All Episodes

Divide-and-Conquer CoT: RL for Reducing Latency via Parallel Reasoning

Owning the AI Pareto Frontier — Jeff Dean

Learning to Reason in 13 Parameters

Nearly Optimal Active Preference Learning and Its Application to LLM Alignment

Language Model Circuits Are Sparse in the Neuron Basis

Rethinking the Trust Region in LLM Reinforcement Learning

Principled Fine-tuning of LLMs from User-Edits: A Medley of Preference, Supervision, and Reward

Self-distillation enables continual learning

Maximum Likelihood Reinforcement Learning

In-Context Algorithm Emulation in Fixed-Weight Transformers

PPI-SVRG: Unifying Prediction-Powered Inference and Variance Reduction for Semi-Supervised Optimization

When Models Don’t Collapse: On the Consistency of Iterative MLE

An orthogonal learner for individualized outcomes In markov decision processes

Shaping capabilities with token-level data filtering

Self-Improving Pretraining: using post-trained models to pretrain better models

Success Conditioning as Policy Improvement: The Optimization Problem Solved by Imitating Success

Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning

GameTalk: Training LLMs for Strategic Multi-Turn Conversation

Reinforcement Learning via Self-Distillation

Self-Supervised Contrastive Learning is Approximately Supervised Contrastive Learning

On the alignment between supervised and self-supervised contrastive learning

Rethinking the value of multi-agent work-flow: a strong single agent baseline

Greedy Sampling Is Provably Efficient for RLHF

A Generalization Theory for Zero-Shot Prediction

Learning to Discover at Test Time

How Does the Pretraining Distribution Shape In-Context Learning? Task Selection, Generalization, and Robustness

Highlighting What Matters: Promptable Embeddings for Attribute-Focused Retrieval

Activation Reward Models for Few-Shot Model Alignment

Reward is enough: LLMs are in-context reinforcement learners

Understanding the Performance Gap in Preference Learning: A Dichotomy of RLHF and DPO

The End of Reward Engineering: How LLMs Are Redefining Multi-Agent Coordination

PRL: Process Reward Learning Improves LLMs’ Reasoning Ability and Broadens the Reasoning Boundary

Coverage Improvement and Fast Convergence of On-policy Preference Learning

Stagewise Reinforcement Learning and the Geometry of the Regret Landscape

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

Learning Latent Action World Models In The Wild

From Unstructured Data to Demand Counterfactuals: Theory and Practice

In-context reinforcement learning through bayesian fusion of context and value prior

Digital RedQueen: Adversarial Program Evolution in Core War with LLMs

Extending the Context of Pretrained LLMs by Dropping Their Positional Embeddings

Representation-Based Exploration for Language Models: from test-time to post-training

NextFlow: Unified Sequential Modeling Activates Multimodal Understanding and Generation

RelayLLM: Efficient Reasoning via Collaborative Decoding

A Unified Definition of Hallucination, Or: It’s the World Model, Stupid

Deep sequence models tend to memorize geometrically; it is unclear why.

From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence

Diffusion Language Models are Provably Optimal Parallel Samplers

Universal Reasoning Model

Recursive language models

Adapting fast and slow: transportable circuits for few shot learning

Position: Probabilistic Modelling is Sufficient for Causal Inference

End-to-End Test-Time Training for Long Context

Parallel Token Generation for Language Models

Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning

Activation oracles: training and evaluating llms as general-purpose activation explainers

Emergent temporal abstractions in autoregressive models enable hierarchical reinforcement learning

Joint-Embedding vs Reconstruction: Provable Benefits of Latent Space Prediction

Monitoring Monitorability/ OpenAI

Detailed Balance in Large Language Model-Driven Agents

Learning to reason in LLMs by expectation maximization

Exploratory Causal Inference in SAEnce

Detailed balance in large language model-driven agents

The Prism Hypothesis: Harmonizing Semantic and Pixel Representations via Unified Autoencoding

Adaptation of Agentic AI

Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning

Let’s (not) just put things in Context: Test-Time Training for Long-Context LLMs

TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models

What’s In My Human Feedback? Learning Interpretable Descriptions of Preference Data

Bolmo: Byteifying the Next Generation of Language Models

What happened with sparse autoencoders?

What Matters Right Now in Mechanistic Interpretability

CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning

Self-Improving AI and Human Co-Improvement for Safer Co-Superintelligence

Towards a Science of Scaling Agent Systems / Google Deepmind

Emergent hierarchical reasoning in LLMs through reinforcement learning

AI revolution finally comes to Relational foundational models for structured data

REFRAG: Rethinking RAG based Decoding

Provable Long-Range Benefits of Next-Token Prediction

Jeff Dean on TPUs, AI Research, and Funding

Latent Debate: surrogate framework for Interpreting LLM Thinking

Distribution-calibrated inference time compute for thinking llm-as-a-judge

Principled RL for diffusion LLMs emerges from sequence level perspective

Algorithmic Thinking Theory

On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models

Natural language actor-critic: Scalable off-policy learning in language space

Beyond the Transformer: Titans, MIRAS, and the Future of Infinite Context

On the Limits of Test-Time Compute: Sequential Reward Filtering for Better Inference

The Universal Weight Subspace Hypothesis

Stabilizing Reinforcement Learning with LLMs: Formulation and Practices

Benchmarking In-context Experiential Learning Through Repeated Product Recommendations

Training LLMs for Honesty via Confessions

STOIC REASONER: Dual-Mode Transformers that Compress to Think and Decompress to Speak

E-GEO: A Testbed for Generative Engine Optimization in E-Commerce

1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities

Treatment Effect Estimation for Optimal Decision-Making

Pass@K Policy Optimization: Solving Harder Reinforcement Learning Problems

Debugging misaligned completions with sparse-autoencoder latent attribution

Building Effective AI Agents \ Anthropic

How to Correctly Report LLM-as-a-Judge Evaluations

In-Context Learning with Hypothesis-Class Guidance