Sanjeev Arora

Why is Your Language Model a Poor Implicit Reward Model?

Goedel-Prover-V2: Scaling Formal Theorem Proving with Scaffolded Data Synthesis and Self-Correction

STAT: Skill-Targeted Adaptive Training

On the Impossibility of Retrain Equivalence in Machine Unlearning

Large Language Models as Improvement Operators: Better Reasoning by Iteration

How Does RL Induce Skill Composition? A Case Study Using Countdown

Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors

What Makes a Reward Model a Good Teacher? An Optimization Perspective

On the Power of Context-Enhanced Learning in LLMs

Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization

Goedel-Prover: A Frontier Model for Open-Source Automated Theorem Proving

AdaptMI: Adaptive Skill-based In-context Math Instructions for Small Language Models

Weak-to-Strong Generalization Even in Random Feature Networks, Provably

Provable unlearning in topic modeling and downstream tasks

Low Compute Unlearning via Sparse Representations

Generalizing from SIMPLE to HARD Visual Reasoning: Can We Mitigate Modality Imbalance in VLMs?

Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning

AI-Assisted Generation of Difficult Math Questions