Contrastive Critics Fail as Value Objectives in Policy Search

Ayushman Singh, Siddharth Aphale· July 31, 2026 View original

Key takeaways

  • Contrastive critics are good for ranking but poor for value-based action selection.
  • Norm drift and score saturation cause misranking and off-support actions.
  • Bellman-trained TD-Q offers superior value calibration for action ordering.
  • Hybrid approaches may be needed to leverage contrastive strengths safely.

Who benefits

RoboticsAutonomous SystemsGamingFinancial ServicesLogistics

Summary

This research finds that while contrastive critics can rank actions, they are unreliable as value-like objectives for action selection in reinforcement learning. Issues like norm drift and score saturation lead to misranking and off-support action selection, even with cosine bounding.

This paper investigates the effectiveness of contrastive critics when used as value-like objectives in reinforcement learning, particularly for tasks like best-of-K selection and critic-guided generation. The findings indicate that despite their ability to rank actions, these critics often fail to provide reliable value estimates. Problems such as unbounded bilinear scores causing large embedding norms to inflate off-support values, and score saturation, lead to inaccurate action selection. The study demonstrates that contrastive scores are often weakly calibrated or even inverted in top score deciles across various navigation tasks. In contrast, Bellman-trained TD-Q methods prove more successful in ordering actions by value. The research suggests that while contrastive critics remain useful for compatibility ranking, they are not suitable for action selection without a value-calibrated scalar due to their susceptibility to false positives and misranking.

Why it matters

Professionals developing or deploying AI systems that rely on contrastive learning for decision-making need to be aware of its limitations as a direct value objective, potentially leading to suboptimal or unsafe actions.

How to implement this in your domain

  1. 1Evaluate existing AI systems using contrastive critics to identify potential vulnerabilities in action selection.
  2. 2Incorporate Bellman-trained TD-Q methods or other value-calibrated scalars for critical decision-making components.
  3. 3Develop robust testing protocols to detect off-support action selection and misranking in contrastive critic applications.
  4. 4Research alternative or hybrid approaches that combine the ranking strengths of contrastive critics with more reliable value estimation.
  5. 5Educate engineering teams on the nuances and limitations of contrastive learning in policy search contexts.

Original post by Ayushman Singh, Siddharth Aphale

"arXiv:2607.27422v1 Announce Type: new Abstract: Good action rankings do not make a contrastive critic safe to maximize. These critics increasingly act as value-like objectives for best-of-$K$ selection, planning, and critic-guided generation. Unbounded bilinear scores can let lar…"

View on X

Originally posted by Ayushman Singh, Siddharth Aphale on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

New Framework Improves Partial Multi-View Clustering Performance.

DAS-PMVC is a novel framework for partial multi-view clustering that addresses view asymmetry and irrelevant samples by leveraging dual alignment and structure enhancement. It uses anchor graph structure alignment, structure-enhanced feature learning, and a dual alignment strategy to achieve superior clustering performance on various datasets.

Shubin Ma, Liang Zhao, Chuanye He, Zhenjiao Liu, Liang Zou, Lin Yuanbo Wu, Yu ShaoJul 31, 2026
AI Engineering & DevToolsAI Research

Dual Teachers Improve Adversarial Robustness and Accuracy.

This work extends Information Bottleneck Distillation (IBD) by introducing a "clean teacher" alongside a robust teacher to improve the robustness/accuracy tradeoff against adversarial attacks. The proposed method transfers features from both teachers to a student model, achieving better clean accuracy while maintaining adversarial robustness, outperforming original IBD and competing with state-of-the-art approaches.

Vincent Ryusuke Takahashi, Yoshinari Takeishi, Jun'ichi Takeuchi, Kave SalamatianJul 31, 2026
AI Engineering & DevToolsAI Research

Dynamic Batch Sizes Improve Large Language Model Training Efficiency.

This paper proposes a new approach to deep learning dynamics, deriving joint scaling laws for loss based on both learning rate and batch size schedules. It introduces an optimal dynamic batch size schedule that consistently outperforms static batch size baselines, highlighting its importance for large language model training.

Jiaxiang Li, Zhiqi Bu, Shiyun XuJul 31, 2026