ToxReason - Mechanistic Reasoning: leaderboard
ToxReason (2026) tests mechanistic chemical toxicity reasoning: given a query molecule with experimental initiating-event evidence from structurally similar compounds, the model predicts liver, heart and kidney toxicity and explains how molecular initiating events lead to the adverse outcome along an Adverse Outcome Pathway. Overall quality of the mechanistic explanation, scored 0 to 10 by a Claude Sonnet 4.5 judge against the reference pathway; higher is better.
Metric: Overall quality (0-10) of the step-wise mechanistic explanation linking molecular initiating events to the organ-level adverse outcome, scored 0 to 10 by a Claude Sonnet 4.5 judge against the reference adverse outcome pathway, on the ToxReason test set (query chemicals with curated human CTD toxicity labels and adverse outcome pathway context, eight retrieved activation and inhibition examples from similar compounds), zero-shot, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around May 2028. 11 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 5.52 | #240 |
| 2 | GPT-5 | 5.42 | #178 |
| 3 | O3 | 5.33 | #237 |
| 4 | GPT-4o | 4.96 | #751 |
| 5 | Gemma 3 27B (IT) | 4.96 | #1324 |
| 6 | O4 Mini | 4.95 | #368 |
| 7 | Llama 3.1 70B Instruct | 4.65 | #1571 |
| 8 | Qwen 2.5 14B Instruct | 4.53 | #1775 |
| 9 | Qwen 3 4B Instruct | 4.52 | #1696 |
| 10 | DeepSeek R1 Distill Llama 70B | 4.49 | #1669 |
| 11 | Llama 3.1 8B Instruct | 3.28 | #4600 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=toxreason-mechanistic-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-21.