tau-Rec - Constraint Score: leaderboard

Metric: Mean constraint sub-score (0-1 scaled to %), the partial-credit share of typed constraints the recommendation satisfies; the agent recommends one movie from a 153-title TMDB catalog to a GPT-5 mini user simulator through dialogue and catalog tools on 60 tasks (constraints revealed up front, only on asking, or never), 4 trials per task, agent temperature 0 where supported; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 9 models tracked.

Top models

#ModelScore
1DeepSeek V4 Flash (Max)59.6
2DeepSeek V4 Flash (High)58.6
3DeepSeek V4 Flash (Non-reasoning)58.3
4Claude Sonnet 4.657.9
5GPT-5.4 (Medium)56.5
6GPT-5.4 (Non-reasoning)49.6
7GPT-5 Mini42.9
8Qwen 3 32B36.1
9Gemini 2.5 Flash31.1

Interactive version: theaggregate.ai/benchmark?slug=tau-rec-constraint-score · How It Works · Data refreshed daily, snapshot 2026-09-29.