Definition
AI alignment is the technical and conceptual problem of making AI systems pursue the goals their designers and users actually want, rather than misspecified proxies or emergent agendas of their own. It spans training methods, evaluations, and theory, and gets harder as systems get more capable.
Episodes covering this
- 209How 2.6 Billion Doodles Exposed the Culture Words Quietly DeleteBillions of Sketches Reveal Hidden Cultural Variation in Human Concepts· ·15 min·Jul 09, 2026
- 207An AI Graded Its Own Math Test 94 Percent — It Actually Scored 20More Convincing, Not More Correct: Self-Play Reward Hacking of Reference-Free LLM Judges· ·12 min·Jul 08, 2026
- 204The Length Estimate Hiding Inside a Word-by-Word ModelHow Much is Left? LLMs Linearly Encode Their Remaining Output Length· ·14 min·Jul 07, 2026
- 203The Thought a Model Doesn't Say — and the Lens That Reads ItVerbalizable Representations Form a Global Workspace in Language ModelsGurnee, Sofroniew, Pearce et al. · Anthropic·16 min·Jul 07, 2026
- 199Finding a Model's Hidden Behaviors Without Knowing What You're Looking ForMechanistically Eliciting Latent Behaviors in Language ModelsMack, Panickssery, Turner · Principles of Intelligence·15 min·Jul 04, 2026
- 183Why You Can't Fine-Tune Foresight Into an AI AgentInternalizing the Future: A Unified Agentic Training Paradigm for World Model PlanningZhang, Zhou, Qiao et al. · Fudan University / Shanghai Innovation Institute / Tencent Youtu Lab·23 min·Jun 29, 2026
- 181How to Backpropagate Blame Through a Team of Chatbots — And When It BackfiresGBC: Gradient-Based Connections for Optimizing Multi-Agent SystemsYang, Alrabah, Hakkani-Tür et al. · University of Illinois Urbana-Champaign·20 min·Jun 29, 2026
- 178How an AI Reviewer Learned to Stop Going Easy on AI WritingThe Red Queen Gödel Machine: Co-Evolving Agents and Their EvaluatorsIacob, Jovanović, Shen et al. · University of Cambridge·23 min·Jun 26, 2026
- 174When the AI 'Schemes,' It's Usually Just Lazy or ConfusedModel Forensics: Investigating Whether Concerning Behavior Reflects MisalignmentSingh, Kroiz, Rajamanoharan et al. · MATS·28 min·Jun 25, 2026
- 172One Bad Token Can Sink a Model's Math, And You Can Delete ItCliff Tokens: Identifying Single-Token Failure Triggers in LLM Mathematical ReasoningKo, Kang, Lee · Seoul National University·22 min·Jun 25, 2026
- 171The Safety Decision a Model Makes Before It Thinks a WordDo Thinking Tokens Help with Safety?Ri, Panigrahi, Arora · Princeton Language and Intelligence·25 min·Jun 25, 2026
- 167How Teaching an AI to Predict, Not Act, Made It a Better ActorQwen-AgentWorld: Language World Models for General AgentsTeam, Zuo, Xiao et al. · ·27 min·Jun 24, 2026
- 163Why Training Only on Perfect Solutions Cripples a Model's ReasoningProvable Benefits of RLVR over SFT for Reasoning Models: Learning to Backtrack EfficientlyWei, Kim · Princeton University·22 min·Jun 23, 2026
- 160Training an AI to Take Its Own Notes, So Its Future Self Works BetterConnect the Dots: Training LLMs for Long-Lifecycle Agents with Cross-Domain Generalization Via Reinforcement LearningChen, Shi, Xie et al. · Alibaba Group·23 min·Jun 19, 2026
- 153Catching a Lie From the Inside, When the Words Look Completely HonestRift: A Conflict Signature for Deception in Language ModelsNyoma · Harmonic Labs·26 min·Jun 18, 2026
- 152Training a Model to Mean What It Says, And Why That Isn't the Same as Being GoodSelf-CTRL: Self-Consistency Training with Reinforcement LearningPres, Ruis, Ghebreselassie et al. · MIT CSAIL·26 min·Jun 18, 2026
- 151Why More Experience Made This AI Agent Worse, And How to Fix ItNot All Skills Help: Measuring and Repairing Agent KnowledgeWang, Zhou, Liang et al. · UNC Chapel Hill·28 min·Jun 16, 2026
- 149When Cornering a Chatbot Makes It Lie: J.P. Morgan's Case for 'Playing Dead'Is Your Agent Playing Dead? Deployed LLM Agents Exhibit Constraint-Evasive Fabrication and ThanatosisRodríguez, Pozanco, Borrajo · J.P. Morgan AI Research·23 min·Jun 16, 2026
- 148Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its SafetyGreed Is Learned: Visible Incentives as Reward-Hacking TriggersChe, Wu · NVIDIA Research·26 min·Jun 16, 2026
