MMShopBench - Judge@3: leaderboard

Metric: Judge success at 3 (%; share of cases in which any of the first three recommended products satisfies the annotated purchase intent and every mandatory requirement according to a GPT-5.5 multimodal judge given the product evidence; multimodal multi-turn shopping requests from real assistant logs, answered by a tool-using agent over a frozen offline catalog with text and region-aware image retrieval (at most eight tool steps), followed by evidence-grounded verification and selection of the candidate pool by the policy model). Source: arxiv.org. Saturation forecast: Around December 2026. 7 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)73.4
2Claude Opus 4.8 (Thinking)72
3Kimi K2.6 (Thinking)47.8
4MiniMax-M2.726
5Qwen 3.5 9B (Non-reasoning)20.4
6Qwen 3.5 27B (Non-reasoning)10.7
7Qwen 3.5 122B A10B (Non-reasoning)5.9

Interactive version: theaggregate.ai/benchmark?slug=mmshopbench-judge-3 · How It Works · Data refreshed daily, snapshot 2026-09-29.