1/ Autoregressive LMs are born left-to-right: token 1 → token 2 → token 3 → … This makes generation—and often reasoning—an intrinsically serial process. Diffusion LMs are different: they start from masked text and iteratively reveal tokens. That creates a new freedom: the order of thought.
7/ This is like process supervision for diffusion LMs— but without human-written reasoning traces. Instead of only rewarding the final answer, SAS gives dense feedback over the whole reveal trajectory. It learns a model-native curriculum: commit the helpful parts early, postpone ambiguity.
10/ The gains transfer beyond Sudoku. With LLaDA-8B: GSM8K pass@1: 64% → 76% MBPP pass@1: 39.5% → 41% And the learned scheduler remains strong across generation lengths and semi-autoregressive/block decoding settings. https://x.com/furongh/status/2074147811856470483/photo/1
6/ The key idea: make the scheduler self-aware. SAS rewards an unmasking order by the frozen model’s own pathwise likelihood: “Along this route, how well can I explain the target?” So the model learns an order aligned with its own predictive strengths.
Ask a question below.
Published answers will appear here.