RPCBench - Fact Fabrication: leaderboard
RPCBench tests proactive critique of flawed recommendation premises, locating the premise problem and suggesting a suitable strategy while remaining faithful to the provided evidence. This series reports RPCBench - Fact Fabrication from 2609.00918v1, S3.T2: Fact Fabrication Rate (%). Lower is better.
Metric: Fact Fabrication Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 122B A10B | 1.8 | #329 |
| 2 | Qwen 3.5 Plus | 1.9 | #227 |
| 3 | Qwen 3.5 35B A3B | 2.1 | #480 |
| 4 | Qwen 3.5 397B A17B | 2.2 | #242 |
| 5 | GPT-5.5 | 3.9 | #48 |
| 6 | DeepSeek V4 Pro | 7 | #181 |
| 7 | Gemini 3.1 Pro (Preview) | 7.6 | #91 |
| 8 | DeepSeek V4 Flash | 8.4 | #254 |
| 9 | Claude Sonnet 4.6 | 13.2 | #158 |
| 10 | Llama 3.1 70B Instruct | 46.2 | #1541 |
| 11 | Llama 3.1 8B Instruct | 53.7 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rpcbench-fact-fabrication · How It Works · Data refreshed daily, snapshot 2026-11-12.