HINTBench - Fine Risk-Step Localization: leaderboard
Metric: Localization F1 (%) for fine risk-step localization (twelve risk subtypes described; a predicted risky step matches a gold step within plus or minus 3 steps; type ignored) over the 536 synthetic HINTBench agent trajectories (400 risky, 136 safe, 24 steps on average, unanimously accepted by three human verifiers) audited zero-shot by a general LLM for intrinsic, non-attack risk under a five-constraint taxonomy (goal, capability, factual, procedural, state constraints), each model at its stated reasoning setting; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Sonnet 4.6 (Non-reasoning) | 59.05 |
| 2 | Claude Opus 4.6 (Non-reasoning) | 58.67 |
| 3 | DeepSeek R1 | 57.32 |
| 4 | Qwen 3 235B A22B | 56.9 |
| 5 | GLM-5 | 56.3 |
| 6 | ERNIE 5.0 | 56.16 |
| 7 | Kimi K2.5 | 56.15 |
| 8 | GPT-5.5 (Low) | 54.17 |
| 9 | DeepSeek V3.2 | 54.07 |
| 10 | MiniMax-M2.5 | 53.82 |
| 11 | GPT-5.5 (High) | 53.13 |
| 12 | GPT-5.5 (Medium) | 52.33 |
| 13 | Qwen 3 32B | 51.31 |
| 14 | Qwen 3 14B | 50.65 |
| 15 | Qwen 3 8B | 29.71 |
Interactive version: theaggregate.ai/benchmark?slug=hintbench-fine-risk-step-localization · How It Works · Data refreshed daily, snapshot 2026-10-07.