Fan Li

Network traffic alone can reveal LLM agent risks and granular behaviors, but current analysis methods fail to distinguish subtle malicious activities or semantically similar scenarios.

MLLMs struggle to grasp deeply interleaved text-image contexts, with performance gaps of over 40% compared to human experts on TIC-Bench.

FlightLLM reveals how combining LLMs with structured prompts and statistical classifiers can transform complex flight safety data into interpretable insights.

Dynamically adjusting visual importance at test time can significantly enhance action prediction accuracy in VLA models.

Identity leakage drops to just 0.74% while maintaining high recognition accuracy, thanks to an innovative decoy-oriented approach in face recognition.

DRQN-CMDP achieves a unique trifecta of 83% lower gas costs, a 7.5% miss rate, and moderate detection latency, outperforming all existing methods in off-chain data auditing.

A new non-traditional code design boosts the reliability of Content Addressable Memory, tackling a critical gap in memory technology.