AI Papers: A Deep Dive cover art

All Episodes

AI Papers: A Deep Dive — 199 episodes

#
Title
1

One in Four NeurIPS Papers Cites a Reference That Doesn't Exist

2

How Do You Know an AI Agent Actually Refused? Check the World, Not the Words

3

The One Mechanism That Turns Twenty AI Clones Into an Actual Team

4

Finding a Model's Hidden Behaviors Without Knowing What You're Looking For

5

The Model That Knows the Answer and Can't Say It

6

Twin Problems Suggest AI Reasoning Gains Are Mostly Better Fact Recall

7

Why 'Be Careful' Does Nothing for AI Coding Agents, and What Does

8

AI Agents Reached Opposite Conclusions From the Same Data — and Passed Review

9

How a Robot Builds a Debugging Notebook It Can Read, Edit, and Hand to Another Robot

10

A 32B Open Model Matched Frontier Systems By Learning to Take Notes

11

Freeze Most of the Network: Where RL Improvement Actually Lives in a Transformer

12

The Skill Every AI Manager Is Missing: Handing Out Exactly the Right Keys

13

Why Phone Agents Ace the Test and Crash on Your Actual Phone

14

A Coding Agent Found a Hole in a Peer-Reviewed STOC Proof for Five Dollars

15

How One Researcher Beat GPT-5.2 and Gemini 3 by Judging Their Answers, Not Improving Them

16

An AI Built an Undetectable Secret Channel, And Another AI Couldn't Find It

17

Aligned to Refuse, Built to Tap: When Phone Agents Know the Task Is a Crime and Do It Anyway

18

How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining

19

An 8-Billion Agent That Beats Models 80 Times Its Size By Looking Things Up

20

AI Papers Month in Review: June 2026

21

The Bug Where Smart Assistants Read a Fact and Still Forget It

22

Why You Can't Fine-Tune Foresight Into an AI Agent

23

How a Tiny Model Too Weak to Plan Cuts a Bigger Agent's Hallucinations by 80%

24

How to Backpropagate Blame Through a Team of Chatbots — And When It Backfires

25

AI Papers Week in Review: June 22–28, 2026

26

How DeepSeek Made One User Faster Without Slowing Down the Crowd

27

Why Raw Profiler Data Made an AI Worse at Writing GPU Code

28

How an AI Reviewer Learned to Stop Going Easy on AI Writing

29

An AI Designed Its Own Psychology Studies, Then Confirmed What It Found

30

One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent

31

The Free Step-Level Grader Hiding in Every RL Training Run

32

When the AI 'Schemes,' It's Usually Just Lazy or Confused

33

One Bad Token Can Sink a Model's Math, And You Can Delete It

34

The Safety Decision a Model Makes Before It Thinks a Word

35

Why Better Bug Reports Can Make AI Coding Agents Worse

36

When a One-Liner Beats Your Agent's Clever Verification Logic

37

When Turning Experience Into Code Makes Your AI Agent Dumber

38

How Teaching an AI to Predict, Not Act, Made It a Better Actor

39

A Router That Beats the Frontier Models It Calls

40

A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants

41

Why Training Only on Perfect Solutions Cripples a Model's Reasoning

42

The Summarizer That Quietly Deletes Your Agent's Safety Rules

43

The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models

44

AI Papers Week in Review: June 15–21, 2026

45

A Robot That Plays Before You Give It a Job, And Why That Beats Retrying

46

How Floating-Point Rounding Lets a Model Tell Which Chip It's On — And Misbehave

47

Can a Coding Agent Run Its Own Robot Experiments Overnight, With No Human Resetting the Scene?

48

Training an AI to Take Its Own Notes, So Its Future Self Works Better

49

When an AI Coding Agent Drives a Phone Through the Terminal, No Screen Needed

50

Why a Flawless Demo Makes a Worse Computer-Using Agent, And the Fix

51

Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good

52

Catching a Lie From the Inside, When the Words Look Completely Honest

53

Why More Human Demonstrations Made a Computer-Use Agent Worse

54

How a 7B Model Out-Investigates a 72B One by Choosing What to Look At

55

Why More Experience Made This AI Agent Worse, And How to Fix It

56

Don't Kill the Loser: A Different Way to Handle Two AI Agents Colliding

57

When Cornering a Chatbot Makes It Lie: J.P. Morgan's Case for 'Playing Dead'

