Selected work

Representative Papers

Boundary-Continuous Cross-Camera RGB Mapping via Hue-Split Model Trees

Aug 11, 2026

This work addresses the problem of RGB color inconsistency across cameras caused by differences in sensor spectral responses and image signal processing pipelines. To tackle this, the authors propose a model tree structure based on recursive hue partitioning, where an affine color correction matrix (CCM) is learned in the logarithmic domain for each node. To mitigate chromatic discontinuities at hue boundaries induced by hard partitioning, they introduce a path-weighted fusion strategy combined with an explicit continuity regularizer defined over boundary prototype pairs and simplex-constrained optimization. Evaluated on the mapping task from Canon EOS-1Ds Mark II to EOS 20D, the method significantly reduces log-RMSE and effectively suppresses false contours near hue thresholds under diverse illumination and exposure conditions, achieving both high colorimetric accuracy and perceptual smoothness.

0 citationsRead paper

Alignment of Similarity-Transformed Images Based on Fourier--Mellin Transform Using Auxiliary Function Method

Aug 11, 2026

This work addresses the challenge of achieving sub-pixel registration accuracy for images undergoing translation, scaling, and rotation. To this end, a two-stage decoupled estimation framework is proposed: first, scale and rotation parameters are estimated in the log-polar domain using the Fourier magnitude spectrum; subsequently, high-precision sub-pixel translation is estimated in the spatial domain by integrating an auxiliary function method with phase correlation. By effectively combining the Fourier–Mellin transform with auxiliary-function-based phase correlation, the approach successfully decouples the parameters of the similarity transformation. Experimental results demonstrate that the proposed method significantly outperforms conventional discrete cross-correlation–based Fourier–Mellin approaches in terms of estimation accuracy for scale, rotation, and translation.

0 citationsRead paper

From Sweep to Seam: Interleaved Cross-Block Post-Training Quantization

Aug 10, 2026

This work addresses the performance degradation in deep neural networks caused by error accumulation in conventional block-wise post-training quantization (PTQ), where a single forward pass prevents correction of early quantization errors. To mitigate this issue, the authors propose Interleaved Cross-Block Quantization (ICBQ), which refines local reconstruction by doubly optimizing adjacent “seam” block pairs at block boundaries. Built upon standard block-wise PTQ, ICBQ incorporates cross-block reconstruction within a sliding window, leveraging local shrinkage and smoothness assumptions, and is compatible with GPTQ scheduling. Experimental results demonstrate that ICBQ significantly reduces perplexity under ternary, 2-bit, and 3-bit quantization, effectively preventing performance collapse at extreme compression ratios while seamlessly integrating into existing PTQ pipelines.

0 citationsRead paper

LOGOS: A Living Logic for AI Agent Teams That Evolve With Humans

Jul 12, 2026

This work addresses the lack of controllability and auditability in continuously evolving AI agent teams by proposing a “verifiable human-in-the-loop engineering” paradigm that introduces a pluggable self-evolution and governance layer. The approach compiles multimodal inputs into versioned agent packages and integrates event sourcing, fail-closed verification, and policy-driven authorization mechanisms to ensure that all evolutionary changes require explicit human approval and empirical validation before taking effect. While enabling agents to learn continuously during operation, the system guarantees that the entire evolution process remains auditable, traceable, and ultimately under human control, thereby achieving a unified balance between operational efficiency and trustworthy governance.

0 citationsRead paper

Signs Beat Floats: Low-Rank Double-Binary Adaptation for On-Device Fine-Tuning

May 21, 2026

This work addresses the substantial storage, communication, and computational overheads incurred by floating-point operations in conventional LoRA during on-device fine-tuning of large language models, despite its small parameter count. The authors propose LoRDBA—the first LoRA-compatible method integrating double binarization—replacing low-rank factors with sign matrices and recovering magnitude information via lightweight channel-wise scaling. LoRDBA achieves accuracy comparable to fp16 LoRA while compressing adapter size by over 10×. It incurs at most an 8% increase in prefill latency during inference and requires approximately 1.6× the training memory of fp16 LoRA, significantly outperforming existing low-bit baselines under the same model scale.

