Research map / Reinforcement learning / RL from feedback
Injecting Distributional Awareness into MLLMs via Reinforcement Learning for Deep Imbalanced Regression
Yao DU, Shanshan Song, Xiaomeng Li
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 CCC batch-level reward to a pairwise rank reward (ablation) | GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 9.35 -> 5.57 | Improved |
| add CCC reward on top of regression MAE reward (ablation) | GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 5.85 -> 5.61 | Improved |
| add CCC reward on top of Spearman correlation reward (ablation) | GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 6.81 -> 5.68 | Improved |
| add CCC reward on top of Spearman reward (ablation) | GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | GM (few-shot): 3.40 -> 4.18 | Hurt |
| use CCC reward instead of per-sample DISCO MAE reward (ablation) | GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 5.95 -> 5.52 | Improved |
| use CCC-GRPO instead of SFT | Qwen2.5-VL-3B · IMDB-Movie-DIR | MAE (few-shot): 21.51 -> 16.35 | Improved |
| use CCC-GRPO instead of SFT | Qwen2.5-VL-3B · IMDB-Movie-DIR | MAE (medium-shot): 11.21 -> 8.12 | Improved |
| use CCC-GRPO instead of SFT | Qwen2.5-VL-7B · BoneAge-DIR | MAE (all): 17.09 -> 13.35 | Improved |
| use DrGRPO instead of GRPO with CCC reward (ablation) | CCC-GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 5.52 -> 5.73 | No clear effect |
| use RegGRPO instead of GRPO with CCC reward (ablation) | CCC-GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 5.52 -> 5.56 | No clear effect |
| increase generations from 4 to 8 (ablation) | CCC-GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 5.52 -> 5.53 | No clear effect |
| increase batch size from 16 to 32 (ablation) | CCC-GRPO · AgeDB-DIR (Qwen2.5-VL-3B) | MAE (all): 5.52 -> 5.56 | No clear effect |
Closest papers by meaning
- Incentivizing Versatile Video Reasoning in MLLMs via Data-Efficient Reinforcement Learning (CVPR 2026 · similarity 0.86)
- Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling (NeurIPS 2025 · similarity 0.86)
- Scalable Power Sampling: Unlocking Efficient, Training-Free Reasoning for LLMs via Distribution Sharpening (ICML 2026 · similarity 0.86)
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- PowerFlow: Unlocking the Dual Nature of LLMs via Principled Distribution Matching (ICML 2026 · similarity 0.85)
- Linking Perception, Confidence and Accuracy in MLLMs (CVPR 2026 · similarity 0.85)
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.
