Beyond Entropy: CPO for correctness-aware RLVR
tencent · hf · 2026-07-20
Beyond Entropy: Correctness-Aware Advantage Shaping via Contrastive Policy Optimization
The paper argues that entropy is a poor correctness signal in reinforcement learning with verifiable rewards (RLVR) because it cannot tell useful uncertainty from harmful confusion.
It proposes Contrastive Policy Optimization (CPO), which shapes advantages using token-level contrastive disagreement between a reference-guided distribution and a vanilla generation distribution. The authors claim this disagreement is a reliable indicator of token-level correctness. They also show:
- On-policy Distillation is a special case of CPO when the posterior distribution comes from an external teacher model.
- CPO addresses the zero-advantage problem.
- Experiments on in-domain and out-of-domain benchmarks show CPO significantly outperforms entropy-based RLVR methods while generalizing well.
- Further analysis suggests correct and incorrect responses naturally support exploration and exploitation respectively, and balancing the two yields the best performance.
Related event: CPO Optimizes Correctness-Aware Advantage Shaping in RLVR(2 posts)→
More from Research
- Proto-tools, a universal infrastructure layer for generative biology, now open source — jmschreiber91 · 2026-10-26
- New paper finds AI agents negotiate badly under uncertainty and break deals on purpose — xuanalogue · 2026-10-26
- MCP fault injection study: agent task scores drop 60% under realistic tool failures, scale doesn't help — dair_ai · 2026-10-26
- Study: AI reviewing AI misses most reward hacking; best models miss 25%+ of hacks — ronbodkin · 2026-10-26
- FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
- NLI alone rejects 5/6 new agent memories; a 0.45 similarity gate fixes it all — Alternative_Pin9598 · 2026-10-26