EntCollabBench - Approval: leaderboard

Task success in percent on 100 policy approval tasks (80 single-step, 20 multi-step) decided by role-specialized LLM agents for 11 enterprise roles that work behind permission-controlled ITSM, HR, CSM, Gitea and document services and delegate to each other (EntCollabBench, 2026), checked against deterministic policy adjudications. Higher is better.

Metric: Task success (%) on the 100 policy approval tasks (80 single-step and 20 multi-step), judged against deterministic policy adjudications, role-specialized LLM agents (11 enterprise roles behind permission-controlled ITSM, HR, CSM, Gitea and document services) delegate over HTTP, temperature 0, three-judge majority vote; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.670#173
2DeepSeek V4 Flash70#267
3DeepSeek V4 Pro68#196
4GPT-5 Mini62#364
5GPT-5.460#135
6Gemini 3.1 Pro (Preview)58#99
7Gemini 3.1 Flash Lite (Preview)46#417
8Qwen 3.5 122B A10B45#346
9Qwen 3.5 35B A3B43#492
10MiniMax-M2.737#539
11Qwen 3.5 9B23#828
12MiMo-V2-Flash7#684

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

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