Selected work

Representative Papers

Accelerating LLM Inference with Lossless Speculative Decoding Algorithms for Heterogeneous Vocabularies

Jan 31, 2025arXiv.org

To address the slow inference speed of large language models (LLMs) and the limitation of existing speculative decoding methods—namely, their reliance on shared vocabularies—this paper proposes the first lossless speculative decoding framework supporting heterogeneous tokenizers. Methodologically, it requires no vocabulary alignment, modifies or retrains neither the draft nor the target model, and achieves cross-vocabulary distribution preservation via three lightweight components: token mapping, dynamic probability projection, and distribution calibration. Its core contribution is the elimination of vocabulary consistency constraints, enabling the first truly plug-and-play speculative decoding across disparate tokenizers. Evaluated on summarization, code generation, and long-context tasks, the framework achieves average speedups of 1.8–2.3× over standard autoregressive decoding, with zero accuracy degradation—outperforming all baseline speculative decoding approaches.

3 citations1 influentialRead paper

ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection

Jul 02, 2026

This work addresses the challenges of industrial anomaly detection in cold-start scenarios, where scarce normal samples and limited anomalies lead to loose normal boundaries and underutilized supervisory signals. To this end, the authors propose ArcAD, a plug-and-play calibration framework that leverages hyperspherical embedding and a push-pull learning paradigm. ArcAD aggregates the limited normal samples into a compact cluster to better cover the normal manifold while simultaneously contracting the decision boundary inward using both real and synthetically generated anomalies, thereby sharpening the boundary and enhancing discriminability. Notably, ArcAD is the first method to effectively model the normal distribution and fully exploit anomaly supervision under cold-start conditions. It achieves state-of-the-art performance across multiple benchmarks—including MVTec-AD, VisA, Real-IAD, and MANTA—outperforming both supervised and unsupervised baselines in both single-class and multi-class cold-start settings.

0 citationsRead paper

DiRotQ: Rotation-Aware Quantization for 4-bit Diffusion Transformers

May 15, 2026

This work addresses the significant degradation in image generation quality commonly observed when applying 4-bit quantization to Diffusion Transformers, which poses a critical challenge in balancing computational efficiency and model performance. The authors propose DiRotQ, a novel framework that uniquely integrates PCA-based orthogonal rotation with quantization: activations are projected onto a principal component subspace, where high-energy coefficients retain higher precision while the remaining components undergo 4-bit quantization. The method further incorporates GPTQ for weight quantization and employs custom Triton kernels to accelerate inference. To comprehensively evaluate perceptual quality and prompt alignment, the study introduces a VLM-as-a-Judge protocol. Experiments demonstrate that DiRotQ achieves an FID of 15.9 and PSNR of 19.1 dB on PixArt-Σ, outperforming SVDQuant, and reduces memory usage by 2.1× while accelerating inference by 2.3× on FLUX.1-dev.

0 citationsRead paper

MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference

Apr 01, 2026

This work addresses the limitation of conventional quantization-aware training (QAT), which supports only a single numerical format and thus struggles to accommodate dynamic precision requirements during inference. To overcome this, the authors propose a multi-format QAT framework integrated with a Slice-and-Scale transformation mechanism, enabling a single model to achieve high performance across diverse MXINT and MXFP formats while allowing runtime precision switching without retraining. Anchored on high-precision formats such as MXINT8 or MXFP8, the method generalizes instantly to unseen lower-precision configurations. Experimental results demonstrate that the proposed approach matches the accuracy of dedicated single-format QAT models across all target precisions, with negligible degradation during format transitions, thereby substantially enhancing deployment flexibility.

0 citationsRead paper

FOCA: Frequency-Oriented Cross-Domain Forgery Detection, Localization and Explanation via Multi-Modal Large Language Model

Feb 21, 2026

This work addresses the limitations of existing image forgery detection methods, which overly rely on semantic content while neglecting low-level textural cues and lacking interpretability. To overcome these issues, we propose a multimodal large language model framework that integrates features from both the RGB spatial domain and the frequency domain, leveraging a cross-attention mechanism to jointly model dual-domain representations. This approach enables high-accuracy forgery detection and localization, along with human-interpretable, cross-domain explanations. To support this research, we introduce FSE-Set, the first large-scale dataset annotated with pixel-level masks and dual-domain labels. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches in both detection accuracy and interpretability, validating the efficacy of the spatial-frequency dual-domain fusion strategy.

