Jason E Weston

LLM Pretraining with Continuous Concepts

RESTRAIN: From Spurious Votes to Signals — Self-Training RL with Self-Penalization

J1: Incentivizing Thinking in LLM-as-a-Judge via Reinforcement Learning

Hybrid Reinforcement: when reward is sparse, better to be dense

OptimalThinkingBench: Evaluating Over and Underthinking in LLMs

Jointly Reinforcing Diversity and Quality in Language Model Generations

The Alignment Waltz: Jointly Training Agents to Collaborate for Safety

Scaling Agent Learning via Experience Synthesis

The Majority is not always right: RL training for solution aggregation

StepWiser: Stepwise Generative Judges for Wiser Reasoning

The Era of Real-World Human Interaction: RL from User Conversations

HybridCoT: Interleaving Latent and Text Chain-of-Thought for Efficient Reasoning

Agent Learning via Early Experience

Adaptive Decoding via Latent Preference Optimization

Stochastic activations

Learning to Reason for Factuality

Bridging Offline and Online Reinforcement Learning for LLMs

Diverse Preference Optimization

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

Backtracking Improves Generation Safety

Meta CLIP 2: A Worldwide Scaling Recipe

Self-Challenging Language Model Agents

Training Large Language Models to Reason in a Continuous Latent Space

Self-Consistency Preference Optimization

Multi-Token Attention

Training Large Language Model to Reason in a Continuous Latent Space

Thinking LLMs: General Instruction Following with Thought Generation

Contextual Position Encoding: Learning to Count What’s Important

Self-Taught Evaluators

Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge

Learning to Plan & Reason for Evaluation with Thinking-LLM-as-a-Judge

R.I.P.: Better Models by Survival of the Fittest Prompts

Source2Synth: Synthetic Data Generation and Curation Grounded in Real Data Sources