Research map / Reinforcement learning / RL from feedback
RLMR: Reinforcement Learning with Mixed Rewards for Creative Writing
JianXing Liao, Tian Zhang, Xiao Feng, Yusong Zhang, Haorui Wang, Bosi Wen, Ziying Wang, Runzhi Shi
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 |
|---|---|---|---|
| train with writing reward model only | Qwen2.5-32B GRPO · ComplexBench | accuracy: 74.78% -> 68.42% | Hurt |
| train with writing reward model only | Qwen2.5-32B GRPO · WritingBench | score: 6.14 -> 7.05 | Improved |
| train with writing reward model only | Qwen2.5-32B GRPO · IFEval | accuracy: 83.36% -> 80.41% | Hurt |
| train with constraint verification only | Qwen2.5-32B GRPO · WriteEval | win rate: 3.93% -> 1.24% | Hurt |
| train with constraint verification only | Qwen2.5-32B GRPO · ComplexBench | accuracy: 74.78% -> 83.94% | Improved |
| use RLMR dynamic penalty instead of fixed linear weighting | Qwen2.5-32B · WritingBench | score: 7.13 -> 7.93 | Improved |
| use RLMR instead of linear weighting | Qwen2.5-32B · ComplexBench | accuracy: 73.91% -> 79.04% | Improved |
| use RLMR instead of linear weighting | Qwen2.5-72B · WriteEval | win rate: 10.22% -> 17.18% | Improved |
| use RLMR instead of linear weighting | Qwen3-8B · WriteEval | win rate: 26.64% -> 31.69% | Improved |
| use RLMR instead of linear weighting | DeepSeek-R1-Distill-Llama-8B · ComplexBench | accuracy: 53.91% -> 52.35% | Hurt |
| use RLMR instead of linear weighting | DeepSeek-R1-Distill-Llama-8B · WritingBench | score: 5.68 -> 7.41 | Improved |
| add dynamic sampling (DAPO-style filtering) to RLMR (ablation) | Qwen2.5-32B · WritingBench | score: 7.34 -> 7.93 | Improved |
Closest papers by meaning
- PoEM: Predicting New RL Outcomes from Existing Policies (NeurIPS 2026 · similarity 0.86)
- Towards Cost-Effective Reward Guided Text Generation (ICML 2025 · similarity 0.86)
- BLEUBERI: BLEU is a surprisingly effective reward for instruction following (NeurIPS 2025 · similarity 0.86)
- RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style (ICLR 2025 · similarity 0.85)
- Reward Is Enough: LLMs Are In-Context Reinforcement Learners (ICLR 2026 · similarity 0.85)
- Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback (NeurIPS 2025 · 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.
