Jiaheng Liu

A general Agent OS can boost long-horizon robotic execution and enable continual learning through structured memory management and self-evolution.

MLLMs can generate 3D models, but they often miss the mark on precise geometry and coherent assemblies, revealing significant limitations in their structural reasoning abilities.

Current video editing models falter under the weight of complex user instructions, often omitting critical edits and introducing artifacts.

Open-weight Omni models struggle with binding accuracy, achieving only 41.55% on a new counterfactual benchmark, highlighting a critical gap in long-video comprehension.

TVIR-Agent reveals that integrating visual elements into report generation can dramatically improve the quality and reliability of analytical outputs.

Converting noisy, human-centric guides into self-evolving agent skills can yield performance improvements of up to 25.3 percentage points across diverse tasks.

Span-level error localization can boost deep-research agent reliability by up to 30 percentage points, revealing critical insights into where agents go wrong.

LLM agent distillation leads to surprisingly high rates of behavioral mimicry, with some student models exhibiting tool-use habits *more* similar to their teachers than the teacher's own family members.

Evaluating web coding LLMs with real-world fidelity reveals that even state-of-the-art models still struggle with aesthetics and framework-specific nuances.