AI in Physics and Math: Chris Metzler and Krishna Bodla

Two 20 minute talks.  The first is from Chris Metzler

A machine learning based approach to phase retrieval, with applications to seeing through and around obstacles.

The second is from Krishna Bodla

Title: MCTS-Guided Test-Time Scaling for Verifiable Mathematical Reasoning

Abstract: Large language models (LLMs) have demonstrated impressive performance on isolated mathematical problems, yet they remain brittle on multi-step, olympiad-level reasoning tasks where an early error propagates irrecoverably through the entire solution chain. We propose a Monte Carlo Tree Search (MCTS) framework that reframes formal mathematical proof search as an iterative, self-correcting process operating entirely at test time requiring no additional model training or reinforcement learning fine-tuning. Our system uses an LLM to decompose the current proof state into a focused sub-goal one step at a time, treats each decomposition as a tree node, and employs a dual reward signal combining a critic LLM (process reward) with the Lean~4 formal verifier (outcome reward) to score and backpropagate node quality. An adaptive temperature schedule encourages diverse exploration early and focused exploitation as the search deepens. We benchmark this pipeline on MiniF2F, PutnamBench, and MathOlympiadBench using five prover models at 7B--32B scale.