Selected work

Representative Papers

ICPO: Illocution-Calibrated Policy Optimization for Multi-Turn Conversation

Jan 20, 2026

This work addresses the challenge of "dialogue drift" in large language models during multi-turn conversations, where ambiguous initial instructions lead to persistent erroneous assumptions that are difficult to correct. To mitigate this, the paper proposes a novel reinforcement learning framework that, for the first time, incorporates pragmatic awareness into the reward mechanism. By detecting instruction ambiguity, the framework dynamically modulates reward signals to encourage the model to express epistemic humility or proactively seek clarification under uncertainty. This approach combines verifiable rewards with data augmentation using ambiguous prompts to enable fine-grained control over response style. Experimental results demonstrate that the proposed method improves average performance by 75% on multi-turn dialogue tasks while maintaining robustness on single-turn benchmarks, significantly enhancing both the cooperativeness and robustness of conversational agents.

1 citationsRead paper

A Time-Frequency Dual-Domain Multi-Scale Convolutional Neural Network for Bearing Fault Diagnosis under Strong Noise

Aug 10, 2026

This work addresses the significant performance degradation of bearing fault diagnosis under strong noise by proposing a lightweight time–frequency dual-domain multi-scale convolutional neural network. The time-domain branch captures impulsive features through parallel multi-scale convolutional kernels, while the frequency-domain branch leverages fast Fourier transform to extract noise-robust spectral structures. A novel dual-domain collaborative fusion mechanism is designed to enable complementary feature integration. With only 110,000 parameters, the model achieves 99.75% accuracy on clean signals from the CWRU dataset and maintains 92.50% accuracy under severe −4 dB noise—outperforming single-domain baselines by 7.25 percentage points and surpassing state-of-the-art methods such as WDCNN and DRSN-CW, thereby demonstrating superior robustness and computational efficiency.

0 citationsRead paper

M$^3$Prune: Hierarchical Collaborative Pruning for Efficient Multi-Modal Multi-Agent Retrieval-Augmented Generation

Aug 06, 2026

Multimodal multi-agent Retrieval-Augmented Generation (mRAG) systems face significant challenges including high communication overhead, substantial computational costs, and limited scalability. To address these issues, this work proposes M³Prune, a novel framework that achieves the first unified approach to both intra- and cross-modal collaborative pruning. By constructing a hierarchical cooperative communication graph and integrating graph sparsification, modality alignment scoring, and a progressive edge-pruning strategy, M³Prune effectively eliminates redundant communication edges while preserving task performance and substantially reducing computational demands. Experimental results demonstrate that M³Prune outperforms existing single-agent and multi-agent methods on both general and domain-specific mRAG benchmarks, achieving significant reductions in token consumption without compromising accuracy.

0 citationsRead paper

Radar4D-VLM: Proposal-Grounded Temporal 4D Radar Reasoning Across Frozen Language Models

Aug 04, 2026

This work addresses the underutilization of 4D millimeter-wave radar in vision-language models for autonomous driving, despite its robustness in adverse weather and unique radial velocity information. The authors propose the first purely 4D radar-driven inference framework leveraging frozen large language models (LLMs). Their approach generates geometrically anchored object proposals from ten consecutive radar point cloud frames and constructs a compact three-level token representation encoding objects, scenes, and kinematics. A parameter-efficient projector interfaces this representation with various frozen LLMs—including Llama and Qwen—to enable multitask reasoning. Evaluated on the K-Radar validation set, the method achieves a 98.13% recall@64 for proposals within 4 meters, substantially outperforming baselines, and demonstrates consistent compatibility across eight diverse LLMs, validating the effectiveness and generalizability of the proposed tokenization strategy.

0 citationsRead paper

Beyond Single-Use Tokens: Durable Authorization State for Replay-Resistant LLM Agent Actions

Aug 03, 2026

This work identifies a critical vulnerability in large language model (LLM) agents: even when using single-use tokens, repeated execution of the same user authorization during replanning, retries, or crash recovery can lead to semantic replay attacks and budget overruns. The study demonstrates for the first time that token-based mechanisms alone are insufficient to prevent such attacks and argues that maintaining a persistent, monotonic state of authorized actions, confirmation events, and remaining budgets is essential for effective replay protection. To address this, the authors propose CapLease, a mechanism that enforces end-to-end idempotency through a transactional Issue-Prepare-Commit protocol, binding user confirmations to actions, and employing idempotent external receivers. Experiments across diverse LLM agent scenarios show that CapLease effectively prevents duplicate authorizations and their unintended side effects, underscoring the necessity and superiority of persistent authorization state over transient token-based approaches.

