A research team has built a self-training method for large language models that skips human labels entirely - and claims to avoid the pitfalls that sank earlier attempts.
The framework, called Neuron On-Policy Self-Distillation (Neuron-OPSD), works by inspecting a model's internal neuron activations to decide which training examples are worth learning from and how to construct a stronger "teacher" version of the same model. The model then trains against that teacher's output distribution rather than human-verified answers. No ground-truth labels are required at any stage. The researchers tested it on specialized-domain benchmarks and report that it improves in-domain performance without degrading the model's ability to generalize elsewhere.
That last part is the harder problem. Existing annotation-free self-training approaches tend to trade one failure mode for another: supervised fine-tuning variants hurt out-of-domain performance, while reinforcement learning approaches inflate calibration error, meaning the model becomes overconfident in its own wrong answers. Neuron-OPSD's use of internal activation signals to steer data selection is the novel lever here - it attempts to make self-improvement less of a lottery.
The practical target is narrow but real: domains like medicine or law where expert annotation is expensive and live interaction data is scarce. Whether activation-based data selection holds up outside controlled benchmarks - and at the scale that production deployments demand - remains the open question this paper does not yet answer.