Asymmetry of Verification and Verifier’s Rule
Source: https://www.jasonwei.net/blog/asymmetry-of-verification-and-verifiers-law Author: Jason Wei Date: 2026-01-22
Summary
Jason Wei (Google DeepMind) articulates the core insight behind why RL with verifiable rewards works so well for reasoning: verification is asymmetrically easier than generation. This asymmetry is what makes scalable oversight possible.
Key Claims
- Verification asymmetry: checking if a proof/code/answer is correct is far cheaper than generating it — this is the foundational insight for RL-based reasoning.
- “Verifier’s rule”: in domains with verification asymmetry, you can train models with RL using only a verifier — no human labels needed.
- Math and code are the canonical examples: running a test suite or checking a proof is fast; writing the code is hard.
- The open question: can we extend verification asymmetry to non-verifiable domains (creative writing, strategy, medical advice)?
- Weak-to-strong generalization: a weaker verifier can still elicit stronger generation if the asymmetry holds — the verifier just needs to distinguish correct from incorrect, not generate.
Connection to Other Sources
This is the theoretical foundation for why GRPO (GRPO++) works. Also explains the mechanism behind AutoEvolver’s algorithm optimization — code correctness is verifiable.
Entities
- Jason Wei — researcher at Google DeepMind
Concepts
- Test-Time Compute — verification asymmetry enables compute scaling at inference
- RL Infrastructure — verifier-based RL depends on this asymmetry
- Reward Hacking — when verification fails to capture true quality
- Autonomous Research — scalable oversight for research tasks