Hi I'm Zhenyu (Fischer) Lei

Large language models can make systematic, pattern-specific reasoning mistakes, and broad training fixes are inefficient and often disrupt other abilities. This paper introduces Reasoning Editing—targeted edits to a specific reasoning pattern—highlighting a key trade-off between generality and locality. To reduce this trade-off, we propose REdit, which reshapes underlying reasoning circuits to limit interference before applying a lightweight edit.

This paper introduces BrainTAP, a transformer-based framework for brain disorder prediction that integrates functional and structural connectivity while preserving their heterogeneous properties. The model employs adaptive mutual distillation to enable controlled cross-modal interaction and selectively incorporates expert-defined neurobiological priors through learnable global and personalized attention mechanisms, improving both representation quality and interpretability.

MolEdit is a framework for precisely editing facts inside multimodal molecule language models so they can fix outdated or wrong chemistry/biomed knowledge without retraining. It uses a Multi-Expert Knowledge Adapter (MEKA) to edit specific facets (e.g., functional groups, properties) and an Expertise-Aware Editing Switcher (EAES) to trigger edits only on closely matching inputs, plus MEBench to evaluate reliability, locality, and generality—showing sizable gains over prior editors.

This work proposes QR-Distill, a distillation framework that (i) filters for correct, LLM-judged high-quality chains-of-thought, (ii) conditionally routes the remaining paths to students based on their current state, and (iii) enables mutual student collaboration via feature-level peer teaching. Across multiple reasoning datasets, it outperforms single- and multi-path baselines.

In this paper, we study an under-explored research problem of inductive forecasting with limited training data, which requires models to generalize the learned spatial-temporal dependencies from the nodes with available training temporal data to those nodes without. To handle this problem, we propose ST-FiT that can achieve superior performance without additional fine-tuning.

While significant progress has been made in understanding brain activity through functional connectivity (FC) graphs, challenges remain in effectively capturing and interpreting the complex, long-range dependencies and multiple pathways that are inherent in these graphs. In this work, we introduce BrainMAP, a novel framework that can extract multiple long-range activation pathways with adaptive sequentialization and pathway aggregation.

We proposed a bot-detection model named BIC. BIC interacts and exchanges information across text modality and graph modality by a text-graph interaction module. BIC contains a semantic consistency module that derives the inconsistency from tweets by the attention weight to identify advanced bots.