EpiQAL - Multi-Step Inference: leaderboard
EpiQAL-B (Emory University, University of Illinois Chicago and Microsoft, 2026): 478 single-answer questions that require linking several findings in a PLOS Neglected Tropical Diseases article with epidemiological knowledge, against distractors built from reasoning flaws such as causal reversal or misattributing a finding to another entity. The score is exact match of the selected option set in percent, higher is better, zero-shot without chain-of-thought prompting; models are not told that one option is correct, so selecting extra options fails the item. Qwen3-30B-A3B-Instruct-2507 generated the items, and those that GPT-5-mini, DeepSeek-V3.2-Thinking, Qwen3-32B and Phi-4-mini-instruct found easy had their question stems rewritten.
Metric: Exact Match (%; zero-shot, no CoT, full article). Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral 7B Instruct (v0.3) | 78.9 | #6698 |
| 2 | Qwen 3 32B | 74.3 | #960 |
| 3 | GPT-4.1 Nano | 67.8 | #1992 |
| 4 | Llama 3.3 70B Instruct [4bit] | 65.1 | #1221 |
| 5 | Mistral Large 2 (Nov) Instruct (2411) | 64.4 | #1199 |
| 6 | Qwen 3 30B A3B 2507 Instruct | 57.3 | #1078 |
| 7 | GPT-5 Mini | 53.3 | #349 |
| 8 | Qwen 3 8B | 50.8 | #1838 |
| 9 | Llama 3.1 8B Instruct | 31.8 | #4277 |
| 10 | Phi-4 Mini Instruct | 24.9 | #6452 |
| 11 | GPT-4o Mini | 22.2 | #1627 |
| 12 | Llama 3.2 3B Instruct | 15.7 | #7185 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=epiqal-multi-step-inference · How It Works · Data refreshed daily, snapshot 2026-11-12.