All three of our submissions were accepted to ICLR 2026! A fantastic accomplishment by our CLeAR members and collaborators. Here’s a snapshot:
-
Know When to Abstain: Optimal Selective Classification with Likelihood Ratios. Led by Alvin, this work applies the Neyman–Pearson lemma to design optimal selector functions for selective classification, enabling models to know when to say “I don’t know.” The framework unifies existing methods and motivates two new selectors, ∆-MDS and ∆-KNN, which consistently outperform baselines under covariate shift across vision and language tasks. A pre-print is available here.
-
Abstracting Robot Manipulation Skills via Mixture-of-Experts Diffusion Policies. Led by Ce, this work introduces Skill Mixture-of-Experts Policy (SMP), a diffusion-based policy for scalable multi-task robot manipulation. SMP learns a compact set of reusable skill experts and uses sticky routing to activate only the task-relevant experts at each step. This allows the robot to reuse learned skills, reduce inference cost, and adapt more efficiently across multi-task and transfer-learning settings. A pre-print is available here.
-
Masked Skill Token Training for Hierarchical Off-Dynamics Transfer. Led by Zeyu, this work tackles policy transfer when the target environment has different dynamics and direct interaction is unavailable. The paper proposes Masked Skill Token Training (MSTT), a fully offline hierarchical reinforcement learning framework that learns discrete skill tokens, uses masked Bellman updates to reason about dynamics shifts, and plans with temporally extended skills. This provides a promising step toward more robust and structure-aware transfer in embodied AI. See the OpenReview page here.
