mlsanityData to 13 Oct 2026

Good Rankers, Bad Objectives: Bilinear Contrastive Critics under Expressive Policy Search

arXiv 2607.27422 ↗First posted 29 Jul 2026cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

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 large embedding norms inflate off-support values, but cosine bounding does not remove the failure. A controlled support decomposition attributes most raw bilinear regret to norm drift. Cosine and hybrid critics nevertheless select off-support actions from most pools and incur comparable regret. Contrastive scores are weakly calibrated or inverted in the top score decile across four OGBench navigation tasks, and they fail to order fixed-query actions by value. Bellman-trained TD-Q succeeds, including in a parameter-matched function-class control. Realized costs depend on the task: simulator rollouts reveal single-step selection costs on PointMaze and the exact-Q^* toy but well-powered nulls on AntMaze and HumanoidMaze, where the controller can self-correct. A training/readout decomposition traces the lost ordering to the cosine training objective; raw-trained embeddings retain weak ordering after inference-time normalization. Candidate maximization can therefore exploit false positives caused by norm drift, score saturation, or in-support misranking. Contrastive critics remain useful compatibility rankers on navigation and manipulation tasks, but action selection requires a value-calibrated scalar.

Key facts. No citing papers in the map yet (data through 13 Oct 2026). Builds on 16 of its 16 references that are in the map. First posted on arXiv 29 Jul 2026.

Builds on

16 of 16 references.

References · 16

  1. 2023Diffusion Policy: Visuomotor Policy Learning via Action Diffusioncited by 2,760
  2. 2020Conservative Q-Learning for Offline Reinforcement Learningcited by 1,782
  3. 2021Offline Reinforcement Learning with Implicit Q-Learningcited by 1,257
  4. 2018Off-Policy Deep Reinforcement Learning without Explorationcited by 1,231
  5. 2022Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learningcited by 547
  6. 2022Contrastive Learning as Goal-Conditioned Reinforcement Learningcited by 244
  7. 2023Cal-QL: Calibrated Offline RL Pre-Training for Efficient Online Fine-Tuningcited by 230
  8. 2024OGBench: Benchmarking Offline Goal-Conditioned RLcited by 224
  9. 2025Flow Q-Learningcited by 164
  10. 2023Contrastive Energy Prediction for Exact Energy-Guided Diffusion Sampling in Offline Reinforcement Learningcited by 145
  11. 2023Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learningcited by 118
  12. 2024Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Makingcited by 55
4 more
  1. 2024Diffusion-DICE: In-Sample Diffusion Guidance for Offline Reinforcement Learningcited by 42
  2. 2025EXPO: Stable Reinforcement Learning with Expressive Policiescited by 32
  3. 2023Stabilizing Contrastive RL: Techniques for Robotic Goal Reaching from Offline Datacited by 20
  4. 2026RankQ: Offline-to-Online Reinforcement Learning via Self-Supervised Action Rankingcited by 2

Built on by

Not cited yet.