Research map / Reinforcement learning / RL from feedback
Rubric Curriculum RL: Exploiting the Generation-Verification Gap in Non-Verifiable Domains
Tejas Krishnan, Sumeet Ramesh Motwani, Charles London, Suhaas M Bhat, Huitian Jiao, Philip Torr, Riashat Islam, Christopher Summerfield, Christian Schroeder de Witt, Qilong Gu, Shital Shah
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pairwise rewards instead of absolute rubric scores (ablation) | RcRL · FictionBench | win rate: 55.7 -> 73.9 | Improved |
| add curriculum over rubric criteria to pairwise RaR (ablation) | Pairwise RaR · FictionBench | win rate: 69.6 -> 73.9 | Improved |
| add rubric criteria curriculum to absolute RaR | Absolute RaR · FictionBench | win rate: 57.5 -> 73.9 | Improved |
| add curriculum to absolute rewards (curriculum alone) (ablation) | Absolute RaR · FictionBench | win rate: 57.5 -> 55.7 | No clear effect |
| add pairwise rewards to curriculum on HealthBench (ablation) | RcRL · HealthBench full | score improvement: +13.40% -> +14.60% | Improved |
| add pairwise rewards to curriculum on HealthBench consensus (ablation) | RcRL · HealthBench consensus | score improvement: +23.87% -> +25.38% | Improved |
| use RcRL with fixed rubric instead of instance-specific rubrics | RcRL · HealthBench full | score improvement: +13.52% -> +14.60% | Improved |
| add weighting of newest criterion (ablation) | RcRL · FictionBench | win rate: 72.1 -> 73.9 | No clear effect |
| add KL term to RcRL (ablation) | RcRL · FictionBench | win rate: 73.9 -> 50.5 | Hurt |
| add exploration bonus to RcRL (ablation) | RcRL · FictionBench | win rate: 73.9 -> 60.0 | Hurt |
| vary curriculum ordering (ablation) | RcRL · FictionBench | win rate: 73.9; 79.6; 81.0 | Mixed |
| remove curriculum (RaR) vs RcRL on policy entropy | RaR · LechMazur writing benchmark | entropy ratio: 0.73 -> 0.93 | Improved |
Closest papers by meaning
- Reward Hacking in Rubric-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- Step-wise Rubric Rewards for LLM Reasoning (NeurIPS 2026 · similarity 0.90)
- RuCL: Stratified Rubric-Based Curriculum Learning for Multimodal Large Language Model Reasoning (ICML 2026 · similarity 0.89)
- Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR (NeurIPS 2026 · similarity 0.89)
- From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation (ICLR 2026 · similarity 0.87)
- Spurious Rewards: Rethinking Training Signals in RLVR (ICML 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
