CPO: Correctness-Aware Advantage Shaping via Contrastive Policy Optimization
Weiwen Xu · hf · 2026-07-20
Background: In Reinforcement Learning with Verifiable Rewards (RLVR), entropy is commonly used for advantage shaping. However, entropy fails to distinguish useful exploration (uncertainty) from detrimental confusion, limiting its effectiveness as a correctness signal.
Method: The authors propose Contrastive Policy Optimization (CPO). It uses token-level contrastive disagreement between reference-guided and vanilla generation distributions for correctness-aware advantage shaping.
Conclusions:
- Both theoretical and empirical results show this disagreement reliably indicates token-level correctness.
- On-policy Distillation is proven to be a special case of CPO.
- CPO resolves the zero-advantage problem. Experiments demonstrate that CPO substantially outperforms entropy-based RLVR methods while maintaining strong generalization. Balancing exploitation of correct responses and exploration of incorrect ones leads to the best performance.
Related event: CPO Optimizes Correctness-Aware Advantage Shaping in RLVR(2 posts)→
More from Research
- Meta FAIR's Self-Direction turns agent failures into reusable LoRA skills, +13.8pp on Terminal-Bench — karen_ullrich · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Why don't more LLMs ship 10M-token context? A researcher asks about attention tradeoffs — agarwl_ · 2026-10-27
- PSNR below 15 dB in papers likely means the metric was misused, researcher warns — kwangmoo_yi · 2026-10-27
- Ortet's first health eval: 34 open-weight models, 50 tasks, generic rankings fall short — thegautamkamath · 2026-10-27
- Scaling-fit discrepancy: output-head FLOPs are ~65% of compute at the smallest rung, ~3% at 535B — dlwh · 2026-10-27