Research map / Reinforcement learning / RL from feedback

DyJR: Preserving Local Policy Plasticity in Reinforcement Learning with Verifiable Rewards via Dynamic Jensen-Shannon Replay

NeurIPS 2026 · Reinforcement learning › RL from feedback · approach: GRPO stability and variants

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DyJR, a replay method for RL with verifiable rewards that reuses past rollouts via a dynamic Jensen-Shannon objective to preserve local policy plasticity and avoid mode collapse.

Problem

On-policy RL algorithms such as GRPO for LLM reasoning are sample-inefficient because they discard past rollouts, and existing experience replay methods that directly reuse accurate samples are computationally costly and cause mode collapse through overfitting.

Approach

Treat historical data as a means to sustain local policy plasticity rather than direct policy updates, using a dynamic Jensen-Shannon replay objective over past rollouts.

Experimental setup

Baselines: GRPO experience replay methods

Builds on: GRPO

Closest papers by meaning

• A Study of Plasticity Loss in On-Policy Deep Reinforcement Learning (NeurIPS 2024 · similarity 0.90)
• Revisiting Plasticity in Visual Reinforcement Learning: Data, Modules and Training Stages (ICLR 2024 · similarity 0.89)
• TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
• LLM-Explorer: A Plug-in Reinforcement Learning Policy Exploration Enhancement Driven by Large Language Models (NeurIPS 2025 · similarity 0.88)
• One-Way Policy Optimization for Self-Evolving LLMs (ICML 2026 · similarity 0.88)
• Multi-Step Likelihood-Ratio Correction for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.88)