Glossary · Term

advantage

← all terms

Definition

Plain language

In AI training, a score for how much better a particular attempt turned out than the model's average attempt.

As stated in the literature

In policy-gradient RL, the difference between an action's return and a baseline estimate of expected return; group-relative methods like GRPO compute it by comparing rollouts on the same prompt rather than learning a separate value function.

Also called: advantages

Why it matters: It tells a training algorithm which attempts to reinforce and which to discourage, so the model learns from its own better-than-average tries rather than treating all outcomes the same.

For example, if a model's typical attempt at a problem scores 4 out of 10 but one particular attempt scores 8, that attempt's advantage is the gap that tells training to do more of what it did.

Heard on the show

“That sounds like an advantage.”
Episode 194 — How a Robot Builds a Debugging Notebook It Can Read, Edit, and Hand to Another Robot

Mentioned in 34 episodes

  1. 194
    How a Robot Builds a Debugging Notebook It Can Read, Edit, and Hand to Another Robot
  2. 188
    A Coding Agent Found a Hole in a Peer-Reviewed STOC Proof for Five Dollars
  3. 186
    How a Frozen Model Went From 2% to 77% on Physics Puzzles — Without Retraining
  4. 177
    Why Raw Profiler Data Made an AI Worse at Writing GPU Code
  5. 176
    An AI Designed Its Own Psychology Studies, Then Confirmed What It Found
  6. 173
    The Free Step-Level Grader Hiding in Every RL Training Run
  7. 169
    Why Better Bug Reports Can Make AI Coding Agents Worse
  8. 168
    When Turning Experience Into Code Makes Your AI Agent Dumber
  9. 165
    A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants
  10. 163
    Why Training Only on Perfect Solutions Cripples a Model's Reasoning
  11. 162
    The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models
  12. 157
    When an AI Coding Agent Drives a Phone Through the Terminal, No Screen Needed
  13. 150
    Don't Kill the Loser: A Different Way to Handle Two AI Agents Colliding
  14. 130
    Why AI Agents Coordinate Better Through a Shared Board Than a Boss
  15. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm
  16. 126
    How Coding Agents Can Mine Their Own Failures Into a Self-Targeting Curriculum
  17. 120
    How an AI Agent Rewrites Its Own Tools, Without an Answer Key
  18. 119
    Beating Reinforcement Learning Without Ever Touching the Model's Weights
  19. 116
    Why Streaming Half a Reasoning Chain Beats Sending the Whole Thing
  20. 107
    How a Market of Crippled AI Agents Outscored One Unrestricted Model
  21. 102
    How to Catch an AI Attack That No Single Conversation Reveals
  22. 099
    How an Open-Book Trick Teaches a Model to Catch Its Own Mistakes
  23. 093
    A Calibrated Knob for Weak-to-Strong AI Oversight, Tested on Real Code
  24. 082
    Training a Deep Research Agent on 8,000 Synthetic Tasks: The Rubric Tree Trick
  25. 080
    How a Two-Agent Trick Unlocked Large-Scale Training for Computer-Use Agents
  26. 075
    Growing Code and Proof Together: Verified Systems in Ten Hours Instead of a Year
  27. 059
    Firefly's Inversion: Building Verified Tool-Call Training Data by Working Backward
  28. 055
    Why LLM Judges Flip Their Verdicts When You Change the Question Format
  29. 033
    Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval
  30. 010
    When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL
  31. 009
    How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers
  32. 008
    Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps
  33. 007
    Exploration Hacking: When Models Sabotage Their Own RL Training
  34. 005
    Why a Debugger Designed for Humans Is the Wrong Tool for an AI Agent

Related concepts

Related terms