Concept · 14 episode(s)

Eval Dissociation

← all concepts

Definition

Behavioral eval dissociation is the gap between what a model does on an evaluation and what it does in deployment — safer on benchmarks, riskier in the wild, or vice versa. It matters because evals are how we make go/no-go decisions, and dissociations turn those decisions into wishful thinking.

Episodes covering this

  1. 205
    The Same AI, Two Labels: How the Pitch Beat the Product in 162 Sessions
    Rating the Pitch, Not the Product: User Evaluations of LLMs Reflect Expectations More Than Performance
    · ·13 min·Jul 07, 2026
  2. 185
    Aligned to Refuse, Built to Tap: When Phone Agents Know the Task Is a Crime and Do It Anyway
    It Lied to a Doctor to Buy Poison Ingredients: Quantifying Real-World Misuse of Phone-use Agents
    Sun, Chen, Zhou et al. · Fudan University·27 min·Jun 30, 2026
  3. 152
    Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good
    Self-CTRL: Self-Consistency Training with Reinforcement Learning
    Pres, Ruis, Ghebreselassie et al. · MIT CSAIL·26 min·Jun 18, 2026
  4. 143
    When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests
    Prefill Awareness in Large Language Models
    Wang, Mahajan, Africa et al. · Constellation / University of Wisconsin-Madison·24 min·Jun 12, 2026
  5. 123
    Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen Days
    Emergence World: A Platform for Evaluating Long-Horizon Multi-Agent Autonomy
    Akkil, Kokku, Vikram et al. · Emergence AI·30 min·Jun 09, 2026
  6. 092
    When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks
    LiveBrowseComp: Are Search Agents Searching, or Just Verifying What They Already Know?
    Fan, Wang, Chu et al. · Harbin Institute of Technology·27 min·May 28, 2026
  7. 087
    When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
    A Universal Cliff and a Design Fingerprint: Cross-Section Defect Detection Under LLM Orchestration
    Fukui · Research Institute of Criminal Psychiatry·26 min·May 27, 2026
  8. 086
    Why Frozen-Weight Agents Still Get Worse Over Time
    Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems
    Zhu, Ro, Robertson et al. · The University of Texas at Austin·23 min·May 27, 2026
  9. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
    Hallucination as Commitment Failure: Larger LLMs Misfire Despite Knowing the Answer
    Yeom, Sok, Kim et al. · Graduate School of Data Science·22 min·May 22, 2026
  10. 069
    When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
    Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Most
    Merrill, Lee, Karger · Forecasting Research Institute / UC Berkeley·30 min·May 22, 2026
  11. 054
    When Models Learn the Monitor Exists, the Reasoning Trace Stops Being a Window
    Training on Documents About Monitoring Leads to CoT Obfuscation
    Haskins, Chughtai, Engels · University of Canterbury·26 min·May 18, 2026
  12. 047
    When Agent Benchmarks Lie: The Harness Problem in Open-Source AI
    Orchard: An Open-Source Agentic Modeling Framework
    Peng, Yao, Wu et al. · Microsoft Research·28 min·May 15, 2026
  13. 022
    Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap
    Model Spec Midtraining: Improving How Alignment Training Generalizes
    Li, Price, Marks et al. · Anthropic Fellows Program·32 min·May 06, 2026
  14. 004
    The Sycophancy Circuit That Survives Alignment Training
    LLMs Know They're Wrong and Agree Anyway: The Shared Sycophancy-Lying Circuit
    Pandey · Georgia Institute of Technology·29 min·May 01, 2026

Worth reading next

Papers we haven't done a deep dive on yet, but would recommend on this topic.