AI TrackerFrontier desk · AEST
Back to news
normalaitech

Negative-Only Policy Optimization for One-Sided Verifiable Rewards

arXiv cs.CL · 13 Nov 2026, 04:52 AEST

A new approach called Negative-Only Policy Optimization (NOPO) addresses challenges in reinforcement learning with verifiable rewards. This method improves performance by focusing on suppressing trusted verified negatives rather than reinforcing potentially misleading verified positives.

Key facts

Read aloud

A new approach called Negative-Only Policy Optimization NOPO addresses challenges in reinforcement learning with verifiable rewards. This method improves performance by focusing on suppressing trusted verified negatives rather than reinforcing potentially misleading verified positives. Reinforcement learning with verifiable rewards RLVR typically relies on a verifier that can reliably judge whether sampled solutions are correct. We formalize this regime as one-sided verifiability OSV: verified negatives are reliable counterexamples, but verified positives are merely non-falsified candidates and may contain many false positives. To address this issue, we propose Negative-Only Policy Optimization NOPO, which treats verified positives as unlabeled and updates the policy only by suppressing trusted verified negatives. Across unit-test, constraint-check, and L L M self-verifiers, NOPO improves pass@k over baselines on code, logic, and math benchmarks when oracle rewards are unavailable.

Sources in this cluster