AI ENGVisual Encyclopedia

MODULE 5: REASONING & VERIFIABLE REWARDS · SCENE 13

RLVR: Rewards You Can Verify

Math checkers, compilers, and unit tests as deterministic reward oracles — no human labels in the loop.

VERIFIER VERDICTS · GROUP OF 6R = 1·[correct] + 0.1·[format only]
R=1.0✓ ANSWERR=0.1FORMAT ONLYR=0.0✗ WRONGR=1.0✓ ANSWERR=1.0✓ ANSWERR=0.1FORMAT ONLY

Rewards are computed, not learned: an exact-match checker, a compiler, a unit-test run. Fluent prose earns nothing — the oracle cannot be charmed.

TECHNICAL BREAKDOWNModule 5: Reasoning & Verifiable Rewards

RLVR: rewards you can verify

Human preference labels are noisy and expensive. For math, code, and structured tasks there is a better oracle: check the answer. A python interpreter, a unit-test suite, or an exact-answer comparator gives a reward signal that is deterministic, cheap, and impossible to charm with fluent prose.

The Oracle Set

Math: exact-match or SymPy equivalence. Code: compile + hidden tests. Agentic tasks: did the tool call succeed? These are rule-based rewards (RLVR).

Format Matters

Answers must be extractable — 'put the final number in \boxed{}'. A correct answer in the wrong format earns partial or zero reward, so format compliance gets its own reward component.

No Reward Model

Because rewards are computed, not learned, there is no reward model to over-optimize. The failure mode shifts from reward hacking to overfitting the train problem distribution.

MATHEMATICAL FORMULATION · VERIFIABLE REWARD
R = 1·[correct] + 0.1·[format_ok] · [not correct]

Binary correctness dominates; a small format reward shapes parseability without letting the model profit from well-formatted wrong answers.

REAL-WORLD PRODUCTION ENGINEERING
  • DeepSeek-R1 and OpenAI's o-series are the canonical demonstrations: RL on verifiable rewards alone elicits long self-correcting chains of thought.
  • Infrastructure is the hard part: sandboxed code execution at thousands of rollouts/step, with timeout and memory limits per sample.