PAST-Bench - Self-Improvement Gain: leaderboard

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 January 2028. 7 models tracked.

Top models

#ModelScore
1GPT-5.4 (Medium)24
2Claude Sonnet 4.620
3GLM-5.120
4Claude Opus 4.6 (Medium)19
5Kimi K2.6 (Thinking)17
6DeepSeek V4 Pro (High)17
7MiniMax-M2.713

Interactive version: theaggregate.ai/benchmark?slug=past-bench-self-improvement-gain · How It Works · Data refreshed daily, snapshot 2026-09-29.