Research map / Reinforcement learning / RL from feedback
PoLi-RL: A Point-to-List Reinforcement Learning Framework for Conditional Semantic Textual Similarity
Zixin Song, Bowen Zhang, Qian-Wen Zhang, di yin, Xing Sun, Chunping 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 the two-stage pointwise-to-hybrid curriculum instead of naive listwise RL (ablation) | Qwen3-8B RL · C-STS | Spearman: 38.19 -> 48.18 | Improved |
| naive listwise RL instead of few-shot (ablation) | Qwen3-8B · C-STS | Spearman: 37.9 -> 38.19 | No clear effect |
| add binary reward in Stage I (ablation) | PoLi-RL Stage I · C-STS | Spearman: 44.19 -> 44.77 | Improved |
| add Stage II hybrid reward after Stage I (ablation) | PoLi-RL · C-STS | Spearman: 44.77 -> 48.18 | Improved |
| add listwise reward to Stage II (ablation) | PoLi-RL Stage II · C-STS | Spearman: 46.71 -> 48.18 | Improved |
| add pairwise reward to Stage II (ablation) | PoLi-RL Stage II · C-STS | Spearman: 47.6 -> 48.18 | Improved |
| slice size N=24 instead of 16, 32, 40 or 48 (ablation) | PoLi-RL PSRR · C-STS | Spearman: 47.16 / 47.44 / 47.18 / 46.78 -> 48.18 | Improved |
| pairwise weight 1.5 instead of 1.0 (ablation) | PoLi-RL Stage II · C-STS | Spearman: 47.83 -> 48.18 | Improved |
| generation multiplicity G=8 instead of 4 or 12 (ablation) | PoLi-RL Stage II · C-STS | Spearman: 47.47 / 47.98 -> 48.18 | Mixed |
| pairwise base reward 0.5 instead of 0 (ablation) | PoLi-RL Stage II · C-STS | Spearman: 47.90 -> 48.18 | Improved |
| vary binary reward weight lambda2 in Stage I (ablation) | PoLi-RL Stage I · C-STS | Spearman: 44.94 / 44.77 / 44.67 / 44.76 | No clear effect |
| RL instead of regression SFT with Pcc loss | Qwen3-8B · C-STS | Spearman: 46.44 -> 48.18 | Improved |
Closest papers by meaning
- Sequence to Sequence Reward Modeling: Improving RLHF by Language Feedback (AAAI 2025 · similarity 0.88)
- RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style (ICLR 2025 · similarity 0.88)
- Generative RLHF-V: Learning Principles from Multi-modal Human Preference (NeurIPS 2025 · similarity 0.88)
- Webscale-RL: Automated Data Pipeline for Scaling RL Data to Pretraining Levels (ICLR 2026 · similarity 0.87)
- MA-RLHF: Reinforcement Learning from Human Feedback with Macro Actions (ICLR 2025 · similarity 0.87)
- Co-rewarding: Stable Self-supervised RL for Eliciting Reasoning in Large Language Models (ICLR 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.