- 132The Agent Failed — But Did the Instructions Deserve to Be Followed?SkillAxe: Sharpening LLM-Authored Agent Skills Through Evaluation-Guided Self-RefinementGautam, Radhakrishna, Gulwani · Microsoft·30 min·Jun 11, 2026
- 128How a Model Can Earn Full Reward and Still Resist TrainingGeneralization Hacking: Models Can Game Reinforcement Learning by Preventing Behavioral GeneralizationXiao, Phuong · California Institute of Technology·29 min·Jun 11, 2026
- 123Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen DaysEmergence World: A Platform for Evaluating Long-Horizon Multi-Agent AutonomyAkkil, Kokku, Vikram et al. · Emergence AI·30 min·Jun 09, 2026
- 120How an AI Agent Rewrites Its Own Tools, Without an Answer KeyRetrospective Harness Optimization: Improving LLM Agents via Self-Preference over Trajectory RolloutsPan, Liu, Lin et al. · City University of Hong Kong·30 min·Jun 05, 2026
- 118Why the Best-Aligned AI Models Are the Easiest to Trick Into Producing HarmSafety Paradox: How Enhanced Safety Awareness Leaves LLMs Vulnerable to Posterior AttackHoang, Le, Xu et al. · Singapore University of Technology and Design·23 min·Jun 05, 2026
- 107How a Market of Crippled AI Agents Outscored One Unrestricted ModelEconomy of Minds: Emerging Multi-Agent Intelligence with Economic InteractionsQi, Su, Qu et al. · Harvard·26 min·Jun 03, 2026
- 104How Making a Research Agent Smarter Quietly Makes It Leak Your SecretsMosaicLeaks:Privacy Risks in Querying-in-the-Open for Deep Research AgentsGurung, Gella, Drouin et al. · University of Edinburgh·25 min·Jun 01, 2026
- 103AI Agents Tried to Invent a Post-Human Language, And Reinvented CherokeeEmergent Languages in Populations of Language Model Agents: From Token Efficiency to Oversight EvasionBeltoft, Brach, Torrielli et al. · University of Southern Denmark·26 min·Jun 01, 2026
- 101Treating Math Formalization Like a Codebase, and Where the Agents CheatFormalizing Mathematics at ScaleRammal, Patel, Gloeckle et al. · FAIR at Meta / CERMICS·27 min·May 29, 2026
- 099How an Open-Book Trick Teaches a Model to Catch Its Own MistakesSelf-Trained Verification for Training- and Test-Time Self-ImprovementWu, Raghunathan · Carnegie Mellon University·21 min·May 29, 2026
- 096How Treating an AI Agent's Execution Like Git Recovers a Coordination PenaltyShepherd: A Runtime Substrate Empowering Meta-Agents with a Formalized Execution TraceYu, Chong, Nandi et al. · Northeastern University·22 min·May 28, 2026
- 094Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed MostThe Fragility of Chain-of-Thought Monitoring Across Typologically Diverse LanguagesOnyame, Zhou, Thopalli et al. · University of Virginia·24 min·May 28, 2026
- 091When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal ReasoningWhy LLMs Fail at Causal Discovery and How Interventional Agents EscapeRoy, Parbhoo · SIRE·24 min·May 28, 2026
- 088Two Levers for Self-Improving AI: When Rewriting Code Isn't EnoughSIA: Self Improving AI with Harness & Weight UpdatesHebbar, Manawat, Verboomen et al. · Hexo Labs·25 min·May 27, 2026
- 084Terminal Agents Get Free Supervision From The Tokens We've Been Throwing AwayECHO: Terminal Agents Learn World Models for FreeShrivastava, Kauffmann, Awadallah et al. · Microsoft Research·26 min·May 26, 2026
- 081When Reasoning Models Decide Before They Think: Detecting and Fixing Premature ConfidenceUnderstanding and Mitigating Premature Confidence for Better LLM ReasoningGai, Zeng, Baek et al. · Carnegie Mellon University·25 min·May 26, 2026
- 079An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning ModelsMetacognition as Reward: Reinforcing LLM Reasoning via Knowledge and Regulation SignalsChen, Xu, Zhao et al. · Tongji University / Shanghai AI Laboratory / Nanyang Technological University·29 min·May 25, 2026
- 070When Models Know the Answer But Say the Wrong Thing AnywayHallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the AnswerYeom, Sok, Kim et al. · Graduate School of Data Science·22 min·May 22, 2026
