Research paper critiques contrastive critics in AI policy search

A new research paper titled "Good Rankers, Bad Objectives: Bilinear Contrastive Critics under Expressive Policy Search" explores the limitations of contrastive critics in AI policy search. The paper demonstrates that while good action rankings do not guarantee safety, unbounded bilinear scores can inflate off-support values. The research indicates that cosine bounding does not resolve this issue, and a decomposition attributes regret to norm drift. The study found that contrastive critics are weakly calibrated or inverted in the top score decile across several OGBench navigation tasks, failing to accurately order actions by value. While Bellman-trained TD-Q shows success, contrastive critics are deemed useful for compatibility ranking but require value-calibrated scalars for action selection. AI

IMPACT Critiques current contrastive critic methods, suggesting a need for value-calibrated scalars for reliable action selection in AI policy search.

RANK_REASON The item is a research paper published on arXiv detailing novel findings in AI policy search. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• AntMaze
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hugging Face
• HumanoidMaze
• IArxiv Recommender
• Influence Flower
• Litmaps
• OGBench
• PointMaze
• ScienceCast
• scite Smart Citations

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →