Research map / Reinforcement learning / RL from feedback
Beyond Normalization: Rethinking the Partition Function as a Difficulty Scheduler for RLVR
Dohyung Kim, Minbeom Kim, Jeonghye Kim, Sangmook Lee, Sojeong Rhee, Kyomin Jung
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 partition-function-guided prompt selection and replay (PACED-RL) over GRPO | Qwen2.5-Math-7B · Math avg of five benchmarks | pass@1: 37.2 -> 40.1 | Improved |
| PACED-RL versus FlowRL | Qwen2.5-Math-1.5B · Math avg of five benchmarks | pass@1: 27.1 -> 32.1 | Improved |
| add accuracy-estimation-error-prioritized replay (ablation) | PACED-RL (Qwen2.5-Math-1.5B) · Math avg of five benchmarks | pass@1: 30.2 -> 32.1 | Improved |
| add replay on AIME 24 (ablation) | PACED-RL (Qwen2.5-Math-1.5B) · AIME 24 Avg@32 | pass@1: 9.8 -> 13.3 | Improved |
| use greedy top-m selection instead of softmax soft sampling (ablation) | PACED-RL (Qwen2.5-Math-1.5B) · MATH-500 Avg@8 | Avg@8: 72.2 -> 73.2 | Improved |
| target accuracy tau=0.5 instead of 0.1 or 0.9 (ablation) | PACED-RL (Qwen2.5-Math-1.5B) · MinervaMath test accuracy | test accuracy | Improved |
| target accuracy tau=0.1 (ablation) | PACED-RL (Qwen2.5-Math-1.5B) · MinervaMath test accuracy | test accuracy | Hurt |
| omit KL term in accuracy estimator (Eq. 8 vs Eq. 7) (ablation) | PACED-RL · prompt selection overlap | selection overlap | No clear effect |
| learn partition function instead of batch-estimate VarGrad (ablation) | FlowRL (Qwen2.5-Math-1.5B) · 5 math benchmarks | test performance | Improved |
| use VarGrad batch-estimate partition function in FlowRL (ablation) | FlowRL (Qwen2.5-Math-1.5B) · 5 math benchmarks | test performance | Improved |
| use partition function accuracy estimates instead of over-sampling (DS/LILO) | PACED-RL (Qwen2.5-Math-7B) · Math avg of five benchmarks | pass@1 and rollouts: LILO 40.1 vs PACED-RL 40.1; rollouts 4.0x -> 1.0x | No clear effect |
| over-sampling prompt selection (DS, LILO) for code generation | DeepSeek-R1-Distill-Qwen-1.5B · HumanEval+ / LiveCodeBench | pass@1 vs wall-clock | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- FlowRL — FlowRL: Matching Reward Distributions for LLM Reasoning (ICLR 2026) · better on 5 of 5 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Exploration (ICML 2026 · similarity 0.89)
- On the Direction of RLVR Updates for LLM Reasoning: Identification and Exploitation (ICLR 2026 · similarity 0.89)
- RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs (ICLR 2026 · similarity 0.88)
- Generalization of RLVR Using Causal Reasoning as a Testbed (ICLR 2026 · similarity 0.88)
- Reshaping Reasoning in LLMs: A Theoretical Analysis of RL Training Dynamics through Pattern Selection (ICLR 2026 · similarity 0.88)
- The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (NeurIPS 2025 · 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.
