Research map / Reinforcement learning / RL from feedback
When Rubrics Fail: Error Enumeration as Reward for Reference-Free RL Post-Training
Wisdom Ikezogwo, Mehmet Saygin Seyfioglu, Ranjay Krishna, Karim Bouyarmane
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 |
|---|---|---|---|
| use implicit error counting reward instead of rubric reward (RaR) | GRPO-trained Kontext VTO · MDressBench (flat references) | CEC: 5.60 -> 5.31 | Improved |
| use implicit error counting reward instead of rubric reward (RaR) | GRPO-trained Kontext VTO · MDressBench (non-flat references) | CEC: 5.533 -> 5.203 | Improved |
| use implicit error counting instead of direct scalar scoring | GRPO-trained Kontext VTO · MDressBench (flat references) | CEC: 5.373 -> 5.312 | Improved |
| add RL post-training with RaR rubric reward vs direct scoring | GRPO-trained Kontext VTO · MDressBench (flat references) | CEC: 5.373 -> 5.596 | Hurt |
| use explicit error counting (EEC) instead of implicit (ablation) | IEC reward · MDressBench (non-flat) | CEC: 5.20 -> 5.72 | Hurt |
| use explicit error counting (EEC) instead of implicit (ablation) | IEC reward · MDressBench (flat references) | CEC: 5.31 -> 6.03 | Hurt |
| add group calibration (ablation) | IEC reward · MDressBench (non-flat references) | CEC: 5.259 -> 5.203 | Improved |
| add group calibration (ablation) | IEC reward · VITON-HD | FID: 5.190 -> 5.120 | Improved |
| add group calibration (ablation) | IEC reward · DressCode | FID: 3.980 -> 3.610 | Improved |
| use CEC as selector instead of Direct or RaR | Human preference selection · Human study (25 samples) | Top-1 accuracy: 0.30 -> 0.60 | Improved |
| increase best-of-n from 8 to 12 (ablation) | IEC GRPO · MDressBench (flat) | CEC: 5.31 -> 4.80 | Improved |
| increase best-of-n from 8 to 12 (ablation) | IEC GRPO · MDressBench (non-flat) | CEC: 5.20 -> 5.27 | Hurt |
Closest papers by meaning
- Step-wise Rubric Rewards for LLM Reasoning (NeurIPS 2026 · similarity 0.91)
- Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR (NeurIPS 2026 · similarity 0.90)
- Reward Hacking in Rubric-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- Post-hoc Reward Calibration: A Case Study on Length Bias (ICLR 2025 · similarity 0.87)
- From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation (ICLR 2026 · similarity 0.87)
- Auto-Rubric as Reward: From Implicit Preference to Explicit Generative Criteria (NeurIPS 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.
