Selected work

Representative Papers

Efficient and Stealthy Jailbreak Attacks via Adversarial Prompt Distillation from LLMs to SLMs

May 26, 2025arXiv.org

Jailbreaking large language models (LLMs) incurs high computational overhead and limited practicality due to reliance on costly LLM-based inference. Method: This paper proposes the first jailbreak-capability-oriented prompt distillation framework, transferring jailbreak knowledge efficiently from LLMs to small language models (SLMs). It integrates masked language modeling, reinforcement learning, and dynamic temperature control to construct a lightweight adversarial prompt generation and distillation pipeline. Contributions/Results: Experiments demonstrate high attack success rates across multiple mainstream LLMs, strong zero-shot cross-model transferability, and enhanced attack harmfulness. The approach reduces computational cost by 73% compared to LLM-based baselines, significantly improving stealthiness and real-world deployability of jailbreak attacks.

4 citationsRead paper

Numina-Lean-Agent: An Open and General Agentic Reasoning System for Formal Mathematics

Jan 20, 2026

Existing formal mathematical reasoning systems rely on task-specific pipelines and specialized models, limiting their generality and reproducibility. This work proposes the first approach that directly leverages general-purpose code agents for formal mathematical reasoning, enabling performance improvements through base model substitution without additional training. Built upon Claude Opus 4.5 and integrated with the Lean theorem prover via an MCP architecture, the system autonomously performs interactive theorem proving, lemma retrieval, and auxiliary reasoning. It achieves a perfect score on all 12 problems of the Putnam 2025 competition and successfully assists in the formalization of the Brascamp–Lieb inequality. The implementation and solutions are publicly released.

2 citationsRead paper

Tumor Detection, Segmentation and Classification Challenge on Automated 3D Breast Ultrasound: The TDSC-ABUS Challenge

Jan 26, 2025

Automatic breast ultrasound (ABUS) tumor detection, segmentation, and classification are challenged by morphological heterogeneity, low signal-to-noise ratio, and scarcity of annotated 3D data. Method: We introduce the first publicly available, high-quality, multi-center ABUS tumor benchmark dataset and the TDSC-ABUS2023 international challenge platform—enabling the first unified three-task evaluation. Our proposed framework integrates multi-scale 3D CNNs, Transformers, semi-supervised learning, and boundary-aware loss to address ABUS-specific challenges including ill-defined tumor boundaries and low contrast. Contribution/Results: Our method achieves state-of-the-art performance: 82.3% mAP@0.5 for detection, 79.6% Dice for segmentation, and 91.4% accuracy for malignancy classification—significantly outperforming baselines. This work fills critical gaps in publicly accessible ABUS benchmarks and standardized multi-task evaluation, advancing intelligent early diagnosis of breast cancer.

2 citationsRead paper

Efficient UAV trajectory prediction: A multi-modal deep diffusion framework

Jan 26, 2026

This work addresses the challenge of insufficient accuracy in predicting trajectories of unauthorized drones in low-altitude airspace, which stems from the limited information provided by single-sensor systems. To overcome this limitation, the authors propose a multimodal deep diffusion framework that fuses point clouds from LiDAR and millimeter-wave radar. The approach employs structurally aligned dual-branch encoders to extract modality-specific features and introduces a bidirectional cross-attention mechanism to achieve semantic alignment and complementary fusion of geometric structures and dynamic reflectivity characteristics. A tailored loss function and post-processing strategy are further integrated to enhance prediction performance. Evaluated on the MMAUD dataset, the proposed method achieves a 40% improvement in trajectory prediction accuracy over baseline models, demonstrating the effectiveness and practicality of the multimodal fusion strategy.

1 citationsRead paper

WenetSpeech-Wu: Datasets, Benchmarks, and Models for a Unified Chinese Wu Dialect Speech Processing Ecosystem

Jan 16, 2026

This work addresses the underdevelopment of robust speech technologies for low-resource dialects such as Wu Chinese, hindered by the scarcity of large-scale datasets, standardized evaluation benchmarks, and open-source models. To bridge this gap, we present WenetSpeech-Wu, the first large-scale, multi-dimensionally annotated open-source Wu Chinese speech corpus comprising approximately 8,000 hours of audio. Building upon this resource, we introduce WenetSpeech-Wu-Bench, the first standardized multi-task benchmark for Wu Chinese, encompassing six core tasks: automatic speech recognition (ASR), Wu-to-Mandarin translation, speaker attribute prediction, emotion recognition, text-to-speech synthesis (TTS), and intent-driven TTS. We also release a suite of strong open-source baseline models trained on this corpus, significantly advancing the Wu Chinese speech processing ecosystem and establishing the first systematic foundation for research in this domain.

1 citationsRead paper

Recent publications

Latest Papers

End-to-End Cell Detection via Instance-aware Graph Modeling

Sep 14, 2026

本文提出一种端到端框架，通过实例感知图建模解决细胞检测和分类问题，动态构建图结构并融合视觉与关系证据，优于现有方法。

0 citationsRead paper

ForeSight: Enhancing Risk Monitoring via Early Safety Signal Distillation

Sep 12, 2026

为解决大型语言模型生成有害内容的问题，提出ForeSight框架，通过提炼早期安全信号，实现高效的风险预测。

0 citationsRead paper

SyRHM: Symbolic-Language-Enhanced Reasoning with Associative Retrieval for Zero-shot Harmful Meme Detection

Sep 12, 2026

为解决隐含恶意的梗图检测问题，提出SyRHM框架，通过关联检索和符号语言增强的多阶段推理来提升检测效果。

0 citationsRead paper

SinoGlyphBench: A Diagnostic Benchmark for Chinese Glyph-Level Obfuscation in Language-Model Moderation

Sep 05, 2026

Glyph-level obfuscation can leave harmful Chinese content readable to humans while degrading automated moderation. We introduce SinoGlyphBench, a diagnostic benchmark that identifies label-critical semantic anchors and creates matched original and glyph-obfuscated inputs in text and image modalities. By perturbing anchors, background context, or both, this design distinguishes corruption of moderation-relevant evidence from general surface variation. Across 176,916 paired evaluations of 12 LLMs and MLLMs, obfuscation increases harmful false-negative and false-positive rates by 6.1 and 4.7 percentage points, respectively, and reduces four-way accuracy by 5.0 points. Models retain 75.7% of the decisions that were correct on the matched original inputs. Full-scope perturbations cause the largest degradation, anchor-only perturbations are more damaging than background-only perturbations, and cross-script substitution is particularly difficult in the text modality. Analysis of structured outputs identifies observable mismatches in visible-form reading, intended-message recovery, and final safety judgment. The evaluated models, therefore, remain brittle to Chinese content written with non-canonical glyphs. Resources are available at https://github.com/fengshun124/SinoGlyphBench.

0 citationsRead paper

Modality Disentangled Learning for Incomplete Multimodal Emotion Recognition: A Primitive Memory Distillation Perspective

Aug 31, 2026

为解决多模态情感识别中模态缺失问题，提出Primitive Memory Distillation框架，通过解耦和记忆库方法提高表示稳定性和鲁棒性。

0 citationsRead paper