Glossary · Term

overfitting

← all terms

Definition

Plain language

When a model learns its practice examples too closely and then does worse on anything new.

As stated in the literature

The failure mode where a model fits idiosyncrasies of the training or selection set rather than the underlying signal, degrading held-out performance; countered with regularization, held-out validation, and early stopping.

Also called: overfit, overfits

Why it matters: It matters because a model that clings to its training data looks great in testing yet fails on the new cases you actually care about.

For example, a model that memorizes every practice question aces the practice test but stumbles on the slightly different real exam.

Heard on the show

“Randomize the codes every training step and the code becomes a pure pointer — nothing to overfit.”
Episode 198 — The Model That Knows the Answer and Can't Say It

Mentioned in 20 episodes

  1. 198
    The Model That Knows the Answer and Can't Say It
  2. 193
    Freeze Most of the Network: Where RL Improvement Actually Lives in a Transformer
  3. 191
    How One Researcher Beat GPT-5.2 and Gemini 3 by Judging Their Answers, Not Improving Them
  4. 176
    An AI Designed Its Own Psychology Studies, Then Confirmed What It Found
  5. 156
    Why More Human Demonstrations Made a Computer-Use Agent Worse
  6. 147
    Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points
  7. 146
    How an Innocent README Can Freeze an AI Agent's Safety Check for an Hour
  8. 132
    The Agent Failed — But Did the Instructions Deserve to Be Followed?
  9. 131
    Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix
  10. 128
    How a Model Can Earn Full Reward and Still Resist Training
  11. 125
    AI Coding Agents Run a Marathon, and Fewer Than One in Three Finish
  12. 096
    How Treating an AI Agent's Execution Like Git Recovers a Coordination Penalty
  13. 095
    Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search
  14. 088
    Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
  15. 082
    Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick
  16. 077
    Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
  17. 033
    Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval
  18. 019
    When the Best Reward Model Trains the Worst Policy: Inside EvoLM
  19. 013
    Why Search Keeps Rediscovering the Same Workflow, and What That Means
  20. 011
    When RL Actually Teaches Agents Something New, And When It Doesn't

Related concepts

Related terms