Selected work

Representative Papers

SENT Map - Semantically Enhanced Topological Maps with Foundation Models

Nov 05, 2025

Indoor autonomous navigation suffers from inadequate semantic representation, inflexible editing of semantic information, and frequent generation of physically infeasible paths during planning. Method: We propose the Semantically Enhanced Topological Map (SENT-Map), a lightweight JSON-based representation unifying human-readable and foundation-model-(FM-)parsable semantic knowledge, enabling natural-language-driven interactive editing. A node-anchoring mechanism constrains the planning space to ensure physical feasibility. SENT-Map integrates vision foundation models for environment perception and semantic mapping, and introduces a two-stage, natural-language-driven planning framework that enables efficient execution of complex tasks using small, localized FMs. Contribution/Results: Experiments demonstrate that SENT-Map significantly improves task success rates while maintaining high robustness and generalization under resource-constrained conditions, establishing a scalable semantic modeling paradigm for lightweight embodied intelligence.

1 citationsRead paper

Harmonizing Real-Time Constraints and Long-Horizon Reasoning: An Asynchronous Agentic Framework for Dynamic Scheduling

May 27, 2026

Dynamic flexible job shop scheduling faces the challenge of simultaneously achieving millisecond-level real-time responsiveness and long-term global optimization. To address this, this work proposes RACE-Sched, a framework featuring an asynchronous dual-stream architecture that decouples execution from reasoning: a reactive stream employs low-latency symbolic heuristics for immediate scheduling decisions, while a deliberative stream leverages large language models to parallelly generate, validate, and evolve scheduling rules. The framework innovatively incorporates a semantic rule repository to enable cross-scale transferability and integrates sandbox validation with atomic update mechanisms to ensure system safety. Experimental results demonstrate that RACE-Sched significantly outperforms existing deep reinforcement learning and LLM-based approaches on GEN-Bench, MK-Bench, and JMS-Bench, achieving superior performance in both scheduling quality and dynamic adaptability.

0 citationsRead paper

Memo2496: Expert-Annotated Dataset and Dual-View Adaptive Framework for Music Emotion Recognition

Dec 15, 2025

Music emotion recognition (MER) faces two key challenges: scarcity of high-quality annotated data and cross-track feature drift. To address these, we introduce Memo2496—the first large-scale expert-annotated instrumental emotion dataset comprising 2,496 tracks with continuous valence-arousal labels—and propose DAMER, a dual-view adaptive framework. Its core contributions are: (1) a dual-stream attention mechanism fusing Mel-spectrogram and cochleagram representations; (2) progressive confidence-based pseudo-labeling, integrating extreme-emotion calibration and consistency filtering with a threshold of 0.25; and (3) style-anchored contrastive memory learning. DAMER further incorporates curriculum-learning-inspired temperature scheduling and Jensen–Shannon divergence to quantify inter-view consistency. Evaluated on Memo2496, 1000songs, and PMEmo, DAMER achieves state-of-the-art arousal classification accuracy—improving by 3.43%, 2.25%, and 0.17%, respectively. Both the Memo2496 dataset and source code are fully open-sourced.

0 citationsRead paper

Detecting Prompt Injection Attacks Against Application Using Classifiers

Dec 14, 2025

To address prompt injection attacks threatening large language model (LLM)-powered web applications, this paper systematically constructs the first high-quality, real-world-oriented prompt injection benchmark dataset—enhanced from the HackAPrompt Playground—and empirically evaluates four classifier families: LSTM, feedforward neural networks, random forests, and naive Bayes. We propose a lightweight detection framework integrating data augmentation and binary classification, achieving significant improvements in detection accuracy and cross-scenario generalization while maintaining low inference overhead. Our key contributions are: (1) the first publicly released, high-fidelity prompt injection benchmark dataset; (2) an empirical characterization of performance boundaries of diverse detectors under realistic deployment conditions; and (3) a plug-and-play, low-latency security module that enables real-time defense—providing actionable, production-ready safeguards for securing LLM-integrated applications.

0 citationsRead paper

LAFA: Agentic LLM-Driven Federated Analytics over Decentralized Data Sources

Oct 21, 2025

Existing LLM-based agent analytics frameworks rely on centralized data and lack privacy guarantees; while federated analytics (FA) supports privacy-preserving computation, it only accepts structured queries and cannot interpret natural language. Method: We propose the first system integrating large language model (LLM) agents with federated analytics, enabling natural language–driven, privacy-preserving analytics over decentralized data sources. Our approach introduces a hierarchical multi-agent architecture—comprising coarse- and fine-grained planners and optimizers—augmented with directed acyclic graph (DAG) modeling, query decomposition/merging optimization, and structured prior knowledge to automatically compile and optimize natural language queries into efficient federated execution plans. Contribution/Results: Experiments demonstrate significant improvements in query execution success rate, alongside substantial reductions in communication and computational overhead compared to baseline approaches.

