Research map / Reinforcement learning / RL from feedback
Beyond Token-level Supervision: Unlocking the Potential of Decoding-based Regression via Reinforcement Learning
Ming Chen, Sheng Tang, Rong-Xi Tan, Ziniu Li, Jiacheng Chen, Ke Xue, Chao Qian
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 sequence-level RL (ReMax) finetuning after CE | decoding-based regressor · TALENT (100 regression tasks, median aggregation) | RMSE: 0.5484 -> 0.5190 | Improved |
| add sequence-level RL (GRPO) finetuning after CE | decoding-based regressor · TALENT (100 regression tasks, median aggregation) | R2: 0.6124 -> 0.6248 | Improved |
| add token-level NTL-WAS finetuning | decoding-based base model · TALENT (100 regression tasks, median aggregation) | RMSE: 0.5484 -> 0.5478 | No clear effect |
| add DIST2 token-level loss | decoding-based base model · TALENT (100 regression tasks, mean aggregation) | R2: 0.6368 -> 0.4678 | Hurt |
| replace pointwise head with GenRe2-ReMax decoder | regression head · TALENT (100 regression tasks, median aggregation) | R2: 0.5708 -> 0.6459 | Improved |
| add reward standardization (GRPO) (ablation) | GenRe2-GRPO · TALENT (100 tasks, digit base 10) | R2: 0.6089 -> 0.5872 | Hurt |
| add importance-sampling clipping (ablation) | GenRe2-GRPO · TALENT (100 tasks, digit base 10) | R2: 0.5854 -> 0.5872 | No clear effect |
| use greedy baseline instead of mean baseline (ablation) | GenRe2-GRPO · TALENT (100 tasks, digit base 10) | R2: 0.5872 -> 0.5876 | No clear effect |
| increase normalized-tokenization digit base to 10 (ablation) | GenRe2-GRPO · TALENT (100 tasks) | R2 | Hurt |
| add GenRe2-ReMax finetuning | RLM code regression · APPS Leetcode | RMSE: 0.493 -> 0.474 | Improved |
| finetune with CE on dataset | RLM code regression · Triton Kernel Latency | RMSE: 1.095 -> 16.37 | Hurt |
| add GenRe2-ReMax finetuning | RLM code regression · Triton Kernel Latency | Rank correlation: 0.536 -> 0.598 | Improved |
Closest papers by meaning
- StepORLM: A Self-Evolving Framework With Generative Process Supervision For Operations Research Language Models (ICLR 2026 · similarity 0.88)
- Controlled Decoding from Language Models (ICML 2024 · similarity 0.88)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · similarity 0.87)
- Reinforcing LLM Agents via Policy Optimization with Action Decomposition (NeurIPS 2024 · similarity 0.87)
- Not Only Where, But When: Temporal Scheduling for RLVR (NeurIPS 2026 · similarity 0.87)
- You Can Learn Tokenization End-to-End with Reinforcement Learning (ICML 2026 · similarity 0.87)
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.
