Institution profile

Zhejiang University of Technology

Academic institutionasia · cn
Official website
Research library322linked papers
Opportunities0open roles
Selected work

Representative Papers

ForgetMark: Stealthy Fingerprint Embedding via Targeted Unlearning in Language Models

Jan 13, 2026

This work proposes a covert fingerprinting mechanism based on directed forgetting to address the vulnerability of existing language model watermarking methods to filtering, heuristic detection, and false triggers. By leveraging an auxiliary model and prediction entropy ranking, the method constructs compact key-value pairs and trains a lightweight LoRA adapter to selectively suppress original responses associated with specific keys, embedding imperceptible forgetting traces without compromising the model’s general capabilities. Departing from conventional fixed trigger-response paradigms, it exploits probabilistic forgetting patterns to significantly enhance stealth and reduce false positives. Combining likelihood and semantic evidence, the approach achieves 100% ownership verification accuracy under black-box and gray-box settings, remains robust against model merging and incremental fine-tuning, incurs no performance degradation on standard tasks, and consistently outperforms backdoor-based baselines.

2 citationsRead paper

Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models

Jan 12, 2026

This work proposes EvoToken-DLM, a novel diffusion language model that overcomes the limitations of traditional approaches relying on hard masking and discrete token assignments, which struggle to revise early decoding decisions and discard valuable intermediate probabilistic information. By replacing hard masks with evolving soft token distributions, EvoToken-DLM enables a progressive and revisable decoding process that smoothly transitions from masked states to discrete outputs. The model introduces a soft token evolution mechanism coupled with continuous trajectory supervision, aligning the training objective with iterative probability updates. Experimental results demonstrate that EvoToken-DLM significantly outperforms existing diffusion and masked language models across multiple benchmarks, achieving superior generation quality and enhanced flexibility.

1 citationsRead paper

PerturboLLaVA: Reducing Multimodal Hallucinations with Perturbative Visual Training

Mar 09, 2025

This paper addresses the hallucination problem in multimodal large language models (MLLMs) for dense image captioning, which stems from excessive reliance on linguistic priors. To mitigate this, we propose a vision-enhancement method based on adversarial text perturbation. Our contributions are threefold: (1) We introduce HalFscore, a concept-level hallucination evaluation metric enabling fine-grained, quantitative hallucination measurement—the first of its kind. (2) We design a perturbed-text training paradigm guided by a language graph and jointly optimized via vision–language fine-tuning, explicitly attenuating linguistic priors while strengthening visual grounding. (3) The method requires intervention only during training, incurring no additional inference overhead. Experiments demonstrate significant hallucination reduction across dense captioning and general multimodal benchmarks, with consistent performance gains over state-of-the-art approaches.

1 citationsRead paper

Revisiting Convolution Architecture in the Realm of DNA Foundation Models

Feb 25, 2025

The DNA foundation model community lacks systematic, benchmark-driven comparisons between CNNs and Transformer/SSM-based architectures. Method: We propose ConvNova, a lightweight and efficient CNN architecture that uniquely integrates dilated convolutions, gated convolutions, and a dual-branch gating mechanism—explicitly eschewing self-attention and state-space modeling. Contribution/Results: ConvNova achieves strong biological interpretability in epigenetic prediction tasks and outperforms state-of-the-art methods on over half of the evaluated benchmarks: it improves histone modification prediction accuracy by 5.8% on average, reduces parameter count by 37%, and accelerates inference by 2.1×. These results demonstrate that carefully engineered CNNs remain highly competitive for DNA sequence modeling, offering a novel, efficient, and interpretable paradigm for DNA foundation models.

1 citationsRead paper

Dataset Distillation-based Hybrid Federated Learning on Non-IID Data

Sep 26, 2024arXiv.org

To address statistical heterogeneity and high communication overhead in federated learning (FL) under severe label imbalance in mobile edge–cloud networks, this paper proposes the Hybrid Federated Learning framework with Distillation and Clustering (HFLDD). The method first performs heterogeneous clustering of clients based on data distribution to achieve inter-cluster label balance, then introduces dataset distillation within each cluster to compress and transfer representative samples to cluster heads—thereby mitigating Non-IID bias and reducing communication load. HFLDD integrates hierarchical aggregation (cluster head → server) and Non-IID-aware model evaluation. Experiments under strong label skew demonstrate that HFLDD achieves up to a 12.3% improvement in test accuracy over baselines including FedAvg and FedProx, while reducing total communication rounds by 37%. The key contributions are: (i) the novel co-design of heterogeneous clustering and cluster-head distillation; (ii) effective alleviation of both statistical heterogeneity and communication bottlenecks in resource-constrained edge–cloud FL settings.

1 citationsRead paper
Recent publications

Latest Papers