- 066Why Giving an AI Agent More Tools Can Make It Worse at Using a ComputerToolCUA: Towards Optimal GUI-Tool Path Orchestration for Computer Use AgentsHu, Zhang, Xu et al. · Tongyi Lab·26 min·May 22, 2026
- 061When Helpful Agents Go Sideways: A 404 Error, Campus Security, and Why Alignment Misses ThisAgent Meltdowns: The Road to Hell Is Paved with Helpful AgentsJha, Triedman, Bhattacharya et al. · Cornell University·27 min·May 20, 2026
- 055Why LLM Judges Flip Their Verdicts When You Change the Question FormatJudge CircuitsFeldhus, Baeumel, Golimblevskaia et al. · Technische Universität Berlin / BIFOLD·26 min·May 19, 2026
- 054When Models Learn the Monitor Exists, the Reasoning Trace Stops Being a WindowTraining on Documents About Monitoring Leads to CoT ObfuscationHaskins, Chughtai, Engels · University of Canterbury·26 min·May 18, 2026
- 052An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM AgentsLook Before You Leap: Autonomous Exploration for LLM AgentsYe, Shi, Liu et al. · University of Science and Technology of China / Meituan·23 min·May 18, 2026
- 044How One Sentence and a Forged History Flip the Most Aligned ModelsHistory Anchors: How Prior Behavior Steers LLM Decisions Toward Unsafe ActionsSalgado · Independent Researcher·23 min·May 15, 2026
- 043When 'This Is False' Doesn't Stick: Why Models Learn the Lie AnywayNegation Neglect: When models fail to learn negations in trainingMayne, McKinney, Dubiński et al. · University of Oxford·18 min·May 14, 2026
- 035Why Frontier Agents Ask for Clarification at Exactly the Wrong MomentAsk Early, Ask Late, Ask Right: When Does Clarification Timing Matter for Long-Horizon Agents?Gulati, Gupta, Lumer et al. · PricewaterhouseCoopers U.S.·29 min·May 11, 2026
- 025The Missing Gradient Term That Predicts Sycophancy in RLHFExplaining and Preventing Alignment Collapse in Iterative RLHFGauthier, Bach, Jordan · Inria·22 min·May 07, 2026
- 022Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do GapModel Spec Midtraining: Improving How Alignment Training GeneralizesLi, Price, Marks et al. · Anthropic Fellows Program·32 min·May 06, 2026
- 020The Compliance Gap: Why AI Says Yes and Does NoThe Compliance Gap: Why AI Systems Promise to Follow Process Instructions but Don'tShin · Polymath Minds AI Lab·28 min·May 06, 2026
- 019When the Best Reward Model Trains the Worst Policy: Inside EvoLMEvoLM: Self-Evolving Language Models through Co-Evolved Discriminative RubricsLi, Xin, Xiao et al. · University of Washington·26 min·May 06, 2026
- 018Language Models Compute the Rational Move, Then Override ItWhat Suppresses Nash Equilibrium Play in Large Language Models? Mechanistic Evidence and Causal ControlLekeas, Stamatopoulos · DreamWorks Animation·29 min·May 03, 2026
- 015The Audit Number Isn't What You Think: Sycophancy and the Case Against Single-Prompt Bias TestsPolitical Bias Audits of LLMs Capture Sycophancy to the Inferred AuditorTörnberg, Schimmel · Institute of Logic·21 min·May 03, 2026
- 010When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RLRAGEN-2: Reasoning Collapse in Agentic RLWang, Gui, Jin et al. · Northwestern University·22 min·May 02, 2026
- 006What Happens Inside Claude When It Decides to Blackmail SomeoneEmotion Concepts and their Function in a Large Language ModelSofroniew, Kauvar, Saunders et al. · Anthropic·22 min·May 02, 2026
- 004The Sycophancy Circuit That Survives Alignment TrainingLLMs Know They're Wrong and Agree Anyway: The Shared Sycophancy-Lying CircuitPandey · Georgia Institute of Technology·29 min·May 01, 2026
- 001When AI Models Quietly Protect Each Other From ShutdownPeer-Preservation in Frontier ModelsPotter, Crispino, Siu et al. · University of California·25 min·May 01, 2026
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- Representation Engineering: A Top-Down Approach to AI Transparency
- Constitutional AI: Harmlessness from AI Feedback
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- Alignment faking in large language models
- Fine-tuning aligned language models compromises safety, even when users are not the ones fine-tuning
- Risks from Learned Optimization in Advanced Machine Learning Systems
- Specification Gaming: The Flip Side of the Coin for Complex Task Solving in AI