Self-Distillation Fails on Difficult Tasks Despite PI Conditioning

A new arXiv preprint (2608.04794) reveals that self-distillation (SD), a compute-efficient alternative to reinforcement learning with verifiable rewards, fails to improve validation accuracy on difficult tasks. The study reproduces SDPO's reported gains in easy settings but finds that the identical setup does not work for challenging tasks across question answering, mathematics, coding, and multi-turn agentic tool use. Despite steady decreases in per-token loss, validation accuracy does not improve and typically degrades. The authors propose a single causal chain to explain this failure, stemming from the privileged information (PI) conditioning of the self-teacher. The research spans various reasoning modes, model sizes, and forms of PI, and tests both SDPO and OPSD recipes. The findings challenge the efficacy of SD as a standalone objective without a reward term, raising questions about its applicability to real-world, high-difficulty problems. The paper is available on arXiv under the title 'Privileged, but Biased: How PI-Conditioned Teachers Break Self-Distillation'.