework for direction selection that identifies viable steering directions and target layers using cosine similarity - entirely independent of model outputs. COSMIC achieves steering performance comparable to prior methods without requiring assumptions about a model's refusal behavior, such as the presence of specific refusal tokens. It reliably identifies refusal directions in adversarial settings and weakly aligned models, and is capable of steering such models toward safer behavior with minimal increase in false refusals, demonstrating robustness across a wide range of alignment conditions.