Selected work

Representative Papers

TopoLoRA-SAM: Topology-Aware Parameter-Efficient Adaptation of Foundation Segmenters for Thin-Structure and Cross-Domain Binary Semantic Segmentation

Jan 05, 2026arXiv.org

This work addresses the challenges faced by foundational segmentation models like SAM in fine-structured and cross-modal binary semantic segmentation tasks, where full fine-tuning is computationally expensive and prone to catastrophic forgetting. To overcome these limitations, we propose TopoLoRA-SAM—the first framework integrating topological awareness with parameter-efficient fine-tuning. Our approach freezes the ViT encoder and injects low-rank adaptation (LoRA) modules alongside lightweight spatial convolutional adapters, optionally augmented with differentiable clDice-based topological supervision. Training only 5.2% of the model parameters (approximately 4.9M), TopoLoRA-SAM achieves state-of-the-art average Dice scores across five benchmark datasets, with particularly significant improvements in segmentation accuracy and robustness on CHASE_DB1, outperforming fully fine-tuned task-specific models.

1 citationsRead paper

Adapting Without Gradients: Affine Statistics Transport and What Its Certificate Can Tell You

Aug 31, 2026

该研究针对冻结模型的测试时适应问题，提出了一种无需梯度的方法CASTER，通过存储源类统计信息并估计目标批次的仿射变换来解决。

0 citationsRead paper

Evaluating RAG for French immigration law: a benchmark and baseline study

Jul 27, 2026

This study addresses the absence of a dedicated AI evaluation benchmark in French immigration law by introducing the first publicly available question-answering benchmark, encompassing three core administrative guidance tasks: residence permit type recommendation, required document retrieval, and citation of relevant legal provisions. Leveraging 52 annotated synthetic user profiles, the work systematically evaluates the performance of parametric large language models—including Qwen3.5-9B and Qwen3.5-27B—alongside dense retrieval-augmented generation (RAG) approaches across varying scales. Experimental results demonstrate that RAG substantially improves answer accuracy, particularly in permit recommendation, thereby underscoring the critical role of retrieval grounding in enhancing the reliability of legal AI systems.

0 citationsRead paper

Thermodynamic Signatures of Reasoning: Free-Energy and Spectral-Form-Factor Diagnostics for Hallucination Detection in Large Language Models

Jun 17, 2026

Detecting hallucinations in large language models urgently requires efficient, fine-tuning-free methods. This work proposes Free-Energy Signatures (FES), which uniquely integrate thermodynamic potentials—such as free energy and spectral entropy—with random matrix theory (RMT). By treating the attention map Laplacian as a Hamiltonian, FES constructs unsupervised spectral descriptors that exhibit Lipschitz stability, strong expressive power, and theoretical Probably Approximately Correct (PAC) guarantees. Empirical evaluation across six open-source large language models and six benchmark datasets demonstrates that FES achieves an average AUROC 6.5 points higher than LapEig and 2.4 points higher than GoR-4, with an RMT-based anomaly score reaching 0.71 AUROC.

0 citationsRead paper

MC-RFM: Geometry-Aware Few-Shot Adaptation via Mixed-Curvature Riemannian Flow Matching

May 08, 2026

This work addresses the limitation of existing parameter-efficient fine-tuning methods, which often overlook the geometric structure of feature shifts in downstream tasks and struggle to balance semantic hierarchy with local discriminability. To overcome this, we propose a mixed-curvature Riemannian flow matching framework that performs task-conditioned continuous feature transport on cached features from a frozen vision backbone, aligning query samples to support-set prototypes via a hyperbolic–Euclidean product manifold. Our approach is the first to integrate mixed-curvature manifolds with flow matching for few-shot adaptation, introducing an adaptive branch gating mechanism, prototype contraction, and a hybrid prototype–linear classifier to explicitly model the geometry of representation transfer. Extensive experiments across seven vision benchmarks, five backbone architectures, and 1/4/16-shot settings demonstrate consistent state-of-the-art performance, with particularly notable gains on Transformer backbones and fine-grained datasets.

