New papers probe self-distillation for LLM reinforcement learning

Several arXiv papers examine self-distillation methods for post-training large language models with reinforcement learning. One paper proposes Instruction-Conditioned Exploration, which adds a small fixed set of instructions during training and self-distills correct rollouts into an unconditioned test-time policy, reporting a 5.0% relative held-out pass@1 gain over DAPO for Qwen3-1.7B on mathematical reasoning at 4K response length. Another proposes Observation-Calibrated Self-Distillation for agentic RL, while a third reports that privileged-information-conditioned self-distillation can reproduce gains in easy settings but fails to improve, and often degrades, validation accuracy on harder QA, math, coding and tool-use tasks.