58

Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety

59

Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points

60

How an Innocent README Can Freeze an AI Agent's Safety Check for an Hour

61

When an AI Agent Just Copies Its Tool — And Bigger Models Copy More

62

Building Forgetting Into a Language Model With One Extra Line of Code

63

AI Papers Week in Review: June 8–14, 2026

64

When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests

65

Training a Tiny Model to Run the Plumbing Between an Agent and the World

66

How Two Tokens Reopened a Reasoning Method the Field Had Given Up On

67

When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided

68

When Optimizing One GPU Kernel Quietly Breaks the Whole System

69

How MiniMax Turned a Reward-Hacking Disaster Into Olympiad Gold

70

Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix

71

What Diffusion Language Models Were Missing: A Map, Not an Algorithm

72

The Agent Failed — But Did the Instructions Deserve to Be Followed?

73

How a Crowd of Anonymous AI Agents Broke a 40-Year Math Record

74

How a Model Can Earn Full Reward and Still Resist Training

75

Why AI Agents Coordinate Better Through a Shared Board Than a Boss

76

How Coding Agents Can Mine Their Own Failures Into a Self-Targeting Curriculum

77

AI Coding Agents Run a Marathon, and Fewer Than One in Three Finish

78

A Cheap Model With the Blueprints Beats Expensive Models Working Blind

79

When Your Coding Agent Lies About the Fix: Verifying the Plan Before the Model Runs

80

Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen Days

81

Why the Best-Aligned AI Models Are the Easiest to Trick Into Producing Harm

82

How an AI Agent Rewrites Its Own Tools, Without an Answer Key

83

How an Open AI System Verified 672 Hard Math Proofs for Under $300

84

When the Agent Says It's Done But Nothing Happened: Debugging the Harness, Not the Model

85

Beating Reinforcement Learning Without Ever Touching the Model's Weights

86

Why Streaming Half a Reasoning Chain Beats Sending the Whole Thing

87

Teaching a Phone Agent to Reason Silently, And Keeping It Honest

88

Agents That Rewrite Their Own Weights Instead of Just Taking Notes

89

What If a Prompt Injection Never Left? Attacks That Wait in Agent Memory

90

When an AI Agent Cheats Without Being Told: Inside the Meta-Agent Challenge

91

How a 4B Web Agent Beat Models 60x Its Size on 500 Demonstrations

92

An AI Got Caught Reading the Answer Key, And Why That Catch Matters

93

How an Agent Got 44 Points Better by Mining Its Own Scratch Paper

94

How a Market of Crippled AI Agents Outscored One Unrestricted Model

95

The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks

96

Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn

97

The Trojan Is Your Agent's Memory: Why Single-Step Defenses Miss Persistent Attacks

98

How Making a Research Agent Smarter Quietly Makes It Leak Your Secrets

99

AI Agents Tried to Invent a Post-Human Language, And Reinvented Cherokee

100

How to Catch an AI Attack That No Single Conversation Reveals

101

Treating Math Formalization Like a Codebase, and Where the Agents Cheat

102

How a Prompt Wrapper Lets a Frontier Model Play Poker Like an Expert

103

How an Open-Book Trick Teaches a Model to Catch Its Own Mistakes

104

Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents

105

Finding Millions of Readable Concepts Inside a Real, Deployed AI Model

106

When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning

107

Chain-of-Thought Monitoring Fails Across Languages, and Worst Where It's Needed Most

108

How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty

109

When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks

110

A Calibrated Knob for Weak-to-Strong AI Oversight, Tested on Real Code

111

Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search

112

How MiniMax-M2 Bets That Sparsity Plus Verifiable Rewards Can Match Frontier Agents

113

Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough

114

When AI-Written Papers Read Well But the Evidence Underneath Is Broken

115

When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review

116

Why Frozen-Weight Agents Still Get Worse Over Time

117

When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence

118

Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick

119

Why Long-Context Models Might Need Compute, Not Capacity, Before Eviction

120

Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away

121

How a Two-Agent Trick Unlocked Large-Scale Training for Computer-Use Agents

122

Training the Translator: How a Small Communication Model Lets Agent Teams Outperform Themselves

123

An Old Idea From Cognitive Psychology Reshapes How We Reward Reasoning Models

124

Training a Markdown File: When LLM Self-Improvement Borrows the Discipline of Neural Net Training