0 citationsRead paper

Recent publications

Latest Papers

Boundary-Continuous Cross-Camera RGB Mapping via Hue-Split Model Trees

Aug 11, 2026

This work addresses the problem of RGB color inconsistency across cameras caused by differences in sensor spectral responses and image signal processing pipelines. To tackle this, the authors propose a model tree structure based on recursive hue partitioning, where an affine color correction matrix (CCM) is learned in the logarithmic domain for each node. To mitigate chromatic discontinuities at hue boundaries induced by hard partitioning, they introduce a path-weighted fusion strategy combined with an explicit continuity regularizer defined over boundary prototype pairs and simplex-constrained optimization. Evaluated on the mapping task from Canon EOS-1Ds Mark II to EOS 20D, the method significantly reduces log-RMSE and effectively suppresses false contours near hue thresholds under diverse illumination and exposure conditions, achieving both high colorimetric accuracy and perceptual smoothness.

0 citationsRead paper

Alignment of Similarity-Transformed Images Based on Fourier--Mellin Transform Using Auxiliary Function Method

Aug 11, 2026

This work addresses the challenge of achieving sub-pixel registration accuracy for images undergoing translation, scaling, and rotation. To this end, a two-stage decoupled estimation framework is proposed: first, scale and rotation parameters are estimated in the log-polar domain using the Fourier magnitude spectrum; subsequently, high-precision sub-pixel translation is estimated in the spatial domain by integrating an auxiliary function method with phase correlation. By effectively combining the Fourier–Mellin transform with auxiliary-function-based phase correlation, the approach successfully decouples the parameters of the similarity transformation. Experimental results demonstrate that the proposed method significantly outperforms conventional discrete cross-correlation–based Fourier–Mellin approaches in terms of estimation accuracy for scale, rotation, and translation.

0 citationsRead paper

From Sweep to Seam: Interleaved Cross-Block Post-Training Quantization

Aug 10, 2026

This work addresses the performance degradation in deep neural networks caused by error accumulation in conventional block-wise post-training quantization (PTQ), where a single forward pass prevents correction of early quantization errors. To mitigate this issue, the authors propose Interleaved Cross-Block Quantization (ICBQ), which refines local reconstruction by doubly optimizing adjacent “seam” block pairs at block boundaries. Built upon standard block-wise PTQ, ICBQ incorporates cross-block reconstruction within a sliding window, leveraging local shrinkage and smoothness assumptions, and is compatible with GPTQ scheduling. Experimental results demonstrate that ICBQ significantly reduces perplexity under ternary, 2-bit, and 3-bit quantization, effectively preventing performance collapse at extreme compression ratios while seamlessly integrating into existing PTQ pipelines.

0 citationsRead paper

LOGOS: A Living Logic for AI Agent Teams That Evolve With Humans

Jul 12, 2026

This work addresses the lack of controllability and auditability in continuously evolving AI agent teams by proposing a “verifiable human-in-the-loop engineering” paradigm that introduces a pluggable self-evolution and governance layer. The approach compiles multimodal inputs into versioned agent packages and integrates event sourcing, fail-closed verification, and policy-driven authorization mechanisms to ensure that all evolutionary changes require explicit human approval and empirical validation before taking effect. While enabling agents to learn continuously during operation, the system guarantees that the entire evolution process remains auditable, traceable, and ultimately under human control, thereby achieving a unified balance between operational efficiency and trustworthy governance.

0 citationsRead paper

Signs Beat Floats: Low-Rank Double-Binary Adaptation for On-Device Fine-Tuning

May 21, 2026

This work addresses the substantial storage, communication, and computational overheads incurred by floating-point operations in conventional LoRA during on-device fine-tuning of large language models, despite its small parameter count. The authors propose LoRDBA—the first LoRA-compatible method integrating double binarization—replacing low-rank factors with sign matrices and recovering magnitude information via lightweight channel-wise scaling. LoRDBA achieves accuracy comparable to fp16 LoRA while compressing adapter size by over 10×. It incurs at most an 8% increase in prefill latency during inference and requires approximately 1.6× the training memory of fp16 LoRA, significantly outperforming existing low-bit baselines under the same model scale.

0 citationsRead paper