0 citationsRead paper

Recent publications

Latest Papers

A Time-Frequency Dual-Domain Multi-Scale Convolutional Neural Network for Bearing Fault Diagnosis under Strong Noise

Aug 10, 2026

This work addresses the significant performance degradation of bearing fault diagnosis under strong noise by proposing a lightweight time–frequency dual-domain multi-scale convolutional neural network. The time-domain branch captures impulsive features through parallel multi-scale convolutional kernels, while the frequency-domain branch leverages fast Fourier transform to extract noise-robust spectral structures. A novel dual-domain collaborative fusion mechanism is designed to enable complementary feature integration. With only 110,000 parameters, the model achieves 99.75% accuracy on clean signals from the CWRU dataset and maintains 92.50% accuracy under severe −4 dB noise—outperforming single-domain baselines by 7.25 percentage points and surpassing state-of-the-art methods such as WDCNN and DRSN-CW, thereby demonstrating superior robustness and computational efficiency.

0 citationsRead paper

M$^3$Prune: Hierarchical Collaborative Pruning for Efficient Multi-Modal Multi-Agent Retrieval-Augmented Generation

Aug 06, 2026

Multimodal multi-agent Retrieval-Augmented Generation (mRAG) systems face significant challenges including high communication overhead, substantial computational costs, and limited scalability. To address these issues, this work proposes M³Prune, a novel framework that achieves the first unified approach to both intra- and cross-modal collaborative pruning. By constructing a hierarchical cooperative communication graph and integrating graph sparsification, modality alignment scoring, and a progressive edge-pruning strategy, M³Prune effectively eliminates redundant communication edges while preserving task performance and substantially reducing computational demands. Experimental results demonstrate that M³Prune outperforms existing single-agent and multi-agent methods on both general and domain-specific mRAG benchmarks, achieving significant reductions in token consumption without compromising accuracy.

0 citationsRead paper

Radar4D-VLM: Proposal-Grounded Temporal 4D Radar Reasoning Across Frozen Language Models

Aug 04, 2026

This work addresses the underutilization of 4D millimeter-wave radar in vision-language models for autonomous driving, despite its robustness in adverse weather and unique radial velocity information. The authors propose the first purely 4D radar-driven inference framework leveraging frozen large language models (LLMs). Their approach generates geometrically anchored object proposals from ten consecutive radar point cloud frames and constructs a compact three-level token representation encoding objects, scenes, and kinematics. A parameter-efficient projector interfaces this representation with various frozen LLMs—including Llama and Qwen—to enable multitask reasoning. Evaluated on the K-Radar validation set, the method achieves a 98.13% recall@64 for proposals within 4 meters, substantially outperforming baselines, and demonstrates consistent compatibility across eight diverse LLMs, validating the effectiveness and generalizability of the proposed tokenization strategy.

0 citationsRead paper

Beyond Single-Use Tokens: Durable Authorization State for Replay-Resistant LLM Agent Actions

Aug 03, 2026

This work identifies a critical vulnerability in large language model (LLM) agents: even when using single-use tokens, repeated execution of the same user authorization during replanning, retries, or crash recovery can lead to semantic replay attacks and budget overruns. The study demonstrates for the first time that token-based mechanisms alone are insufficient to prevent such attacks and argues that maintaining a persistent, monotonic state of authorized actions, confirmation events, and remaining budgets is essential for effective replay protection. To address this, the authors propose CapLease, a mechanism that enforces end-to-end idempotency through a transactional Issue-Prepare-Commit protocol, binding user confirmations to actions, and employing idempotent external receivers. Experiments across diverse LLM agent scenarios show that CapLease effectively prevents duplicate authorizations and their unintended side effects, underscoring the necessity and superiority of persistent authorization state over transient token-based approaches.

0 citationsRead paper

MNC: Scope-Bound Semantic Declassification for Private LLM-Agent Communication

Aug 03, 2026

This work addresses the vulnerability of multi-agent large language model (LLM) systems to unintended disclosure of protected states during communication, where existing privacy mechanisms struggle to precisely govern what is disclosed and how it is subsequently used. To tackle this, the paper introduces the Minimal Necessary Communication (MNC) protocol—the first approach to integrate scope-bound, type-based semantic declassification into LLM agent communication. MNC employs a reference monitor to enforce fine-grained policies that constrain the recipient, purpose, forwarding, lifetime, logging, and memory retention of disclosed information. Coupled with history-aware reasoning for risk assessment, MNC provides composable privacy guarantees. Experimental results demonstrate that MNC effectively prevents unauthorized forwarding, logging, persistent storage, and stale retrieval while preserving task utility, ensuring secure propagation of disclosed information across planning, tool invocation, and memory operations.

0 citationsRead paper