EComAgentBench - Negative Constraints: leaderboard

Metric: Accuracy (%) on the negative constraint (57 tasks) intent of 662 long-horizon shopping tasks over a catalog of real Amazon products and reviews, where requirements are split across the visible query, a tool-gated user profile and scripted clarification; up to 100 tool calls; graded with typed rubrics by a gemini-3.1-pro-preview judge: a task is correct when the agent recommends the exact target product or one that satisfies every rubric; higher is better. Source: arxiv.org. Saturation forecast: Around May 2028. 7 models tracked.

Top models

#ModelScore
1Claude Opus 4.647.4
2Kimi K2.642.1
3MiniMax-M2.740.4
4GPT-5.4 (Medium)36.8
5GPT-535.1
6GPT-5 Mini24.6
7Qwen 3 30B A3B21.1

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