Glossary · Term

evaluator

← all terms

Definition

Plain language

A function that scores how good a candidate solution is and what was wrong with it.

As stated in the literature

In optimize-anything, a domain-specific function returning both a scalar score and structured side information about failures, replacing the gradient signal used in numerical optimization.

Also called: evaluators

Why it matters: Structured feedback about why a candidate failed lets a search procedure improve directionally, which is the difference between blind sampling and meaningful optimization.

For example, when optimizing a SQL query, the evaluator might return a runtime score plus a note saying 'failed on rows with NULL in column X.'

Heard on the show

“But that categorization came from the same commercial evaluator the paper elsewhere shows is twenty-five percent self-consistent on open-format failures.”
Episode 233 — Why a Model Can Grade an Answer But Not Write the Answer Key

Mentioned in 22 episodes

  1. 233
    Why a Model Can Grade an Answer But Not Write the Answer Key
  2. 232
    Coding Models Can Find the Bad Line, They Just Won't Delete It
  3. 178
    How an AI Reviewer Learned to Stop Going Easy on AI Writing
  4. 173
    The Free Step-Level Grader Hiding in Every RL Training Run
  5. 154
    How a 7B Model Out-Investigates a 72B One by Choosing What to Look At
  6. 143
    When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests
  7. 133
    How MiniMax Turned a Reward-Hacking Disaster Into Olympiad Gold
  8. 131
    Why Autonomous Research Agents Forget Their Own Lessons, and Arbor's Fix
  9. 121
    When the Agent Says It's Done But Nothing Happened: Debugging the Harness, Not the Model
  10. 120
    How an AI Agent Rewrites Its Own Tools, Without an Answer Key
  11. 118
    Why the Best-Aligned AI Models Are the Easiest to Trick Into Producing Harm
  12. 097
    Same Tokens, Same Cost, Wildly Different Results: What Actually Scales in AI Agents
  13. 089
    When AI-Written Papers Read Well But the Evidence Underneath Is Broken
  14. 067
    An AI Just Solved a 1996 Erdős Problem—and the Simplest Agent Won
  15. 065
    One Loop to Optimize Them All: A Universal API for LLM-Driven Discovery
  16. 059
    Firefly's Inversion: Building Verified Tool-Call Training Data by Working Backward
  17. 055
    Why LLM Judges Flip Their Verdicts When You Change the Question Format
  18. 046
    When the AI Optimizer Edits the Grade Book: Why Harnessing Evolution Needs a Wall
  19. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
  20. 019
    When the Best Reward Model Trains the Worst Policy: Inside EvoLM
  21. 012
    Why AI Coding Agents Keep Trying to Debug Without a Debugger
  22. 007
    Exploration Hacking: When Models Sabotage Their Own RL Training

Related concepts

Related terms