Kai Chen

This work introduces physical rank consistency (PRC) to measure how well tokenization preserves local physical distance rankings after reconstruction, and presents ActionPiece, which preserves physical action relationships through joint supervision of representation learning and quantization.

Turn-level Multiscale Density Ratio Estimation (tlm-DRE), which assigns different weights on corresponding turns and proposes asymmetric token-level training based on the positive-negative space gaps across multiple turns of tasks, enables LLMs to perform robustly in multi-turn reasoning tasks with both in-domain and out-of-domain conditions.

Qualitative examples show the PhysBrain 1.5 model's ability to produce end-effector trajectories and predict future scenes through spatially aligned RGB, depth, and robot-mask outputs.

Breaking the privacy-utility bottleneck, DP-NGD achieves state-of-the-art accuracy and a 10x speedup in convergence for differentially private training.

SIEVE reveals that leveraging reusable structures in demonstration data can lead to more efficient and effective imitation learning, outperforming full-data training with significantly less input.

M-agents harbor 158 unique implementation bugs that could lead to critical failures in real-world applications, and a new tool can identify over 60% of these issues automatically.

Achieve >97.5% of full-data VIT performance with only 16% of the data using ScalSelect, a surprisingly effective and scalable training-free data selection method.