Institution profile

Xinjiang University

Academic institutionasia · cn
Official website
Research library178linked papers
Opportunities0open roles
Selected work

Representative Papers

Membox: Weaving Topic Continuity into Long-Range Memory for LLM Agents

Jan 07, 2026arXiv.org

This work addresses the challenge that existing large language model agents struggle to maintain thematic continuity in dialogue, often resulting in fragmented narratives and broken causal chains. To overcome this, the authors propose Membox, a novel hierarchical memory architecture that introduces thematic continuity modeling at the storage stage. Membox employs Topic Loom to aggregate thematically related dialogue segments into coherent “memory boxes” and utilizes Trace Weaver to construct long-range event timelines across conversational discontinuities, enabling cognitively inspired, efficient memory organization. Departing from the conventional “fragmented storage–retrieval reconstruction” paradigm, Membox achieves up to a 68% relative improvement in temporal reasoning F1 score on the LoCoMo benchmark, significantly outperforming baselines such as Mem0 and A-MEM while using fewer context tokens, thereby achieving both higher efficiency and superior performance.

1 citationsRead paper

Multi-scale Decomposed Convolution Refinement Network for Visible-Infrared Person Re-Identification

Aug 16, 2026

This study addresses the challenges of significant cross-modality discrepancies and insufficient feature discriminability in visible-infrared person re-identification by proposing MDCRNet. The method introduces a hierarchical decomposition convolutional attention module and a granularity discriminative loss, integrating multi-scale decomposition convolutions, channel attention, and spatial awareness blocks. Furthermore, it employs a joint metric loss to simultaneously optimize intra-modal compactness and inter-modal separability, thereby significantly enhancing cross-modal feature learning and discrimination. Experimental results demonstrate that MDCRNet achieves state-of-the-art performance on both the SYSU-MM01 and RegDB datasets, effectively improving accuracy in cross-modality person re-identification tasks.

0 citationsRead paper
Recent publications

Latest Papers

Multi-scale Decomposed Convolution Refinement Network for Visible-Infrared Person Re-Identification

Aug 16, 2026

This study addresses the challenges of significant cross-modality discrepancies and insufficient feature discriminability in visible-infrared person re-identification by proposing MDCRNet. The method introduces a hierarchical decomposition convolutional attention module and a granularity discriminative loss, integrating multi-scale decomposition convolutions, channel attention, and spatial awareness blocks. Furthermore, it employs a joint metric loss to simultaneously optimize intra-modal compactness and inter-modal separability, thereby significantly enhancing cross-modal feature learning and discrimination. Experimental results demonstrate that MDCRNet achieves state-of-the-art performance on both the SYSU-MM01 and RegDB datasets, effectively improving accuracy in cross-modality person re-identification tasks.

0 citationsRead paper

A Parameter-Specific Retrieval and Knowledge-Guided Reasoning Framework for LLM-Based GPSR Optimization in FANETs

Aug 06, 2026

This work addresses the challenge that existing GPSR-based routing protocols for flying ad hoc networks (FANETs) struggle to adaptively tune critical parameters—such as hello interval, number of multipath routes, and greedy forwarding weights—in highly dynamic environments. To overcome this limitation, the paper proposes the PMKR-GPSR framework, which innovatively integrates large language models with a knowledge-guided mechanism. Specifically, it employs parameter-specific multi-index retrieval to acquire relevant optimization experiences from historical data and constructs a constrained knowledge graph to ensure that parameter adjustments adhere to protocol semantics and feasibility constraints. Experimental results demonstrate that the proposed approach significantly improves packet delivery ratio and reduces end-to-end delay in high-mobility FANET simulations compared to baseline methods.

0 citationsRead paper