PSN-RLVR Enhances LLM Reasoning Diversity via Parameter Perturbation

1 reports archived1 independent sourcesupdated 10/29/2026, 11:40:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Researchers from Shanghai Institute of AI and others proposed PSN-RLVR, which applies consistent small random perturbations to model parameters throughout the reasoning trajectory. This method addresses the issue where RL-trained models become accurate but lose diversity in solution paths, significantly improving Pass@K success rates and out-of-distribution generalization. The paper has been accepted by NeurIPS 2026.

LATEST/Researchers from Shanghai Institute of AI and others proposed PSN-RLVR, which applies consistent small random perturbations to model parameters throughout the reasoning trajectory. This method addresses the issue where RL-trained models become accurate but lose diversity in solution paths, significantly improving Pass@K success rates and out-of-distribution generalization. The paper has been accepted by NeurIPS 2026.

Hourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first

量子位 (微信公众号)T2Oct 29, 11:40·78 pts
PSN-RLVR Enhances LLM Reasoning Diversity via Parameter Perturbation
- Highlights that while Reinforcement Learning improves single-shot accuracy (Pass@1), it often suppresses solution path diversity, leading to lower multi-attempt success rates (Pass@256).
- PSN-RLVR introduces fixed parameter perturbations before generation, allowing the model to maintain distinct thinking tendencies throughout the entire reasoning trajectory to generate coherent new solutions.
- Validated on the GRPO algorithm, the method not only boosts multi-sampling success rates but also demonstrates superior out-of-distribution generalization in scientific QA and coding tasks.
View Related Reports & Analysis Original Source