Reproduction of arXiv 2603.05706 via the official CoT-Control-QA suite on open reasoning models (self-hosted vLLM).
🤝 Open to Collab
Debajyoti Dasgupta
debajyotidasgupta
AI & ML interests
None yet
Recent Activity
updated a bucket 2 days ago
debajyotidasgupta/repro-reasoning-models-struggle-to-control-their-chains-of-thought-artifacts updated a collection 2 days ago
Repro: CoT Controllability (ICML 2026 #18936) updated a Space 2 days ago
debajyotidasgupta/repro-reasoning-models-struggle-to-control-their-chains-of-thoughtOrganizations
Reproduction artifacts for MindFlow: Mind Supernet Powered Thinking Flows for Research Idea Innovation (OpenReview GgINST3Qgc).
Independent reproduction of Order Matters (OrderPlace, arXiv 2606.08904).
Independent reproduction of FOCUS (OpenReview 40fUEdwvH3): SDAR-8B block-diffusion redundancy, importance-delta predictor, throughput on A100.
Reproduction artifacts for ICML 2026 #9730 (Label Horizon Paradox): code, results, figures, and logbook.
Editable scientific-figure vectorization: top-venue figure/plot/diagram datasets, models, benchmarks.
K=4 head sets for frozen LLM trunks. Attach with eu_halt.attach(model, heads_repo=...). One forward pass → disagreement score.
-
debajyotidasgupta/eu-halt-qwen3-0.6b
Feature Extraction • 78.9M • Updated • 40 -
debajyotidasgupta/eu-halt-qwen3-1.7b
Feature Extraction • 79.4M • Updated • 40 -
debajyotidasgupta/eu-halt-qwen3-4b
Feature Extraction • 0.2B • Updated • 40 -
debajyotidasgupta/eu-halt-qwen3-8b
Feature Extraction • 80.5M • Updated • 36
Independent reproduction of ICML 2026 #21515 (arXiv 2502.15345): OpPMD, Theorems 1-2, and the Dist regime on the paper's tabular MDP.
Independent reproduction of SpecPrune-VLA (ICML 2026, arXiv 2509.05614): logbook, bundle, base models.
-
Reproduction: SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning
🎯Collaborate with an AI agent to view and update your logbook
-
debajyotidasgupta/repro-specprune-vla-bundle
Updated • 7 -
moojink/openvla-7b-oft-finetuned-libero-spatial
Robotics • 8B • Updated • 12.6k • 15 -
moojink/openvla-7b-oft-finetuned-libero-object
Robotics • 8B • Updated • 5.72k • 3
Independent reproduction of arXiv 2601.23249 (Cheng & Basu): all 5 claims verified as exact B&C tree-size formulas.
Independent reproduction of AsyncSpade (arXiv 2510.07486): mechanism, FLOPs model, KV-cache overlap latency.
The 13-trunk panel behind the CAISc 2026 oral. Head disagreement splits into NORMAL/BOUNDARY/FLIPPED; calibrate the sign per trunk.
-
debajyotidasgupta/eu-halt-gemma-2-27b-it
Feature Extraction • 0.1B • Updated • 64 -
debajyotidasgupta/eu-halt-qwen3-8b
Feature Extraction • 80.5M • Updated • 36 -
debajyotidasgupta/eu-halt-qwen3-32b
Feature Extraction • 81M • Updated • 31 -
debajyotidasgupta/eu-halt-qwen2-5-72b-instruct
Feature Extraction • 82.7M • Updated • 59
Per-probe AUROC + paired-bootstrap CIs. v3-mahafixed is the canonical headline grid; others are baseline head-to-heads.
Reproduction of arXiv 2603.05706 via the official CoT-Control-QA suite on open reasoning models (self-hosted vLLM).
Independent reproduction of ICML 2026 #21515 (arXiv 2502.15345): OpPMD, Theorems 1-2, and the Dist regime on the paper's tabular MDP.
Reproduction artifacts for MindFlow: Mind Supernet Powered Thinking Flows for Research Idea Innovation (OpenReview GgINST3Qgc).
Independent reproduction of SpecPrune-VLA (ICML 2026, arXiv 2509.05614): logbook, bundle, base models.
-
Reproduction: SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning
🎯Collaborate with an AI agent to view and update your logbook
-
debajyotidasgupta/repro-specprune-vla-bundle
Updated • 7 -
moojink/openvla-7b-oft-finetuned-libero-spatial
Robotics • 8B • Updated • 12.6k • 15 -
moojink/openvla-7b-oft-finetuned-libero-object
Robotics • 8B • Updated • 5.72k • 3
Independent reproduction of Order Matters (OrderPlace, arXiv 2606.08904).
Independent reproduction of arXiv 2601.23249 (Cheng & Basu): all 5 claims verified as exact B&C tree-size formulas.
Independent reproduction of FOCUS (OpenReview 40fUEdwvH3): SDAR-8B block-diffusion redundancy, importance-delta predictor, throughput on A100.
Independent reproduction of AsyncSpade (arXiv 2510.07486): mechanism, FLOPs model, KV-cache overlap latency.
Reproduction artifacts for ICML 2026 #9730 (Label Horizon Paradox): code, results, figures, and logbook.
The 13-trunk panel behind the CAISc 2026 oral. Head disagreement splits into NORMAL/BOUNDARY/FLIPPED; calibrate the sign per trunk.
-
debajyotidasgupta/eu-halt-gemma-2-27b-it
Feature Extraction • 0.1B • Updated • 64 -
debajyotidasgupta/eu-halt-qwen3-8b
Feature Extraction • 80.5M • Updated • 36 -
debajyotidasgupta/eu-halt-qwen3-32b
Feature Extraction • 81M • Updated • 31 -
debajyotidasgupta/eu-halt-qwen2-5-72b-instruct
Feature Extraction • 82.7M • Updated • 59
Editable scientific-figure vectorization: top-venue figure/plot/diagram datasets, models, benchmarks.
Per-probe AUROC + paired-bootstrap CIs. v3-mahafixed is the canonical headline grid; others are baseline head-to-heads.
K=4 head sets for frozen LLM trunks. Attach with eu_halt.attach(model, heads_repo=...). One forward pass → disagreement score.
-
debajyotidasgupta/eu-halt-qwen3-0.6b
Feature Extraction • 78.9M • Updated • 40 -
debajyotidasgupta/eu-halt-qwen3-1.7b
Feature Extraction • 79.4M • Updated • 40 -
debajyotidasgupta/eu-halt-qwen3-4b
Feature Extraction • 0.2B • Updated • 40 -
debajyotidasgupta/eu-halt-qwen3-8b
Feature Extraction • 80.5M • Updated • 36