Zekai Li

LiquiLM is a novel framework that combines Large Language Models (LLMs) with a Dynamic Co-Attention Network (DCN) to effectively bridge the semantic gap between complex smart contract code and high-level liquidity flaw descriptions, achieving high accuracy in auditing DeFi systems.

The paper proposes a novel trace-aware decoding framework, combining Temporal-Spatial Parallel Decoding (TSPD) and Confidence Extrapolation (CE), to significantly accelerate the inference of diffusion-based LLMs by identifying and fixing converged tokens early.