Xinghao Chen

MATCH reveals that integrating in-context retrieval can dramatically boost the performance of sparse attention models without sacrificing efficiency.

Rigid geometric compression can collapse reasoning space, while generative reconstruction preserves semantic integrity and enhances reasoning accuracy.

Semantic acceptance rates can be misleading, with up to 44.2% of models failing to prevent observable harm even when they pass initial checks.

Skill rewriting can reduce operational costs by over 14% without sacrificing performance, challenging the notion that shorter skills are always better.

One-step image super-resolution can now achieve state-of-the-art realism and fidelity thanks to a novel framework that guides generative flows to avoid trajectory drift and prior collapse.

Current video editing AIs still struggle to balance visual quality, instruction adherence, and localized edits, as revealed by a new benchmark designed to disentangle these factors.

LLMs can achieve 93% precision in single-turn tool use, surpassing GPT, Gemini, and Claude, when trained on a unified dataset that standardizes tool interaction patterns and enforces cross-turn dependencies.

Current video editing methods still struggle to maintain physical realism when removing objects, often leaving behind telltale signs like lingering shadows that betray the edit.

Text-to-image synthesis just got almost 4x faster without sacrificing image quality, thanks to a clever twist on Speculative Jacobi Decoding that keeps the generation process moving even when initial drafts are rejected.