Mr Dre - Format Feedback Break Rate: leaderboard
Multi-turn report revision by deep research agents on Mr Dre (Chen et al., 2026): after a first report on an expert-written research question from ResearchRubrics, RigorousBench or ResearcherBench, the agent receives simulated user feedback asking for a change in presentation and revises the report. Break rate in percent is the share of checklist criteria covered in the first report whose coverage score drops after the revision, judged by GPT-4.1-mini and averaged over the three question sets; lower is better.
Metric: Break rate (%; share of checklist criteria covered in the turn-1 report whose coverage score drops after the turn-2 revision following format feedback; averaged over the three question sets; GPT-4.1-mini judge). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | dr-tulu | 14 | |
| 2 | OpenAI o4-mini Deep Research | 14.8 | |
| 3 | Tongyi-DeepResearch-30B-A3B | 23.2 | |
| 4 | LangChain Open Deep Research (GPT-4.1-mini) | 23.9 | |
| 5 | Sonar Deep Research | 27.8 |
Interactive version: theaggregate.ai/benchmark?slug=mr-dre-format-feedback-break-rate · How It Works · Data refreshed daily, snapshot 2026-11-12.