Jun-Chi Yan

Multi-head attention applied to network depth unlocks substantial performance gains and compute efficiency in large MoE models.

Even state-of-the-art LMMs struggle with reasoning-informed visual editing, with the best model only achieving 56.6% accuracy on a new comprehensive benchmark.

Bench2Dex is meant as a platform for studying visuo-tactile learning across dexterous hands and offers a shared setting for algorithm development while tactile hardware and simulation models are still evolving.

Native memory in foundation models can significantly enhance efficiency and performance, as shown by the innovative Metis architecture.

Scaling Hyper-Connections beyond four streams is now feasible, yielding substantial performance gains without prohibitive costs.

Saliency-guided sparse updates, focusing on high-magnitude activations in query and key vectors, unlock significant performance gains in long-context RL, outperforming uniform update strategies.

Closed-loop evaluation reveals how VLMs for autonomous driving handle the messy reality of off-road deviations and out-of-distribution states, something static QA datasets can't capture.

LLMs can now synthesize high-performance kernels for niche hardware like NPUs, even with limited data, thanks to a self-evolving agent that bootstraps and refines code via value-driven reinforcement learning.

FineRMoE achieves 6x higher parameter efficiency, 281x lower prefill latency, and 136x higher decoding throughput compared to strong baselines, demonstrating a significant leap in MoE performance.