Research map / Reinforcement learning / RL from feedback
Smarter Not Harder: Generative Process Evaluation with Intrinsic-Signal Driving and Ability‑Adaptive Reward Shaping
Tao He, Rongchuan Mu, Lizi Liao, Yixin Cao, Yang Li, Yijia Luo, Weixun Wang, Ming Liu, Bing Qin
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 intrinsic-signal-driven evaluation (vs LLM-as-a-judge) (ablation) | TP-GRPO · AIME 24 (1.5B) | Avg@16 accuracy: 31.04 -> 33.12 | Improved |
| add intrinsic-signal-driven evaluation (vs LLM-as-a-judge) (ablation) | TP-GRPO · AMC 23 (1.5B) | Avg@16 accuracy: 63.93 -> 64.01 | No clear effect |
| add thought-level step merging (ablation) | TP-GRPO · AIME 25 (1.5B) | Avg@16 accuracy: 22.29 -> 25.63 | Improved |
| add thought-level step merging (ablation) | TP-GRPO · AIME 24 (1.5B) | Avg@16 accuracy: 31.66 -> 33.12 | Improved |
| add difficulty-aware adaptive reward (vs fixed +1/-1) (ablation) | TP-GRPO · AIME 25 (1.5B) | Avg@16 accuracy: 22.92 -> 25.63 | Improved |
| add difficulty-aware adaptive reward (vs fixed +1/-1) (ablation) | TP-GRPO · AIME 24 (1.5B) | Avg@16 accuracy: 32.71 -> 33.12 | Improved |
| add process rewards for correct solutions (ablation) | TP-GRPO · AIME 24 (1.5B) | Avg@16 accuracy: 30.00 -> 33.12 | Improved |
| add process rewards for incorrect solutions (ablation) | TP-GRPO · AIME 25 (1.5B) | Avg@16 accuracy: 22.29 -> 25.63 | Improved |
| add process rewards for incorrect solutions (ablation) | TP-GRPO · AIME 24 (1.5B) | Avg@16 accuracy: 32.01 -> 33.12 | Improved |
| add process rewards for incorrect solutions (ablation) | TP-GRPO · AIME 25 vs AIME 24 | Avg@16 accuracy: 30.00 -> 33.12 | Mixed |
| raise alpha from 0 to 1 (ablation) | TP-GRPO · AIME24/AIME25/AMC23 avg (1.5B) | average accuracy: 39.17 -> 40.92 | Improved |
| raise alpha from 1 to 4 (ablation) | TP-GRPO · AIME24/AIME25/AMC23 avg (1.5B) | average accuracy: 40.92 -> 39.72 | Hurt |
Closest papers by meaning
- GenPRM: Scaling Test-Time Compute of Process Reward Models via Generative Reasoning (AAAI 2026 · similarity 0.92)
- Think-RM: Enabling Long-Horizon Reasoning in Generative Reward Models (NeurIPS 2025 · similarity 0.90)
- From Mathematical Reasoning to Code: Generalization of Process Reward Models in Test-Time Scaling (AAAI 2026 · similarity 0.90)
- Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning (ICLR 2025 · similarity 0.89)
- AdaptiveStep: Automatically Dividing Reasoning Step through Model Confidence (ICML 2025 · similarity 0.88)
- Boosting LLM Reasoning via Human-Inspired Reward Shaping (NeurIPS 2026 · similarity 0.88)
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.
