Instruction-Conditioned Exploration Enhances RL for LLMs

A new paper on arXiv (2608.02087) introduces Instruction-Conditioned Exploration (ICE), a method to improve reinforcement learning (RL) for large language models (LLMs). The authors propose supplementing task prompts with distinct instructions during training to increase behavioral coverage, and combine RL with self-distillation (Asymmetric-RL/SD) to transfer explored behaviors to the test-time policy. Experiments on Qwen3-1.7B show improved held-out pass@1 performance at 4K response length on mathematical reasoning tasks. The paper addresses challenges in LLM action-space structure that differ from classical RL, emphasizing the need for deliberate exploration leveraging pre-trained knowledge.