AgentCollabBench: leaderboard
Failure to preserve constraints across peer-agent hops. Four-backbone diagnostic protocol over 900 human-validated collaboration tasks; Table 9 reports percentages with standard errors. Judge: DeepSeek V4 Flash. Mean scores, not confidence intervals. This series reports Instruction Decay Rate (%). Source: https://arxiv.org/html/2605.08647v1.
Metric: Instruction Decay Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 35B A3B | 0.9 | #256 |
| 2 | Gemini 2.5 Flash Lite | 1.5 | #422 |
| 3 | GPT-4.1 Mini | 3.4 | #355 |
| 4 | Llama 3.1 8B Instruct | 10.1 | #1035 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=agentcollabbench · How It Works · Data refreshed daily, snapshot 2026-10-20.