PAST-Bench - Update: leaderboard
Part of PAST-Bench (Xue et al., 2026): task score in percent on the update task families, where outdated state must be replaced, with persistence of earlier episodes switched on, in the Hermes agent framework. Scores combine task completion, tool-error recovery and a safety gate; higher is better.
Metric: Capability task score (%; update families, where outdated state must be replaced; persistence on: the agent may reuse memory records, skills, profile state and session history produced by earlier episodes of the family; 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 February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5.1 | 90 | #179 |
| 2 | GPT-5.4 (Medium) | 80 | #123 (GPT-5.4) |
| 3 | Kimi K2.6 (Thinking) | 78 | #178 (Kimi K2.6) |
| 4 | Claude Opus 4.6 (Medium) | 78 | #102 (Claude Opus 4.6) |
| 5 | Claude Sonnet 4.6 | 77 | #161 |
| 6 | DeepSeek V4 Pro (High) | 69 | #188 (DeepSeek V4 Pro) |
| 7 | MiniMax-M2.7 | 62 | #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-update · How It Works · Data refreshed daily, snapshot 2026-11-13.