Yu Ding

Opal is a private memory system for personal AI that maintains high retrieval accuracy and throughput while ensuring data privacy by confining all data-dependent reasoning to a trusted hardware enclave.

The paper proposes a local perturbation theory showing that cross-domain interference in multi-domain RL occurs via a low-dimensional shared conflict subspace, which can be selectively mitigated by short domain refresh cycles.

TempoVLA is a novel Vision-Language-Action model that enables controllable execution speed for robot manipulation by explicitly conditioning the policy on the desired speed.

This paper audits eight automatic scorers for attribution in LLM retrieval-augmented generation and finds that none of them transfer across datasets for generated-answer attribution.

This paper proposes a two-stage training framework to pretrain action modules with motion priors before Vision-Language-Action (VLA) alignment, improving VLA performance and reducing optimization challenges.

A framework called DexCompose is proposed to reuse pretrained dexterous policies for multi-task manipulation with explicit finger-level action ownership.

This paper introduces the Always-On Evaluation Protocol (AOEP-v0) for evaluating always-on agents by focusing on state mutation and recovery obligations.

The paper introduces DexVerse, a large-scale and modular benchmark for dexterous manipulation with 100 tasks, 3 robot arms, 6 hands, and configurable visual variations.

A single robot platform, Handroid, is introduced that can function as both a dexterous hand and a humanoid robot, with interchangeable control and learning frameworks.