Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
On the Plasticity and Stability for Post-Training Large Language Models
Wenwen Qiang, Ziyin Gu, Jiahuan Zhou, Jie Hu, Jingyao Wang, Changwen Zheng, Hui Xiong
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 |
|---|---|---|---|
| replace GRPO with PCR (probabilistic soft projection) | DeepScaleR-1.5B-Preview · Math pass@1 average | accuracy: 55.0 -> 57.7 | Improved |
| replace GRPO with PCR | DeepSeek-R1-Distill-Qwen-1.5B · Math pass@1 average | accuracy: 46.0 -> 48.5 | Improved |
| replace GRPO with PCR | DeepSeek-R1-Distill-Qwen-7B · Math pass@1 average | accuracy: 65.1 -> 67.6 | Improved |
| replace GCPO with PCR | DeepSeek-R1-Distill-Qwen-7B · Math pass@1 average | accuracy: 66.5 -> 67.6 | Improved |
| apply plain GRPO | DeepScaleR-1.5B-Preview · AMC 2023 | pass@1: 83.0 -> 81.5 | Hurt |
| apply plain GRPO | DeepScaleR-1.5B-Preview · MATH500 | pass@1: 85.2 -> 84.9 | No clear effect |
| apply GRPO to base model | DeepSeek-R1-Distill-Qwen-7B · Math pass@1 average | accuracy: 64.1 -> 65.1 | Improved |
| use PCR adaptive alpha instead of fixed soft projection alpha in {0.2,0.5,0.8} (ablation) | PCR · math reasoning (Figure ablation) | accuracy | Improved |
| apply PCR only to MLP layers instead of all layers (ablation) | PCR · math reasoning (Figure ablation) | accuracy and training time | No clear effect |
| add PCR to GRPO on Qwen2-7B-Instruct | GRPO · HumanEval 0-shot and 5-shot | pass@1: almost 1.2% over baselines | Improved |
| vary KL coefficient beta of GRPO slightly lower (ablation) | GRPO · DeepSeek-R1-Distill-Llama-8B AIME vs WikiText-2 PPL | PPL | Hurt |
| increase PCR parameter learning rate eta (ablation) | PCR · larger update magnitudes | performance | Improved |
Closest papers by meaning
- Optimizing Language Models for Inference Time Objectives using Reinforcement Learning (ICML 2025 · similarity 0.87)
- On Predictability of Reinforcement Learning Dynamics for Large Language Models (ICLR 2026 · similarity 0.87)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · similarity 0.87)
- WS-GRPO: Weakly-Supervised Group-Relative Policy Optimization for Rollout-Efficient Reasoning (ICML 2026 · similarity 0.87)
- SR-GRPO: Stable Rank as an Intrinsic Geometric Reward for Large Language Model Alignment (NeurIPS 2026 · similarity 0.87)
- Compute Aligned Training: Optimizing for Test Time Inference (NeurIPS 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.
