Chang Liu

LLMs can achieve state-of-the-art results on complex reasoning tasks with far fewer parameters by iteratively excavating and reasoning over external knowledge.

Current LLM agents stumble when vital information isn't indexed by search engines, but a new multi-agent framework, UIS-Digger, shows how proactive browsing and file parsing can overcome this limitation.

Achieve automated dental crown design with significantly improved geometric accuracy and clinical feasibility by leveraging margin-aware template deformation and refinement.

Today's code-generating AI falls apart when faced with real-world software engineering tasks that demand cross-repository reasoning and external knowledge, achieving less than 45% success on the new BeyondSWE benchmark.

LinkVLA tackles the language-action misalignment problem in autonomous driving by unifying language and action tokens in a shared space, leading to faster and more accurate instruction following.

End-to-end autonomous driving can ditch expert demonstrations and still achieve state-of-the-art performance, thanks to a risk-aware world model that learns to predict and avoid hazardous outcomes.

Video-LLMs can spot a shape-shifting object, but still fail at basic physics, revealing a critical gap in predictive world modeling that HOCA-Bench exposes.

Domain-specific visual tasks get a 13.51% performance boost in LVLMs thanks to a new codebook-anchored adaptation method that decouples vision encoder tuning from the language model.

Amphibious robots are getting smarter and more adaptable thanks to advances in locomotion, actuation, and control, paving the way for more effective environmental monitoring and disaster response.

Forget GPT-4o, the secret to better robot manipulation might be an agentic framework that generates diverse, physically plausible tasks, leading to superior VLA pre-training.

Get 50% shorter LLM responses without sacrificing accuracy using a new RL method that dynamically balances task reward and length constraints.