GroupTravelBench - LLM Judge: leaderboard

GroupTravelBench LLM Judge rates the planning process of an LLM agent in group travel chats on hallucination, tool-use reasoning, interaction, conflict coordination and plan humanization, with a Gemini 3 Flash judge calibrated by a meta-judge (GroupTravelBench, 2026). The score runs from 0 to 100, higher is better.

Metric: LLM-judge process score (0-100): mean of five 1-5 ratings (hallucination, tool-use reasoning, interaction, conflict coordination, plan humanization) rescaled to 0-100 and multiplied by a 1-5 meta-judge calibration over 5, Gemini-3-Flash-Preview judge and meta-judge at temperature 0, over all 650 tasks, on GroupTravelBench, multi-user multi-turn group travel planning in a group-chat sandbox with 10 cached travel tools (DeepSeek-V4-Flash user simulator at temperature 0, agent temperature 0.7, mean of 3 trials per task); GPT-5.1 in instruct mode, the other models with reasoning enabled; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1DeepSeek V4 ProThinking92.3#196
2GPT-5.1Non-reasoning85.2#240
3Qwen 3.5 PlusThinking62.2#242
4Qwen 3.6 Max (Preview)Thinking57.3
5GroupTravelBench Qwen3-235B-A22B-Thinking (checkpoint unspecified)45.4
6Qwen 3.5 FlashThinking43.7#441
7GroupTravelBench Qwen3-4B-Thinking (checkpoint unspecified)24.8
8Qwen 3 4B Instruct22.9#1696
9GroupTravelBench Qwen3-30B-A3B-Thinking (checkpoint unspecified)21.3

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=grouptravelbench-llm-judge · How It Works · Data refreshed daily, snapshot 2026-11-21.