0 citationsRead paper

Recent publications

Latest Papers

Harmonizing Real-Time Constraints and Long-Horizon Reasoning: An Asynchronous Agentic Framework for Dynamic Scheduling

May 27, 2026

Dynamic flexible job shop scheduling faces the challenge of simultaneously achieving millisecond-level real-time responsiveness and long-term global optimization. To address this, this work proposes RACE-Sched, a framework featuring an asynchronous dual-stream architecture that decouples execution from reasoning: a reactive stream employs low-latency symbolic heuristics for immediate scheduling decisions, while a deliberative stream leverages large language models to parallelly generate, validate, and evolve scheduling rules. The framework innovatively incorporates a semantic rule repository to enable cross-scale transferability and integrates sandbox validation with atomic update mechanisms to ensure system safety. Experimental results demonstrate that RACE-Sched significantly outperforms existing deep reinforcement learning and LLM-based approaches on GEN-Bench, MK-Bench, and JMS-Bench, achieving superior performance in both scheduling quality and dynamic adaptability.

0 citationsRead paper

Memo2496: Expert-Annotated Dataset and Dual-View Adaptive Framework for Music Emotion Recognition

Dec 15, 2025

Music emotion recognition (MER) faces two key challenges: scarcity of high-quality annotated data and cross-track feature drift. To address these, we introduce Memo2496—the first large-scale expert-annotated instrumental emotion dataset comprising 2,496 tracks with continuous valence-arousal labels—and propose DAMER, a dual-view adaptive framework. Its core contributions are: (1) a dual-stream attention mechanism fusing Mel-spectrogram and cochleagram representations; (2) progressive confidence-based pseudo-labeling, integrating extreme-emotion calibration and consistency filtering with a threshold of 0.25; and (3) style-anchored contrastive memory learning. DAMER further incorporates curriculum-learning-inspired temperature scheduling and Jensen–Shannon divergence to quantify inter-view consistency. Evaluated on Memo2496, 1000songs, and PMEmo, DAMER achieves state-of-the-art arousal classification accuracy—improving by 3.43%, 2.25%, and 0.17%, respectively. Both the Memo2496 dataset and source code are fully open-sourced.

0 citationsRead paper

Detecting Prompt Injection Attacks Against Application Using Classifiers

Dec 14, 2025

To address prompt injection attacks threatening large language model (LLM)-powered web applications, this paper systematically constructs the first high-quality, real-world-oriented prompt injection benchmark dataset—enhanced from the HackAPrompt Playground—and empirically evaluates four classifier families: LSTM, feedforward neural networks, random forests, and naive Bayes. We propose a lightweight detection framework integrating data augmentation and binary classification, achieving significant improvements in detection accuracy and cross-scenario generalization while maintaining low inference overhead. Our key contributions are: (1) the first publicly released, high-fidelity prompt injection benchmark dataset; (2) an empirical characterization of performance boundaries of diverse detectors under realistic deployment conditions; and (3) a plug-and-play, low-latency security module that enables real-time defense—providing actionable, production-ready safeguards for securing LLM-integrated applications.

0 citationsRead paper

SENT Map - Semantically Enhanced Topological Maps with Foundation Models

Nov 05, 2025

Indoor autonomous navigation suffers from inadequate semantic representation, inflexible editing of semantic information, and frequent generation of physically infeasible paths during planning. Method: We propose the Semantically Enhanced Topological Map (SENT-Map), a lightweight JSON-based representation unifying human-readable and foundation-model-(FM-)parsable semantic knowledge, enabling natural-language-driven interactive editing. A node-anchoring mechanism constrains the planning space to ensure physical feasibility. SENT-Map integrates vision foundation models for environment perception and semantic mapping, and introduces a two-stage, natural-language-driven planning framework that enables efficient execution of complex tasks using small, localized FMs. Contribution/Results: Experiments demonstrate that SENT-Map significantly improves task success rates while maintaining high robustness and generalization under resource-constrained conditions, establishing a scalable semantic modeling paradigm for lightweight embodied intelligence.

1 citationsRead paper

LAFA: Agentic LLM-Driven Federated Analytics over Decentralized Data Sources

Oct 21, 2025

Existing LLM-based agent analytics frameworks rely on centralized data and lack privacy guarantees; while federated analytics (FA) supports privacy-preserving computation, it only accepts structured queries and cannot interpret natural language. Method: We propose the first system integrating large language model (LLM) agents with federated analytics, enabling natural language–driven, privacy-preserving analytics over decentralized data sources. Our approach introduces a hierarchical multi-agent architecture—comprising coarse- and fine-grained planners and optimizers—augmented with directed acyclic graph (DAG) modeling, query decomposition/merging optimization, and structured prior knowledge to automatically compile and optimize natural language queries into efficient federated execution plans. Contribution/Results: Experiments demonstrate significant improvements in query execution success rate, alongside substantial reductions in communication and computational overhead compared to baseline approaches.

0 citationsRead paper