AJ-Bench (LLM-as-a-Judge) - Deep Search: leaderboard

Verifying agent trajectories on AJ-Bench (2026), the Deep Search subdomain (156 trajectories of multi-hop web research tasks): the judge decides whether each trajectory succeeded and is scored against binary human-verified labels. In this setting the judge reads the trajectory only, with no tools or environment access. F1 in percent; higher is better.

Metric: F1 (%) on Deep search (Mind2Web 2-derived items; 52 tasks, 156 trajectories), averaged over three runs, with the judge reading the trajectory only (no tools or environment access), of AJ-Bench, judging whether agent trajectories succeeded against binary human-verified labels (155 tasks, 516 trajectories), default model configurations; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1LongCat-Flash-Chat81.8#820
2Claude Sonnet 4.581.34#267
3Qwen 3 235B A22B81.33#633
4Gemini 3 Pro (Preview)81.26#123
5Gemini 2.5 Pro81.22#280
6Claude Opus 4.581.11#165
7GPT-580.37#185
8Kimi K2 090580.17#578
9Grok 478.32#321
10GLM-4.677.88#496
11GPT-5.170.9#248
12GPT-5 MiniLow68.42#373
13DeepSeek V3.262.91#425

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=aj-bench-llm-as-a-judge-deep-search · How It Works · Data refreshed daily, snapshot 2026-11-22.