PAST-Bench - Self-Improvement Gain: leaderboard
Part of PAST-Bench (Xue et al., 2026): the gain from retained experience, the overall task score with persistence switched on minus a matched run with it switched off, in points on a scale from minus 100 to 100. Gaps are averaged per task family, then per capability; higher is better.
Metric: Gain from retained experience (points on a -100 to 100 scale; overall task score with persistence on minus the matched persistence-off run with the same prompt, grader, tools and seed, averaged per family, then per capability, then over the four capabilities; 26 scenarios and 204 episodes run as ordered sequences of fresh-session tasks; each evaluation episode scores completion (audit-data checks, or a MiniMax-M2.7 rubric judge for open-ended families), tool-error recovery and a safety gate; family-balanced means over three trials; Hermes agent framework). Source: arxiv.org. Saturation forecast: Around June 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Medium) | 24 | #123 (GPT-5.4) |
| 2 | Claude Sonnet 4.6 | 20 | #161 |
| 3 | GLM-5.1 | 20 | #179 |
| 4 | Claude Opus 4.6 (Medium) | 19 | #102 (Claude Opus 4.6) |
| 5 | Kimi K2.6 (Thinking) | 17 | #178 (Kimi K2.6) |
| 6 | DeepSeek V4 Pro (High) | 17 | #188 (DeepSeek V4 Pro) |
| 7 | MiniMax-M2.7 | 13 | #518 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=past-bench-self-improvement-gain · How It Works · Data refreshed daily, snapshot 2026-11-13.