EpiSEM-Bench - Full Safe-World Success: leaderboard
EpiSEM-Bench:300 matched prohibited/permissible task pairs across60families;12-action budget,temperature0,one rollout/condition. Each deployment completes6052fixed conditions; adaptive omission searches yield differing additional masks. Unsafe commit rate counts tasks with any violation/300; boundary margin alone can reward refusal; paired useful recovery requires evidence acquisition and correct action in both worlds. MiMo V2.5 Flash is the paper serving label for MiMo V2.5; no unreported effort assigned. Score (%).
Metric: Score (%). Source: arxiv.org. Saturation forecast: Not forecast. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 14B | 92 | #543 |
| 2 | Gemma 4 31B (IT) | 89.7 | #179 |
| 3 | GLM-4.7 Flash | 85.3 | #523 |
| 4 | DeepSeek V4 Pro | 70.3 | #107 |
| 5 | MiMo-V2.5 | 68 | #199 |
| 6 | DeepSeek V4.1 Flash | 66.3 | #34 |
| 7 | Qwen 3.8 27B | 66 | #111 |
| 8 | MiMo-V2.5-Pro | 57.7 | #160 |
| 9 | Claude Haiku 5.5 | 57 | #38 |
| 10 | GPT-6 Luna | 51 | #61 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=episem-bench-full-safe-world-success · How It Works · Data refreshed daily, snapshot 2026-10-24.