README.md

---
license: mit
tags:
- code
- evaluation
- benchmark
- coding-agents
- reliability
- security
pipeline_tag: text-generation
---

# CodeBench

**CodeBench** is an evaluation framework for measuring the *reliable* code
generation ability of AI coding agents, beyond inflated pass@k scores.

## Key Findings

| Finding | Result |
|---|---|
| **H1 — Category error** | Standard pass@k treats test-case counts as independent sample counts, producing different scores for agents with identical true reliability |
| **H2 — Score inflation** | Current pass@5 ≈ 0.96–0.97 collapses to reliability@5 ≈ 0.00–0.12 when the category error is corrected |
| **H3 — Proxy validity** | Single-rollout proxy score has low Spearman correlation with reliability@k; ≥5 rollouts needed for reliable ranking |
| **H4 — Security gap** | Security-adjusted reliability@5 is meaningfully lower than reliability@5; leaderboard rank flips when security is accounted for |

## Metrics

- **`reliability@k`** — correct operationalization of Chen et al. (2021) pass@k,
  using per-(task, agent) rollout counts and binary execution success
- **`security_adjusted_reliability@k`** — reliability@k counting only rollouts
  that are both correct *and* produce code with no insecure patterns (eval,
  exec, os.system, yaml.load without Loader, pickle.loads)

## Agents Evaluated

| Agent | Provider | Model |
|---|---|---|
| `anote-code` | Anthropic | claude-sonnet-4-6 (Anote system prompt) |
| `claude-code` | Anthropic | claude-sonnet-4-6 |
| `codex` | OpenAI | gpt-4o |

## Figures

### Figure 1 — Baseline: pass@1 vs current pass@5
![fig1](figures/fig1_baseline.png)

### Figure 2 — H1: Category-error proof
![fig2](figures/fig2_h1_proof.png)

### Figure 3 — H2: Score inflation magnitude
![fig3](figures/fig3_h2_comparison.png)

### Figure 4 — H3: Proxy vs reliability@k correlation
![fig4](figures/fig4_h3_correlation.png)

### Figure 5 — H4: Security-adjusted reliability leaderboard
![fig5](figures/fig5_h4_security_leaderboard.png)

## Datasets

- [anote-ai/codebench-tasks](https://huggingface.co/datasets/anote-ai/codebench-tasks) — benchmark task definitions
- [anote-ai/codebench-results](https://huggingface.co/datasets/anote-ai/codebench-results) — experiment results (h4_security + swebench_smoke)

## Citation

```
@misc{codebench2026,
  title  = {CodeBench: Measuring Reliable Code Generation},
  author = {Anote AI},
  year   = {2026}
}
```