EvoClawBench - Baseline (nanobot): leaderboard
EvoClawBench (Shanghai Jiao Tong University and collaborators, 2026) asks whether agents can turn evidence from their own runs into reusable skills, over 100 fixture-backed tasks with 502 repeated sub-problems in coding, data, office, security, operations and document work. In this baseline mode the agent solves each task directly and may not create skills. Execution-phase mean task score in percent in the nanobot agent runtime, higher is better.
Metric: Execution-only mean task score (%; 100 fixture-backed tasks plus a sanity task, 502 sub-problems across coding, data, office, security, operations and document workflows, automated or hybrid graders; one run per task in the nanobot agent runtime, execution phase only; the agent solves each task directly and may not create skills). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 96.13 | #123 |
| 2 | MiniMax-M2.7 | 90.97 | #518 |
| 3 | GPT-5.4 Mini | 88.71 | #389 |
| 4 | DeepSeek V4 Pro | 77.77 | #188 |
| 5 | Qwen 3.6 Plus | 56.45 | #222 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=evoclawbench-baseline-nanobot · How It Works · Data refreshed daily, snapshot 2026-11-13.