Jiacheng Liu

A simple training framework boosts pixel-level tampering detection performance in VLMs by over 26%, showcasing the power of balanced sampling and late-injection strategies.

AOHP redefines how AI agents interact with operating systems, achieving a 21% boost in task completion and a dramatic cut in execution costs.

MAGIS transforms strabismus diagnosis from a black-box process into a transparent, evidence-driven framework that boosts accuracy and clinical reliability.

Mixed-authorship documents can be harder to detect than purely human or AI-generated texts, challenging existing assumptions about AI-text detection.

Novelty-driven interaction enables agents to explore more effectively while using memory efficiently, outperforming traditional methods in open-ended environments.

Uncover a model's "digital DNA" – its pretraining data mixture – from its outputs alone, even without access to the training data.

LLMs can maintain long-context performance even with aggressive KV-cache eviction by learning to predict token importance and compressing evicted tokens into a latent memory.

Claude Code's architecture reveals a surprising amount of complexity outside the core LLM loop, with most code dedicated to safety, context management, and extensibility.