Why verifiable rewards mattered for language models
Karpathy's 2025 review explains training against checkable results and the limits of uneven model abilities.
Originally posted · Published here · 1 source · Resume reading
In his December 2025 review, Andrej Karpathy highlighted reinforcement learning from verifiable rewards as a major change in language-model training. Examples include tasks where a program can check a mathematical answer or whether code passes a test. [1]
Separate the answer from its explanation
A checkable outcome gives training a concrete target. Karpathy describes models learning intermediate calculations and revision strategies while optimizing for such rewards. The visible explanation alone is not the evidence that the answer works. [1]
Keep verification attached to the task
For a coding task, run checks against the actual behavior you need. A plausible explanation of a fix should lead to an inspectable change and a relevant test result.
Karpathy also discusses uneven capabilities: a model can perform strongly on one kind of work and fail unexpectedly on another. His review is an account of trends and examples, rather than proof that every model reasons reliably on every task. [1]
Choose a result you can check before deciding how much responsibility to give a model.
Keep this guide
Filed under AI agents
Sources
- Karpathy 2025 LLM Year in Reviewkarpathy.bearblog.dev
Next in this stack
Keyword LOOP
Five rules for building an AI agent that works
Keyword STARTER
Write skill metadata an agent can recognize
Keyword SIMPLEFIRST
Choose an AI workflow before building an agent