0 citationsRead paper

Recent publications

Latest Papers

Adapting Without Gradients: Affine Statistics Transport and What Its Certificate Can Tell You

Aug 31, 2026

该研究针对冻结模型的测试时适应问题，提出了一种无需梯度的方法CASTER，通过存储源类统计信息并估计目标批次的仿射变换来解决。

0 citationsRead paper

Evaluating RAG for French immigration law: a benchmark and baseline study

Jul 27, 2026

This study addresses the absence of a dedicated AI evaluation benchmark in French immigration law by introducing the first publicly available question-answering benchmark, encompassing three core administrative guidance tasks: residence permit type recommendation, required document retrieval, and citation of relevant legal provisions. Leveraging 52 annotated synthetic user profiles, the work systematically evaluates the performance of parametric large language models—including Qwen3.5-9B and Qwen3.5-27B—alongside dense retrieval-augmented generation (RAG) approaches across varying scales. Experimental results demonstrate that RAG substantially improves answer accuracy, particularly in permit recommendation, thereby underscoring the critical role of retrieval grounding in enhancing the reliability of legal AI systems.

0 citationsRead paper

Thermodynamic Signatures of Reasoning: Free-Energy and Spectral-Form-Factor Diagnostics for Hallucination Detection in Large Language Models

Jun 17, 2026

Detecting hallucinations in large language models urgently requires efficient, fine-tuning-free methods. This work proposes Free-Energy Signatures (FES), which uniquely integrate thermodynamic potentials—such as free energy and spectral entropy—with random matrix theory (RMT). By treating the attention map Laplacian as a Hamiltonian, FES constructs unsupervised spectral descriptors that exhibit Lipschitz stability, strong expressive power, and theoretical Probably Approximately Correct (PAC) guarantees. Empirical evaluation across six open-source large language models and six benchmark datasets demonstrates that FES achieves an average AUROC 6.5 points higher than LapEig and 2.4 points higher than GoR-4, with an RMT-based anomaly score reaching 0.71 AUROC.

0 citationsRead paper

MC-RFM: Geometry-Aware Few-Shot Adaptation via Mixed-Curvature Riemannian Flow Matching

May 08, 2026

This work addresses the limitation of existing parameter-efficient fine-tuning methods, which often overlook the geometric structure of feature shifts in downstream tasks and struggle to balance semantic hierarchy with local discriminability. To overcome this, we propose a mixed-curvature Riemannian flow matching framework that performs task-conditioned continuous feature transport on cached features from a frozen vision backbone, aligning query samples to support-set prototypes via a hyperbolic–Euclidean product manifold. Our approach is the first to integrate mixed-curvature manifolds with flow matching for few-shot adaptation, introducing an adaptive branch gating mechanism, prototype contraction, and a hybrid prototype–linear classifier to explicitly model the geometry of representation transfer. Extensive experiments across seven vision benchmarks, five backbone architectures, and 1/4/16-shot settings demonstrate consistent state-of-the-art performance, with particularly notable gains on Transformer backbones and fine-grained datasets.

0 citationsRead paper

AdapterTune: Zero-Initialized Low-Rank Adapters for Frozen Vision Transformers

Mar 15, 2026

This work addresses the optimization instability and lack of theoretical capacity guidance associated with inserting adapters into frozen vision Transformer backbones during transfer learning. The authors propose the Zero-initialized Residual Low-rank Adapter, which introduces a low-rank bottleneck structure in each Transformer block, with the up-projection layer initialized to zero to ensure that fine-tuning starts identically to the pretrained model, thereby preventing early representation drift. For the first time, the adapter rank is theoretically modeled as a capacity budget tied to the feature shift of downstream tasks, revealing an “elbow”-shaped accuracy gain as rank increases. Experiments across nine datasets and three backbone scales show that the method improves top-1 accuracy by 14.9% on average over training only the classification head, using just 0.92% of the parameters required for full fine-tuning, and outperforms full fine-tuning in 10 out of 15 dataset-backbone combinations.

0 citationsRead paper