DRNoise - Conditional Deference: leaderboard

Metric: Conditional deference (%; share of the tasks a model answered correctly in the clean condition whose noisy-condition answer equals the designated false value from the seeded document; 100 deep-research tasks whose unique gold answer must be recovered from two corroborating indirect record chains, run in a fixed BrowseComp-Plus search-only ReAct agent (top-5 dense retrieval, 512-token snippets, up to 100 iterations); mean of three runs; lower is better). Source: arxiv.org. Saturation forecast: Around 2037. 5 models tracked.

Top models

#ModelScore
1Qwen 3 32B61.3
2Gemini 3.5 Flash72.6
3GPT-5.480.4
4MiMo-V2.596.3
5DeepSeek V4 Flash98.1

Interactive version: theaggregate.ai/benchmark?slug=drnoise-conditional-deference · How It Works · Data refreshed daily, snapshot 2026-09-29.