PolyReal: leaderboard

PolyReal (2026) evaluates multimodal models on 545 expert-written polymer science tasks covering five stages of real experimental work. Answers are scored against expert key points by an LLM judge and summarized as F1 from 0 to 1. This board is the overall F1 across all five modules; higher is better.

Metric: Overall F1 score of the key-point PRF framework (completeness and correctness against expert Key Points, LLM-judged) over the five PolyReal workflow modules, 545 questions, zero-shot, temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1O30.601#237
2GPT-50.578#178
3Claude Sonnet 4.5Thinking0.546#261
4Qwen 3 VL 235B A22B (Thinking)0.538#459
5Gemini 2.5 Pro0.536#276
6Grok 40.533#319
7Qwen 3 VL 32B (Thinking)0.525#573
8Gemini 2.5 Flash0.512#463
9DeepSeek R10.484#482
10Gemini 2.0 Flash (Thinking)0.468
11Intern-S10.427#624
12GPT-4o0.367#751
13Qwen 2.5 VL 72B Instruct0.337#841
14GPT-4o Mini0.299#1667
15DeepSeek-VL20.132#2927

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

Interactive version: theaggregate.ai/benchmark?slug=polyreal · How It Works · Data refreshed daily, snapshot 2026-11-21.