A new training technique lets an AI model learn from a version of itself that hasn't been built yet.
Researchers describe Bootstrapped On-Policy Self-Distillation, or B-OPSD, in a new paper. The method temporarily pushes a language model's training forward, freezes that more-advanced checkpoint as a 'future teacher,' then rewinds the actual model back to its starting point and has the future teacher supervise it token by token. It builds on on-policy self-distillation, where a more-informed version of a model helps train a less-informed one on its own generated outputs. The team tested B-OPSD on Qwen3-4B and Qwen3-8B, two open-source language models, on math reasoning problems. In the paper's rollout-privileged setting - where the teacher gets extra information while generating example solutions - accuracy climbed from 27.50 percent to 41.30 percent on the 4-billion-parameter model, and from 48.80 percent to 64.44 percent on the 8-billion-parameter model, compared with standard self-distillation.
That's a real jump for a technique that adds no new external data - it just recycles the model's own optimization progress. If a model can reliably distill its future gains back into its present state, it chips away at the assumption that self-improvement needs bigger models, more compute, or fresh human-labeled feedback to keep climbing.
Math problems are also the easiest place for a trick like this to shine, since answers are checkable and rewards are unambiguous. Whether the same bootstrapping holds up on messier, harder-to-grade tasks is the real test.