Full Professor, Nanjing University
3 papers at NeurIPS 2025
This paper presents a simple OOD detection approach by mixing multi-layer representations of Transformer.
We propose Curriculum Abductive Learning (C-ABL), a method that explicitly leverages structural properties of domain knowledge bases to improve training efficiency, stability and reasoning accuracy in ABL.
We introduce the first theoretical framework for analyzing LLM reasoning errors, and bridge two typical sampling-based test-time scaling methods to achieve both low error and fast convergence.
