RPCBench - Error Group - Inconsistent Premise - Evidence Faithfulness: leaderboard
RPCBench tests proactive premise critique in recommendation requests. This series measures evidence faithfulness on the Inconsistent Premise subset (1,867 questions), using the same eleven-model, three-judge experiment as the headline paper. Published EFI fractions from the authors' per-model CSV are scaled once to Faithfulness Index (%). Higher is better; domains and nested premise-error groups are related slices of the same experiment.
Metric: Faithfulness Index (%). Source: github.com. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 Plus | 93.47 | #136 |
| 2 | Qwen 3.5 397B A17B | 92.85 | #155 |
| 3 | Qwen 3.5 122B A10B | 92.66 | #186 |
| 4 | GPT-5.5 | 92.21 | #34 |
| 5 | Qwen 3.5 35B A3B | 90.6 | #270 |
| 6 | Gemini 3.1 Pro (Preview) | 89.29 | #66 |
| 7 | DeepSeek V4 Pro | 85.97 | #113 |
| 8 | DeepSeek V4 Flash | 84.92 | #156 |
| 9 | Claude Sonnet 4.6 | 84.49 | #97 |
| 10 | Llama 3.1 70B Instruct | 39.1 | #570 |
| 11 | Llama 3.1 8B Instruct | 32.16 | #1021 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rpcbench-error-group-inconsistent-premise-evidence-faithfulness · How It Works · Data refreshed daily, snapshot 2026-11-07.