arXiv subjects

Cangqi Zhou

Publications and source records attributed to Cangqi Zhou.

5 recordsLinked to original sources

arXiv2026

VERPO: Verified Evidence Regularized Policy Optimization

Verifiable outcome rewards guide language-model post-training, but sequence-level advantages do not identify which token-level decisions should be preserved or revised. Evidence-conditioned Teachers provide denser supervision by replaying sampled trajectories with privileged feedback. Yet indiscriminate imitation risks transferring formatting or reasoning-style shifts that do not support task success. We introduce VERPO, a Verified Evidence Regularized Policy Optimization framework that treats evidence as a proposal for policy correction while retaining the outcome objective. It separates evidence-free reference restoration from signed token-level evidence corrections. Fisher Evidence Contrast attenuates corrections along an estimated evidence-presence direction. A stopped token-wise ZPD controller scales acceptance according to local reward alignment and Fisher movement cost, while the reference channel remains independent of acceptance. Across five scientific-reasoning and tool-use tasks, the best variant on each backbone exceeds the strongest compared baseline in average score. The averages rise from 0.6826 to 0.6857 on Qwen3-4B, from 0.6895 to 0.7058 on Qwen3-8B, and from 0.4751 to 0.5657 on Llama-3.2-1B.

cs.LG↗

arXiv2026

Optimizing RAG Rerankers with LLM Feedback via Reinforcement Learning

Rerankers play a pivotal role in refining retrieval results for Retrieval-Augmented Generation. However, current reranking models are typically optimized on static human annotated relevance labels in isolation, decoupled from the downstream generation process. This isolation leads to a fundamental misalignment: documents identified as topically relevant by information retrieval metrics often fail to provide the actual utility required by the LLM for precise answer generation. To bridge this gap, we introduce ReRanking Preference Optimization (RRPO), a reinforcement learning framework that directly aligns reranking with the LLM's generation quality. By formulating reranking as a sequential decision-making process, RRPO optimizes for context utility using LLM feedback, thereby eliminating the need for expensive human annotations. To ensure training stability, we further introduce a reference-anchored deterministic baseline. Extensive experiments on knowledge-intensive benchmarks demonstrate that RRPO significantly outperforms strong baselines, including the powerful list-wise reranker RankZephyr. Further analysis highlights the versatility of our framework: it generalizes seamlessly to diverse readers (e.g., GPT-4o), integrates orthogonally with query expansion modules like Query2Doc, and remains robust even when trained with noisy supervisors.

cs.CL↗

arXiv2024

Heterogeneous Graph Contrastive Learning with Spectral Augmentation

Heterogeneous graphs can well describe the complex entity relationships in the real world. For example, online shopping networks contain multiple physical types of consumers and products, as well as multiple relationship types such as purchasing and favoriting. More and more scholars pay attention to this research because heterogeneous graph representation learning shows strong application potential in real-world scenarios. However, the existing heterogeneous graph models use data augmentation techniques to enhance the use of graph structure information, which only captures the graph structure information from the spatial topology, ignoring the information displayed in the spectrum dimension of the graph structure. To address the issue that heterogeneous graph representation learning methods fail to model spectral information, this paper introduces a spectral-enhanced graph contrastive learning model (SHCL) and proposes a spectral augmentation algorithm for the first time in heterogeneous graph neural networks. The proposed model learns an adaptive topology augmentation scheme through the heterogeneous graph itself, disrupting the structural information of the heterogeneous graph in the spectrum dimension, and ultimately improving the learning effect of the model. Experimental results on multiple real-world datasets demonstrate substantial advantages of the proposed model.

cs.LG↗

arXiv2023

Network Intrusion Detection with Edge-Directed Graph Multi-Head Attention Networks

A network intrusion usually involves a number of network locations. Data flow (including the data generated by intrusion behaviors) among these locations (usually represented by IP addresses) naturally forms a graph. Thus, graph neural networks (GNNs) have been used in the construction of intrusion detection models in recent years since they have an excellent ability to capture graph topological features of intrusion data flow. However, existing GNN models treat node mean aggregation equally in node information aggregation. In reality, the correlations of nodes and their neighbors as well as the linked edges are different. Assigning higher weights to nodes and edges with high similarity can highlight the correlation among them, which will enhance the accuracy and expressiveness of the model. To this end, this paper proposes novel Edge-Directed Graph Multi-Head Attention Networks (EDGMAT) for network intrusion detection. The proposed EDGMAT model introduces a multi-head attention mechanism into the intrusion detection model. Additional weight learning is realized through the combination of a multi-head attention mechanism and edge features. Weighted aggregation makes better use of the relationship between different network traffic data. Experimental results on four recent NIDS benchmark datasets show that the performance of EDGMAT in terms of weighted F1-Score is significantly better than that of four state-of-the-art models in multi-class detection tasks.

cs.CR↗

arXiv2019

Semi-Supervised Graph Embedding for Multi-Label Graph Node Classification

The graph convolution network (GCN) is a widely-used facility to realize graph-based semi-supervised learning, which usually integrates node features and graph topologic information to build learning models. However, as for multi-label learning tasks, the supervision part of GCN simply minimizes the cross-entropy loss between the last layer outputs and the ground-truth label distribution, which tends to lose some useful information such as label correlations, so that prevents from obtaining high performance. In this paper, we pro-pose a novel GCN-based semi-supervised learning approach for multi-label classification, namely ML-GCN. ML-GCN first uses a GCN to embed the node features and graph topologic information. Then, it randomly generates a label matrix, where each row (i.e., label vector) represents a kind of labels. The dimension of the label vector is the same as that of the node vector before the last convolution operation of GCN. That is, all labels and nodes are embedded in a uniform vector space. Finally, during the ML-GCN model training, label vectors and node vectors are concatenated to serve as the inputs of the relaxed skip-gram model to detect the node-label correlation as well as the label-label correlation. Experimental results on several graph classification datasets show that the proposed ML-GCN outperforms four state-of-the-art methods.

cs.LG↗