AutoRLBench 40 Issue / 20 Solution - Issue: leaderboard
Issue-diagnosis split (40 tasks) of Thinking-Space AutoRLBench 40-issue / 20-solution edition measuring whether agents isolate and verify root causes of RL training failures; pass rate (%).
Metric: Root cause found, 40 failed training runs (%). Source: thinking-space.github.io. Saturation forecast: Around July 2028. 9 models tracked. Updated 19 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | GPT-6 Astra | Max | 30 | #10 |
| 2 | GPT-6.1 Sol | Max | 27.5 | #15 |
| 3 | GLM-5.3 | Max | 20 | #73 |
| 4 | Claude Fable 5.1 | Max | 15 | #20 |
| 5 | Kimi K3.1 Pro | Max | 15 | #35 |
| 6 | Grok 4.7 | Max | 15 | #51 |
| 7 | Gemini 4 Argon | Max | 15 | #18 |
| 8 | Qwen 3.8 Max | Max | 12.5 | #66 |
| 9 | Kimi K3 | Max | 0 | #61 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=autorlbench-40-issue-20-solution-issue · How It Works · Data refreshed daily, snapshot 2026-11-20.