EpiBench (Research Workflows) - Final-Turn Accuracy: leaderboard

EpiBench (2026, research workflows) is an episodic benchmark in which multimodal agents search the literature, extract evidence from figures and tables across papers and reuse it over several turns to answer research questions. This board is accuracy in percent on the final turn, which must be answered from evidence gathered earlier; higher is better.

Metric: Final-turn accuracy (%): the multi-evidence fusion turn answered from memory, over the 102 EpiBench episodes (39 easy, 63 hard), each a multi-turn research workflow in which the agent searches papers, extracts figures and tables with tools and answers objective questions, the final turn from stored evidence only; answers judged by GPT-5.2; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.259.41#185
2Kimi K2.557.58#266
3GPT-5 Mini54.9#364
4Gemini 2.5 Pro52.94#276
5Qwen 3 VL 235B A22B (Thinking)51.96#459
6Grok 4.150#462
7Qwen 3 VL 235B A22B Instruct46.08#532
8GLM-4.5V38.24#754

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=epibench-research-workflows-final-turn-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-21.