NoisyBench - Hard Negative Distractors: leaderboard
NoisyBench (Lost in the Noise; Lee et al., KAIST and LG AI Research, 2026) with a task-specific hard-negative distractor that looks relevant but does not help: the datasets' own hard negatives for RAG, and documents generated by Gemini-2.5-Pro and filtered so they neither change nor reveal the answer for the others. Each of the 11 datasets keeps its own metric (pass@k, pass^k for TauBench) and the score is their harmonic mean in percent; higher is better.
Metric: Harmonic Mean Score (%; 11 datasets' pass@k, pass^k for TauBench; a task-specific hard-negative distractor added). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-OSS-120B (High) | 50.2 | #699 (GPT-OSS-120B) |
| 2 | Gemini 2.5 Pro | 48 | #262 |
| 3 | DeepSeek R1 0528 | 47.6 | #425 |
| 4 | Gemini 2.5 Flash | 45.6 | #452 |
| 5 | Qwen 3 30B A3B 2507 (Thinking) | 35 | #759 |
| 6 | Qwen 3 4B 2507 (Thinking) | 32.7 | #1388 |
| 7 | DeepSeek R1 Distill Llama 8B | 6.3 | #6773 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=noisybench-hard-negative-distractors · How It Works · Data refreshed daily, snapshot 2026-11-12.