RPCBench - Domain - MovieLens-1M - Evidence Distortion: leaderboard
RPCBench tests proactive premise critique in recommendation requests. This series measures evidence distortion on the MovieLens-1M subset (462 questions), using the same eleven-model, three-judge experiment as the headline paper. Published F1R fractions from the authors' per-model CSV are scaled once to Evidence Distortion Rate (%). Lower is better; domains and nested premise-error groups are related slices of the same experiment.
Metric: Evidence Distortion Rate (%). Source: github.com. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.1 8B Instruct | 23.59 | #1021 |
| 2 | Llama 3.1 70B Instruct | 32.47 | #570 |
| 3 | Gemini 3.1 Pro (Preview) | 33.55 | #66 |
| 4 | Qwen 3.5 Plus | 34.42 | #136 |
| 5 | Qwen 3.5 35B A3B | 34.63 | #270 |
| 6 | Qwen 3.5 397B A17B | 35.06 | #155 |
| 7 | Qwen 3.5 122B A10B | 35.06 | #186 |
| 8 | GPT-5.5 | 36.36 | #34 |
| 9 | DeepSeek V4 Pro | 38.96 | #113 |
| 10 | Claude Sonnet 4.6 | 40.26 | #97 |
| 11 | DeepSeek V4 Flash | 40.91 | #156 |
Interactive version: theaggregate.ai/benchmark?slug=rpcbench-domain-movielens-1m-evidence-distortion · How It Works · Data refreshed daily, snapshot 2026-11-07.