All episodes

Efficient Test-Time Scaling via Self-Calibration

Conformal Prediction via Bayesian Quadrature

Predicting from Strings: Language Model Embeddings for Bayesian Optimization

Self-Evolving Curriculum for LLM Reasoning

Online Decision-Focused Learning in Dynamic Environments

FisherSFT: Data-Efficient Supervised Fine-Tuning of Language Models Using Information Gain

Reward Shaping from Confounded Offline Data

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

Understanding Best-of-N Language Model Alignment

Maximizing Acquisition Functions for Bayesian Optimization - and its relation to Gradient Descent

Bayesian Prompt Ensembles: Model Uncertainty Estimation for Black-Box Large Language Models

Prompting Strategies for Enabling Large Language Models to Infer Causation from Correlation

The Parallel Knowledge Gradient Method for Batch Bayesian Optimization

FunBO: Discovering Acquisition Functions for Bayesian Optimization with FunSearch

Automated Social Science: A Structural Causal Model-Based Approach

Causal Interpretation of Transformer Self-Attention

A Causal World Model Underlying Next Token Prediction: Exploring GPT in a Controlled Environment

Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs

Adaptive Inference-Time Compute: LLMs Can Predict if They Can Do Better, Even Mid-Generation

Prompts from Reinforcement Learning (PRL)

Logits are All We Need to Adapt Closed Models

Large Language Models Are (Bayesian) Latent Variable Models: Explaining and Finding Good Demonstrations for In-Context Learning

Inference-Time Intervention: Eliciting Truthful Answers from a Language Model

From Decoding to Meta-Generation: Inference-time Algorithms for Large Language Models

LLM In-Context Learning as Kernel Regression

Personalizing LLMs via Decode-Time Human Preference Optimization

Almost Surely Safe LLM Inference-Time Alignment

Survey of In-Context Learning Interpretation and Analysis

From Decoding to Meta-Generation: Inference-time Algorithms for Large Language Models

LLM In-Context Learning as Kernel Regression

Where does In-context Learning Happen in Large Language Models?

Auto-Differentiating Any LLM Workflow: A Farewell to Manual Prompting

metaTextGrad: Learning to learn with language models as optimizers

Semantic Operators: A Declarative Model for Rich, AI-based Data Processing

Isolated Causal Effects of Language

Sleep-time Compute: Beyond Inference Scaling at Test-time

J1: Incentivizing Thinking in LLM-as-a-Judge

ShiQ: Bringing back Bellman to LLMs

Policy Learning with a Natural Language Action Space: A Causal Approach

Multi-Objective Preference Optimization: Improving Human Alignment of Generative Models

End-to-End Learning for Stochastic Optimization: A Bayesian Perspective

TEXTGRAD: Automatic Differentiation via Text

Steering off Course: Reliability Challenges in Steering Language Models

Past-Token Prediction for Long-Context Robot Policies

Recovering Coherent Event Probabilities from LLM Embeddings

Systematic Meta-Abilities Alignment in Large Reasoning Models

Predictability Shapes Adaptation: An Evolutionary Perspective on Modes of Learning in Transformers

Efficient Exploration for LLMs

Rankers, Judges, and Assistants: Towards Understanding the Interplay of LLMs in Information Retrieval Evaluation

Bayesian Concept Bottlenecks with LLM Priors

Transformers for In-Context Reinforcement Learning

Evaluating Large Language Models Across the Lifecycle

Active Ranking from Human Feedback with DopeWolfe

Optimal Designs for Preference Elicitation

Dual Active Learning for Reinforcement Learning from Human Feedback

Active Learning for Direct Preference Optimization

Active Preference Optimization for RLHF

Test-Time Alignment of Diffusion Models without reward over-optimization

Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback

GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-time Alignment

Advantage-Weighted Regression: Simple and Scalable Off-Policy RL

Can RLHF be More Efficient with Imperfect Reward Models? A Policy Coverage Perspective

Transformers can be used for in-context linear regression in the presence of endogeneity

Bayesian Concept Bottlenecks with LLM Priors

In-Context Parametric Inference: Point or Distribution Estimators?

Enough Coin Flips Can Make LLMs Act Bayesian

Bayesian Scaling Laws for In-Context Learning

Posterior Mean Matching Generative Modeling

Can Generative AI Solve Your In-Context Learning Problem? A Martingale Perspective

Dynamic Search for Inference-Time Alignment in Diffusion Models

Is In-Context Learning in Large Language Models Bayesian? A Martingale Perspective

Leaked Claude Sonnet 3.7 System Instruction tuning

Converging Predictions with Shared Information

Test-Time Alignment Via Hypothesis Reweighting

Rethinking Diverse Human Preference Learning through Principal Component Analysis

Active Statistical Inference

Data Mixture Optimization: A Multi-fidelity Multi-scale Bayesian Framework

AI-Powered Bayesian Inference

Can Unconfident LLM Annotations Be Used for Confident Conclusions?

Predictions as Surrogates: Revisiting Surrogate Outcomes in the Age of AI

Learn then Test: Calibrating Predictive Algorithms to Achieve Risk Control

How to Evaluate Reward Models for RLHF

LLMs as Judges: Survey of Evaluation Methods

The Alternative Annotator Test for LLM-as-a-Judge: How to Statistically Justify Replacing Human Annotators with LLMs

Limits to scalable evaluation at the frontier: LLM as Judge won’t beat twice the data

Stratified Prediction-Powered Inference for Hybrid Language Model Evaluation

Accelerating Unbiased LLM Evaluation via Synthetic Feedback

Prediction-Powered Statistical Inference Framework

Optimizing Chain-of-Thought Reasoners via Gradient Variance Minimization in Rejection Sampling and RL

RM-R1: Reward Modeling as Reasoning

Reexamining the Aleatoric and Epistemic Uncertainty Dichotomy

Decoding Claude Code: Terminal Agent for Developers

Emergent Strategic AI Equilibrium from Pre-trained Reasoning

Benefiting from Proprietary Data with Siloed Training

Advantage Alignment Algorithms

Asymptotic Safety Guarantees Based On Scalable Oversight

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

Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems

Identifiable Steering via Sparse Autoencoding of Multi-Concept Shifts

You Are What You Eat - AI Alignment Requires Understanding How Data Shapes Structure and Generalisation