RecRM-Bench - Factual Consistency: leaderboard

RecRM-Bench (Meituan, 2026) tests models as reward models for an agentic recommender system, on real query-response logs. This board covers the factual-consistency judgement of item and content hallucinations; the score is the percent of judgements that match the gold label; higher is better.

Metric: Accuracy (%) of the factual-consistency judgement (item and content hallucinations) the model gives for a recommender response, zero-shot, the model acting as a reward model for an agentic recommender on RecRM-Bench (real Meituan query-response logs), its parsed score compared with the gold label; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1LongCat-Flash-Chat67.34#820
2LongCat-Flash-Thinking64.98#762
3GPT-4.162.96#506
4Qwen 3 MaxThinking56.57#420
5DeepSeek V3.2Non-reasoning43.43#425

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=recrm-bench-factual-consistency · How It Works · Data refreshed daily, snapshot 2026-11-22.