All Episodes
Best AI papers explained — 841 episodes · page 1 of 9
Scaling Empowered Agents
Hybrid Reasoning Agents
Strategic Self-Consistency
Retrofitting language models to operate over bytes
Priced Guidance: Can Language Models Generate Future Research Ideas?
Context Language Models
Position: AI-Agent Pricing Should Become More Outcome-Dependent: An Economic Perspective
Emergent Capabilities Arise Randomly from Learning Sparse Attention Patterns
Pragmatic Causal Inference with AI-Learned Representations
Pointwise or pairwise: when do pairwise losses help reward learning, provably?
When to Trust the Cheap Check: Weak and Strong Verification for Reasoning
Theoretical Limits of Language Model Alignment
DT2: Decision-Targeted Digital Twins
Self-Play Pretraining with Zero Data
Language models need sleep: learning to self-modify and consolidate memories
Jev Creator: System One models for Prod, not God
Detecting and countering misuse of AI: September 2026
Position: LLMs can’t jump
Jailbreaking Jailbreaks: A Proactive Defense for LLMs
When Agents Slow Down: Understanding LLM Agents’ Test-Time Strategies via Elo-per-token Analysis
Thinking with Looped Flows
Multi-Turn LLM Conversations under the Least-Recently-Used Policy: Mean-Field Asymptotics
Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models
Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs
Tail-Likelihood Reinforcement Learning
Next-Latent Prediction Transformers Learn Compact World Models
Language Models Can Control Their Own Attention
AI Finds A Way
Superposition Without Interference? Towards Isolated Interventions via Almost Orthogonal Features in Language Models
TTPO: Test-Time Policy Optimization
Demystifying Reinforcement Learning Post-Training of Language Models
Recursive Experiential–Working Memory Evolution for Long-Horizon Agent Harnesses
TailSFT: Filtered Fine-Tuning Improves Post-Training Performance
SPADE: Self-Play in Adaptive Synthetic Executable Environments
Evaluating Skills, Not Just Agents: Agentic Continuous Evaluation of Skills
Impression Share Prediction: An Offline Evaluation Task for Ranking Systems
Q-Learning with World Models
Conformal Language Modeling via Posterior Sampling
BoNVoyage: Learning Better Rewards without Ranking
Demystifying Agent Skills: Why They Work—Until They Don’t
Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence
Predicting Neural Scaling Laws without Training: A Data Manifold Oracle
Making Interpretable Discoveries from Unstructured Data: A High-Dimensional Multiple Hypothesis Testing
Overcoming the Incentive Collapse Paradox
Position: Modular Memory is the Key to Continual Learning Agents
Harness RL is Meta-Learning: Training to Self-Improve at Test Time
Escaping the Nash Trap: Structural Estimation and Alignment of Strategic Reasoning in Large Language Models
When Does LeJEPA Learn a World Model?
Do Modules Stay in Their Lane? Role Drift in Compound LLM Systems
Do you really need to pretrain Q-functions for online RL fine-tuning?
The Evolution of Digital Search: From Blue Links to Delegated Decision-Making
Ask, Don’t Judge: Binary Questions for Interpretable LLM Evaluation and Self-Improvement
From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning
Understanding Reasoning from Pretraining to Post-Training
A Positive Case for Faithfulness: LLM Self-Explanations Help Predict Model Behavior
Reject, Resample, Repeat: Understanding Parallel Reasoning in Language Model Inference
Rethinking the Evaluation of Harness Evolution for Agents
From Reasoning Traces to Reusable Modules: Understanding Compositional Generalization in Language Model Reasoning
Position: Interpretability can be actionable
High-accuracy sampling for diffusion models and log-concave distributions
Causal Inference with Video Features as Treatments
What Does Thompson Sampling Optimize?
Globally Convergent Offline Reinforcement Learning with Smoothed Bellman Residual Minimization
LLM-as-a-Verifier: A General-Purpose Verification Framework
How Much Do Language Models Memorize?
Position: Uncertainty Quantification in LLMs is Just Unsupervised Clustering
Position: Agents Should Invoke External Tools ONLY When Epistemically Necessary
From conversations to mechanisms: aligning advertiser Incentives in ai-powered product recommendations
Is one layer enough? Training a single transformer layer can match full-parameter RL training
RL Excursions during Pre-Training: Re-examining Policy Optimization for LLM training
Language Generation with Feedback: Queries and Mistakes
Quantifying Theoretical AI Alignment Guarantees: Receiver-Utility Bounds in Bayesian Persuasion
SPIRAL: Learning to search and aggregate
Qwen-AgentWorld: Language World Models for General Agents
When Does Trajectory-Level Supervision Permit Efficient Offline Reinforcement Learning?
SuperThoughts: Reasoning Tokens in Superposition
First-Explore PPO : Learning Meta-Exploration with Proximal Policy Optimization
Self-Distillation for Data-Scarce Language Model Pretraining
Meta-Harness for Agent-State Construction
ExpRL: Using Reference Solutions as Rewards for LLM Mid-Training
Valid Inference with Synthetic Data via Task Exchangeability
GRPO is Secretly a Process Reward Model
Agentic Interactions
A Unifying View of Attention Sinks: Two Algorithms, Two Solutions
From AGI to ASI
Correct Looks Better: Pairwise Comparisons Reveal Accuracy Rankings
Critical Batch Size for LLM Policy Optimization
Self-supervised User Profile Generation for Personalization
From Augmentation to Reconstruction: Guiding the AI Disruption to the Good Place
Self-Distilled Agentic Reinforcement Learning
Subliminal Learning Is Steering Vector Distillation
Subsidizing Sequential Search
Meta-Harness: End-to-End Optimization of Model Harnesses
Self-Improving Language Models with Bidirectional Evolutionary Search
Generative Modeling via Drifting
Instance-Optimal Estimation with Multiple LLM Judges on a Budget
Robust AI Personalization Will Require a Human Context Protocol
Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning
Position: The Pre/Post-Training Boundary Should Govern IP in Industry–Academia ML Collaborations
MEMO: Memory as a Model