ActiveSciBench-Chem (LLM-AutoSciLab): leaderboard
Active discovery of enzyme-kinetic rate laws from ActiveSciBench (Kabra et al., 2026): the system designs experiments on a simulated enzyme-kinetics oracle with a seven-variable interface under a fixed query budget, works out which variables matter and recovers the hidden rate law, over 57 tasks. Symbolic accuracy in percent, an LLM judge checking equivalence to the hidden law up to fitted constants; higher is better. The listed model drives the LLM-AutoSciLab closed-loop framework, with Qwen2.5-7B-Instruct as its fixed ensemble helper.
Metric: Symbolic accuracy (%; share of the 57 enzyme-kinetics tasks whose recovered rate law an LLM judge rates equivalent to the hidden law up to fitted constants; budget-limited active experiments driven by the LLM-AutoSciLab framework with this backbone). Source: arxiv.org. Saturation forecast: Estimated already saturated. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o Mini | 35.09 | #1627 |
| 2 | Qwen 3 32B | 25.56 | #960 |
| 3 | Qwen 3 14B | 23.88 | #1235 |
| 4 | Qwen 3 4B | 12.92 | #2297 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=activescibench-chem-llm-autoscilab · How It Works · Data refreshed daily, snapshot 2026-11-12.