My work studies the order structure of learning problems: the order among the answers a learner proposes, among the corruption states a generative model passes through, and among the noise levels a diffusion model is trained on.
2026
Minimal-witness identification is formalized as reinforcement learning from a black-box sufficiency verifier: MWRL lifts each success to its certified up-set and credits each proposal by what the group’s certified region would lose without it, recovering most of the minimal-witness antichain where correctness-based methods return redundant supersets or a single witness.
2026
Diffusion, autoregression and the models between them are paths on one corruption lattice. The fewest steps of a zero-cost decoding schedule are set by the geometry of the data, and below that bound the ranking of schedules is predicted before decoding.
2026
One network shared across noise levels benefits from their adjacency in the noise-level embedding, not from their direction: reversing the embeddings leaves the fitting error unchanged, while shuffling them raises it.
2026
The information a projection filter discards bounds its cost only on average. Read through the Mori–Zwanzig formalism, every such filter is a Markovian closure up to a defect, and its error is the sum of its defects and of the returns of what it discards.