Self-training AI models have a cheating problem, and a new paper proposes a fix.
When a large language model acts as its own teacher — a technique called on-policy self-distillation — it can accidentally pass hints to the student version of itself that won't exist at test time. The student learns to exploit those hints rather than reason from scratch, a flaw the researchers call privileged information leakage. DemoPSD, introduced in a preprint this week, addresses this by selectively filtering how much teacher guidance the student absorbs at each token position, using a blended target distribution that keeps the student's own reasoning intact. Tested across four scientific fields on the SciKnowEval benchmark, it outperformed both GRPO and SDPO while generalizing better to out-of-distribution questions on the GPQA benchmark.
The detail worth noting is what this reveals about the self-distillation trend itself. Training a single model to teach itself is appealing because it cuts costs and sidesteps the need for separate teacher models — but dense token-level supervision, the researchers argue, quietly trades generalization for in-domain pattern matching. DemoPSD's fix is not to abandon self-distillation but to make the student more skeptical of its teacher.
This lands in a crowded field of reasoning-improvement techniques — RLHF, chain-of-thought fine-tuning, process reward models — all competing to close the gap between benchmark performance and real-world reliability. Whether selectively ignoring your own teacher scales beyond scientific question-answering remains the obvious open question.