AutoRLBench 32 Issue / 18 Solution - Issue: leaderboard

AutoRLBench native 32-issue/18-solution task-set edition. Same published root-cause and repaired-workspace protocols, one run per model/task/level. Issue is an LLM judgment against hidden oracle; solution is the harness rerun metric threshold, with reward hacking failing. Overall counts passes over all50runs, not an average of two task-level rates. The source README announces32issue tasks and24solution-capable tasks, but exactly18 have scored solution runs; do not use24as the reported denominator. Exact task-ID sets are pinned and all rates reproduce the native table to its stated one decimal. Separate from the previously reviewed33/20task-set edition. This board reports Root cause found, 32 failed training runs (%). Higher is better.

Metric: Root cause found, 32 failed training runs (%). Source: thinking-space.github.io. Saturation forecast: Around October 2027. 10 models tracked. Updated 14 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-6 AstraMax46.9#6
2GPT-6.1 SolMax40.6#10
3GLM-5.3Max34.4#68
4Claude Fable 5.1Max28.1#15
5Qwen 3.8 MaxMax28.1#59
6Kimi K3.1 ProMax25#29
7Grok 4.7Max25#47
8Gemini 4 ArgonMax25#12
9Kimi K3Max0#56
10DeepSeek V4 ProMax0#191

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=autorlbench-32-issue-18-solution-issue · How It Works · Data refreshed daily, snapshot 2026-11-16.