Xiaohongshu and Fudan Propose DR-Rubric: Agent-Written Rubrics as RL Rewards
jiqizhixin · x · 2026-10-28
Xiaohongshu and Fudan University introduce DR-Rubric, reported accepted to NeurIPS 2026, tackling who writes grading rubrics for open-ended questions.
- Bottleneck: generic instructions like "be clear and complete" miss the facts and constraints that make an answer useful; training against a weak rubric rewards wrong behavior.
- Mechanism: an agent first researches the question, collecting external evidence and likely failure modes, converts them into individually checkable constraints, and uses the fraction satisfied as an RL reward; a bootstrapping loop lets the trained model generate rubrics for later rounds.
- Results: with just 1,000 RL instances on a shared Qwen3-8B-SFT init, the GPT-5-rubric variant scores 73.7 on ResearchQA and 43.0 on DeepResearchBench, 3.8 and 3.2 points above the best baselines.
More from Research
- François Fleuret Proposes Tagging Every Token 'Heard/Thought/Said' to Solve Alignment — francoisfleuret · 2026-10-28
- AI verifies the 15,000-page finite simple groups classification in 24 days with 50M lines of Lean — FuSheng_0306 · 2026-10-28
- VibeVideoBench: multimodal agents coding videos often fail on time management — Muxi Diao · 2026-10-28
- Nearly 600 runs show full BPTT beats truncated backprop in looped transformers — Xuyuan Xiong · 2026-10-28
- Memory heuristics hurt LLM group decisions: full transcripts cut accuracy too — Anouar Imel · 2026-10-28
- Google's AI Mode feeds uncited Knowledge Graph biographies into answers, analysis shows — dejanseo · 2026-10-28