We propose a sampling algorithm for base models that gives single-shot reasoning boosts on par with RL-posttraining, without compromising generation diversity and multi-shot (pass@k) performance. Crucially, our method is training-free, dataset-free, and verifier-free.

The field of RNA structure prediction has long lacked a reliable, flexible, and self-contained method to detect the presence of key 3D building blocks in folded RNAs. We provide an algorithm that does exactly this, predicting a near exhaustive set of 3D configurations (everything), at any location (everywhere), fully end-to-end over multiple structural hierarchies (all-at-once).

We discover that strong latent initializations in noise space offer a new axis for desiging inference-time algorithms that steer diffusion models towards reward functions. We propose a simple algorithm (ReGuidance) leveraging this insight and both empirically and theoretically demonstrate its superiority over prior techniques.

We present a unifying theory of LLMs and diffusion models explaining the sudden emergence of high-level semantic features during generation. We introduce the notion of critical windows for LLMs and empirically correlate them to reasoning failures.

We present a theory of loss prediction in machine learning and establish an equivalence with algorithmic fairness. In particular, we prove that for binary classifiers, we can learn a nontrivial loss predictor only when the base model is not multicallibrated.

We propose a theory-inspired neural network architecture for solving inverse problems that provably learns optimal Bayesian inference. Our network mirrors diffusion models by learning noisy data priors and runs a Bayes-optimal algorithm on top. For the first time in the literature, we prove that the score-matching objective is learnable in one dimension.

We examine a pair of linear maps with a particular property (tridiagonal pairs) inspired from statistical physics, and we demonstrate that linear perturbations preserve this property under a very simple polynomial condition.

We resolve an open problem proposed by David Mullins in 1993 regarding the recursive computability of crucial topological quantities known as nonzero determinant link invariants.