Jack Lanchantin

LLM Pretraining with Continuous Concepts

LLM Output Homogenization is Task Dependent

OptimalThinkingBench: Evaluating Over and Underthinking in LLMs

Jointly Reinforcing Diversity and Quality in Language Model Generations

Bridging Offline and Online Reinforcement Learning for LLMs

Adaptive Decoding via Latent Preference Optimization

Diverse Preference Optimization

CoT-Self-Instruct: Building high-quality synthetic prompts data for reasoning and non-reasoning tasks

Compositional Interfaces for Compositional Generalization