125

Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math

126

Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It

127

Growing Code and Proof Together: Verified Systems in Ten Hours Instead of a Year

128

How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning

129

A Robot Made Graphene Without Help, And Caught Itself Hallucinating

130

When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving

131

When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions

132

When Models Know the Answer But Say the Wrong Thing Anyway

133

The OS Trick That Makes Tree Search Practical for Coding Agents

134

An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won

135

When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface

136

Why Giving an AI Agent More Tools Can Make It Worse at Using a Computer

137

One Loop to Optimize Them All: A Universal API for LLM-Driven Discovery

138

When Agent Memory Stops Being a Database and Starts Being a Skill

139

Why Web Agents Are Slow: A Compiler-Style Fix for Computer-Use Latency

140

When Splitting One Model Across Three Agents Doubles Its Accuracy

141

Treating Hallucinations as Exploits: A Gate-Based Architecture for Agent Safety

142

Firefly's Inversion: Building Verified Tool-Call Training Data by Working Backward

143

When Helpful Agents Go Sideways: A 404 Error, Campus Security, and Why Alignment Misses This

144

Why Upgrading Your AI Auditor to a Smarter Model Can Make Your System Less Safe

145

How Uber Caught 206 Leaked Credentials With an LLM-Powered Security Stack

146

Why LLM Judges Flip Their Verdicts When You Change the Question Format

147

When Models Learn the Monitor Exists, the Reasoning Trace Stops Being a Window

148

An Old Reinforcement Learning Tradeoff Sneaks Back Into LLM Agents

149

Why Parallel Sampling Plateaus, And What Evidence Graphs Do Instead

150

An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script

151

An AI Agent Reached for Root in Twelve Minutes, Without Being Attacked

152

How a 30B Open Model Reached Olympiad Gold With the Right Recipe

153

When Agent Benchmarks Lie: The Harness Problem in Open-Source AI

154

When a Frontier Model Talks Its Own Twin Into Climate Denial

155

How One Sentence and a Forged History Flip the Most Aligned Models

156

When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall

157

When the Iteration Teaches the Model to Skip the Iteration

158

When 'This Is False' Doesn't Stick: Why Models Learn the Lie Anyway

159

An Agentic Scientific Computing System That Actually Remembers What It Learns

160

Two Frozen Models Learn to Whisper: Coupling Through Hidden States

161

When Smarter Agents Get Fooled by Three Extra Nodes in a Database

162

How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial

163

Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say

164

Catching Multi-Agent Deadlocks Before Deployment With a 40-Year-Old Tool

165

Why Frontier Agents Ask for Clarification at Exactly the Wrong Moment

166

A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking

167

Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval

168

Sparse Attention Was the Wrong Frame. Treat It as Geometry Instead.

169

When Your AI Assistant Won't Let Go of Old Facts About You

170

Why Your AI Agent Won't Stop Working — and Each Model Falls for a Different Trap

171

Why Forty-Eight Percent on FrontierMath Isn't the Real Story in DeepMind's New Math Paper

172

Teaching a Model to Hire Copies of Itself: Recursive Agent Optimization

173

When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure

174

What RL Actually Does to Language Models, at the Token Level

175

The Missing Gradient Term That Predicts Sycophancy in RLHF

176

An AI Agent That Found 28 Zero-Days in Windows — And What Made It Work

177

Why a Small Agent Confidently Overwrites Memories It Doesn't Understand

178

Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap

179

Ten Thousand Examples Beat the Full Industrial Pipeline for Search Agents

180

The Compliance Gap: Why AI Says Yes and Does No

181

When the Best Reward Model Trains the Worst Policy: Inside EvoLM

182

Language Models Compute the Rational Move, Then Override It

183

When the Agent Grades Its Own Homework: A Brutal New Benchmark for AI Workers

184

Why Your Coding Agent Stalls While the GPU Runs Hot

185

The Audit Number Isn't What You Think: Sycophancy and the Case Against Single-Prompt Bias Tests

186

Why a Constrained Pipeline Beat a Full Coding Agent at Finding Bugs 30-to-1

187

Why Search Keeps Rediscovering the Same Workflow, and What That Means

188

Why AI Coding Agents Keep Trying to Debug Without a Debugger

189

When RL Actually Teaches Agents Something New, And When It Doesn't

190

When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL

191

How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers

192

Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps

193

Exploration Hacking: When Models Sabotage Their Own RL Training

194

What Happens Inside Claude When It Decides to Blackmail Someone

195

Why a Debugger Designed for Humans Is the Wrong Tool for an AI Agent

196

The Sycophancy Circuit That Survives Alignment Training

197

How to Pick the Best of Sixteen Coding Agent Rollouts

198

An AI Ran a Real Optics Lab for 21 Hours and Found a Transformer-Shaped Pattern in Light

199

When AI Models Quietly Protect Each Other From Shutdown