0 citationsRead paper

Recent publications

Latest Papers

ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection

Jul 02, 2026

This work addresses the challenges of industrial anomaly detection in cold-start scenarios, where scarce normal samples and limited anomalies lead to loose normal boundaries and underutilized supervisory signals. To this end, the authors propose ArcAD, a plug-and-play calibration framework that leverages hyperspherical embedding and a push-pull learning paradigm. ArcAD aggregates the limited normal samples into a compact cluster to better cover the normal manifold while simultaneously contracting the decision boundary inward using both real and synthetically generated anomalies, thereby sharpening the boundary and enhancing discriminability. Notably, ArcAD is the first method to effectively model the normal distribution and fully exploit anomaly supervision under cold-start conditions. It achieves state-of-the-art performance across multiple benchmarks—including MVTec-AD, VisA, Real-IAD, and MANTA—outperforming both supervised and unsupervised baselines in both single-class and multi-class cold-start settings.

0 citationsRead paper

DiRotQ: Rotation-Aware Quantization for 4-bit Diffusion Transformers

May 15, 2026

This work addresses the significant degradation in image generation quality commonly observed when applying 4-bit quantization to Diffusion Transformers, which poses a critical challenge in balancing computational efficiency and model performance. The authors propose DiRotQ, a novel framework that uniquely integrates PCA-based orthogonal rotation with quantization: activations are projected onto a principal component subspace, where high-energy coefficients retain higher precision while the remaining components undergo 4-bit quantization. The method further incorporates GPTQ for weight quantization and employs custom Triton kernels to accelerate inference. To comprehensively evaluate perceptual quality and prompt alignment, the study introduces a VLM-as-a-Judge protocol. Experiments demonstrate that DiRotQ achieves an FID of 15.9 and PSNR of 19.1 dB on PixArt-Σ, outperforming SVDQuant, and reduces memory usage by 2.1× while accelerating inference by 2.3× on FLUX.1-dev.

0 citationsRead paper

MF-QAT: Multi-Format Quantization-Aware Training for Elastic Inference

Apr 01, 2026

This work addresses the limitation of conventional quantization-aware training (QAT), which supports only a single numerical format and thus struggles to accommodate dynamic precision requirements during inference. To overcome this, the authors propose a multi-format QAT framework integrated with a Slice-and-Scale transformation mechanism, enabling a single model to achieve high performance across diverse MXINT and MXFP formats while allowing runtime precision switching without retraining. Anchored on high-precision formats such as MXINT8 or MXFP8, the method generalizes instantly to unseen lower-precision configurations. Experimental results demonstrate that the proposed approach matches the accuracy of dedicated single-format QAT models across all target precisions, with negligible degradation during format transitions, thereby substantially enhancing deployment flexibility.

0 citationsRead paper

FOCA: Frequency-Oriented Cross-Domain Forgery Detection, Localization and Explanation via Multi-Modal Large Language Model

Feb 21, 2026

This work addresses the limitations of existing image forgery detection methods, which overly rely on semantic content while neglecting low-level textural cues and lacking interpretability. To overcome these issues, we propose a multimodal large language model framework that integrates features from both the RGB spatial domain and the frequency domain, leveraging a cross-attention mechanism to jointly model dual-domain representations. This approach enables high-accuracy forgery detection and localization, along with human-interpretable, cross-domain explanations. To support this research, we introduce FSE-Set, the first large-scale dataset annotated with pixel-level masks and dual-domain labels. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches in both detection accuracy and interpretability, validating the efficacy of the spatial-frequency dual-domain fusion strategy.

0 citationsRead paper

CRoPE: Efficient Parametrization of Rotary Positional Embedding

Jan 06, 2026arXiv.org

This work addresses the parameter redundancy in conventional Rotary Position Embedding (RoPE) within the query, key, and value projections, which fails to fully exploit the expressive efficiency of complex linear transformations. The authors propose a novel RoPE reformulation grounded in genuine complex linear mappings, seamlessly integrating positional encoding into linear transformations in the complex domain. This approach substantially reduces the number of parameters in the attention module by nearly 50% while preserving model performance almost unchanged, thereby enhancing representational conciseness and interpretability. Empirical evaluations demonstrate that the parameter reduction incurs negligible performance degradation on both in-distribution and out-of-distribution tasks, confirming the method’s superiority in parameter efficiency.

0 citationsRead paper