Rerankers are critical in Retrieval-Augmented Generation (RAG) for filtering evidence that enhances the accurate generation of LLMs. With the extension to open-domain scenarios, rerankers are inevitably deployed on mixed-style corpora, whereas most existing rerankers are mainly trained on well-edited texts. A rarely explored issue lies in enabling rerankers to maximally capture the effective knowledge for downstream LLMs without being misled by stylistic features. To address this issue, we propose SARK ( S tyle-A daptive R eranker with K nowledge Prioritization), a style-augmented multi-task framework that prioritizes effective knowledge over stylistic perturbations. SARK performs multi-granular knowledge mining by using an LLM to derive passage-level super-vision on whether a passage helps or harms answer correctness, and list-level relative ranking preferences over candidate passages. It then jointly optimizes the reranker model with passage-level classification and list-level ranking objectives via style-augmented multi-task learning, encouraging the model to focus on the information needed for answering under mixed-style scenarios. Extensive experiments demonstrate that SARK improves generation performance across multiple LLMs under mixed-style conditions.
Ruwen Zhang, Bo Liu, Zhang Sheng Xiang et al.· Annual Meeting of the Associ...· 0 citations
With the widespread adoption of open-source software (OSS), software supply chain attacks have become increasingly sophisticated, posing severe threats to digital infrastructure. Traditional Software Composition Analysis (SCA) tools and Software Bill of Materials (SBOM) systems face critical limitations, including disconnection between static metadata and dynamic code evolution, inability to detect silent fixes and backported patches, and LLM-induced hallucinations in vulnerability detection. To address these challenges, this paper proposes the LLM-Enhanced Component Dependency Evolution Graph (CDEG) framework. The system constructs a time-aware heterogeneous graph by parsing SBOMs to establish static dependency edges and linking version lineages to model temporal evolution. Within this framework, we map code patches using a hybrid representation that fuses structural features extracted by Tree-sitter with semantic embeddings derived from a fine-tuned LLM, effectively distinguishing backported patches from code refactoring. Experimental results on the SC-Vuln-Real dataset demonstrate that our framework achieves 88.4% recall for backported patch detection and achieves an F1-Score of 93.8% in container image scanning.
Yuanjun Gao, Hongzhou Wu, Yujia Luo et al.· Mathematics· 0 citations