Verifiable Reward — AI Glossary

Reference details

Related terms

• Reinforcement Learning from Human Feedback (RLHF)
• Reward Model
• Reasoning (LLM)

How it works

1. The reward comes from executing a check rather than from a model's judgement: run the code against tests, compare the final answer to a known solution, validate the output against a schema.
2. That removes the attack surface reward models have. A learned reward can be gamed by output that scores well without being good, while a test suite cannot be flattered — the code either passes or it does not.
3. Training samples many attempts per problem, keeps the ones that verify, and reinforces them. The signal is binary per attempt, which is weak per sample and extremely cheap to produce at volume.
4. Problem difficulty therefore controls everything. A problem the model never solves yields no positive examples however many attempts are drawn, and a problem it always solves yields no gradient worth having.
5. Process reward models are added to credit intermediate steps, because a single bit at the end of a long chain says nothing about which step was wrong.
6. The ceiling is coverage: the method applies exactly as far as cheap automatic verification reaches, and most of what people want from a model — judgement, tone, helpfulness, appropriate refusal — has no verifier.

Example: Why problem difficulty, not sample count, is the binding constraint

Draw 64 attempts per problem. At a 10% per-attempt success rate the expected number that verify is 64 x 0.10 = 6.4, so the problem yields roughly six positive training examples for free — an excellent return, since no human wrote any of them. At a 1% rate it is 64 x 0.01 = 0.64, so most problems in that band yield nothing at all: the probability that all 64 attempts fail is 0.99^64 = 0.526, so about 53% of such problems produce zero signal while costing 64 generations each. At a 90% rate it is 64 x 0.9 = 57.6 positives, which sounds better and is nearly worthless — the model already solves these, and reinforcing them teaches little. The useful band is where the success rate is low enough to be informative and high enough to produce examples, roughly the middle of the range, which is why curriculum construction is the real work. Scaling attempts from 64 to 256 at a 1% rate raises expected positives from 0.64 to 2.56 at four times the compute, while moving the problem set into the 10-40% band raises it far more for nothing.

Commonly confused with

A verifiable reward is not a process reward model, and the two are frequently used together precisely because they fail in opposite ways. The verifier is ground truth about the outcome and says nothing about the path, so a lucky guess and a sound derivation score identically. A process reward model judges intermediate steps and is itself learned, which means it can be gamed in exactly the way the verifier cannot. Using the verifier alone gives a correct but uninformative signal on long chains; using the process model alone reintroduces reward hacking.

When to use it

Reach for it when:

• Domains with cheap, reliable automatic checks — code against tests, mathematics with known answers, format and schema conformance.
• Scaling training signal beyond what annotation budgets allow, which is the method's decisive advantage.
• Where reward hacking has already been observed, since a verifier closes that particular hole completely.
• With a curated difficulty distribution, which matters more than the sample count.

Reach for something else when:

• Qualities with no verifier — tone, helpfulness, judgement, appropriate refusal — which is most of what alignment addresses.
• On problems the model almost never solves, where attempts produce no positive examples and the compute is spent for nothing.
• As the only signal on long reasoning chains, where one bit at the end cannot locate the faulty step.
• Where the verifier is a proxy rather than the thing you care about: passing the supplied tests is not the same as being correct, and optimising hard against a thin test suite finds the difference.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Verifiable Reward — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path