Linjie Li

Forget static prompts: LDEPrompt dynamically expands and freezes prompts based on layer importance, achieving state-of-the-art performance in class-incremental learning.

Quantum-inspired gating unlocks better knowledge transfer in class-incremental learning, outperforming existing methods by dynamically modeling task relationships.

Forget text-centric pipelines: FlowInOne achieves SOTA multimodal generation by unifying text, layouts, and instructions into a single visual flow, outperforming both open-source and commercial systems.

LLM agents can appear to reason well (high entropy) while completely ignoring the input, and mutual information is a far better metric for catching this failure.