8 papers
Finetuning with Sampling: SFT Learns Better Than You Think
Aayush Karan, Sitan Chen, Yilun Du
Introducing new capabilities to frontier models has long been the goal of posttraining, which predominantly employs supervised finetuning (SFT) and reinforcement learning (RL) to t…
Reasoning with Sampling: Your Base Model is Smarter Than You Think
Aayush Karan, Yilun Du
Frontier reasoning models have exhibited incredible capabilities across a wide array of disciplines, driven by posttraining large language models (LLMs) with reinforcement learning…
ReGuidance: A Simple Diffusion Wrapper for Boosting Sample Quality on Hard Inverse Problems
Aayush Karan, Kulin Shah, Sitan Chen
There has been a flurry of activity around using pretrained diffusion models as informed data priors for solving inverse problems, and more generally around steering these models u…
When does a predictor know its own loss?
Aravind Gollakota, Parikshit Gopalan, Aayush Karan +2
Given a predictor and a loss function, how well can we predict the loss that the predictor will incur on an input? This is the problem of loss prediction, a key computational task…
Blink of an eye: a simple theory for feature localization in generative models
Marvin Li, Aayush Karan, Sitan Chen
Large language models can exhibit unexpected behavior in the blink of an eye. In a recent computer use demo, a language model switched from coding to Googling pictures of Yellowsto…
Unrolled denoising networks provably learn optimal Bayesian inference
Aayush Karan, Kulin Shah, Sitan Chen +1
Much of Bayesian inference centers around the design of estimators for inverse problems which are optimal assuming the data comes from a known prior. But what do these optimality g…
Tridiagonal pairs of -Serre type and their linear perturbations
Aayush Karan
A tridiagonal pair is an ordered pair of diagonalizable linear maps on a nonzero finite-dimensional vector space, that each act on the eigenspaces of the other in a block-tridiagon…
Generating Set for Nonzero Determinant Links Under Skein Relation
Aayush Karan
Traditionally introduced in terms of advanced topological constructions, many link invariants may also be defined in much simpler terms given their values on a few initial links an…