New papers test self-distillation for LLM reinforcement learning

Several new arXiv papers examine self-distillation methods for reinforcement learning and agent training in large language models. One paper proposes Instruction-Conditioned Exploration, which adds fixed training instructions and distills correct rollouts into an unconditioned policy, reporting a 5.0% relative held-out pass@1 gain over DAPO for Qwen3-1.7B on math reasoning at 4K response length, but no gain for Qwen3-4B at 4K. Another proposes Observation-Calibrated Self-Distillation for agentic RL, while a separate study argues that privileged-information self-distillation alone failed to improve and often degraded validation accuracy on harder QA, math, coding and tool-use tasks.