A verifier checks a model’s answer against a known correctness criterion. When a task has a gold answer, it can parse the model’s final response and compare it with that reference, providing a deterministic reward instead of relying on a learned reward model to estimate preference. section 11.0 section 11.1
Checking a financial answer
For a financial question, the model generates a complete response and places its final answer in a LaTeX box. The verifier extracts that answer and compares it with the gold answer; the reward arrives only at the end, so useful reasoning can still receive zero if the answer is wrong or cannot be parsed. section 11.1
The check depends on what counts as verifiable in the task:
- Compare a financial answer with its reference value. section 11.1
- Run generated code against a test suite and use pass or fail as feedback. section 11.8.4
- Use a molecular docking simulation to score a candidate molecule’s binding affinity. section 11.8.3
Figure 11.1 FinQA financial question answering as a language-model RL environment.

See how the verifier supplies a final reward in the financial question-answering task. section 11.1








