AARRI-Bench (Hermes Agent) - Hands-on: leaderboard
AARRI-Bench (Act As a Real Research Intern, 2026) is a set of 82 hand-crafted AI-research tasks graded by unit tests, testing whether an agent harness and model behave like a careful human researcher. This board is the percentage of the hands-on tasks: coding, experimental setup and data processing solved in full with the Hermes Agent harness; higher is better.
Metric: Task success rate (%) under the 0/1 reward on the hands-on tasks: coding, experimental setup and data processing; Harbor containerized runs, Hermes Agent agent harness; a configuration's tasks without a valid run are left out of its denominator; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 Plus | 64.3 | #222 |
| 2 | Claude Opus 4.7 | 57.1 | #79 |
| 3 | MiniMax-M2.7 | 57.1 | #518 |
| 4 | Claude Sonnet 4.6 | 50 | #161 |
| 5 | DeepSeek V4 Flash | 50 | #258 |
| 6 | Qwen 3 235B A22B 2507 (Thinking) | 38.5 | #463 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=aarri-bench-hermes-agent-hands-on · How It Works · Data refreshed daily, snapshot 2026-11-13.