Selected work

Representative Papers

On Function-Correcting Lee Metric Codes with Data Protection

Oct 08, 2026

This study addresses the joint protection of data and function values under the Lee metric by proposing a framework of function-correcting codes with data protection. Methodologically, it integrates algebraic coding theory, Plotkin bound analysis, and combinatorial construction techniques to design an asymmetric error-correction mechanism that provides stronger protection for functions than for data. The main contributions include establishing theoretical lower and upper bounds on optimal redundancy, deriving explicit redundancy upper bounds for specific function classes, and demonstrating that the proposed framework naturally extends to the Hamming metric. Ultimately, this work achieves a unified coding scheme that simultaneously accommodates data and function error correction, thereby providing both a theoretical foundation and a constructive paradigm for asymmetric protection coding.

0 citationsRead paper

Structured Visual Target Learning For Cross-Subject eeg-to-image retrieval

Sep 29, 2026

This study addresses the challenge in cross-subject EEG-based image retrieval where source-domain neural representations struggle to align with the visual space of unseen subjects, proposing a novel framework optimized from the visual target side. The method transforms perceptual encoder patch grids into compact visual views, which are aggregated via a block-structured content-dependent router and jointly trained with an EEG encoder using contrastive learning and MMD regularization. Additionally, a training-free representation refinement strategy is designed to achieve cross-domain alignment of frozen embeddings without updating the encoders. Evaluated on the THINGS-EEG2 dataset, the proposed approach attains a Top-1 accuracy of 48.1%, outperforming the strongest baseline by 18.5% and significantly improving retrieval performance across all held-out subjects.

0 citationsRead paper

Masking Frequent Tokens Sharpens Direct Preference Optimization

Sep 26, 2026

This study addresses the issue that high-frequency symmetric tokens dilute preference signals in Direct Preference Optimization (DPO), leading to gradient entanglement and degraded optimization efficacy. To mitigate this, we propose Anisotropic DPO, which introduces a vocabulary frequency-based hard masking mechanism coupled with a non-uniform weighting strategy. This approach suppresses the reward contributions of high-frequency shared tokens, thereby amplifying discriminative preference signals. Notably, the method incurs no additional parameters or computational overhead, as it solely adjusts token-level objective weights to alleviate gradient interference. Experimental evaluations on benchmarks such as AlpacaEval demonstrate that Anisotropic DPO significantly outperforms standard DPO, effectively enhancing model robustness against noise. Ultimately, this work provides a zero-cost, plug-and-play solution for preference optimization.

0 citationsRead paper

Understanding and Exploiting Anisotropy in Post-Training

Sep 26, 2026

In the post-training of large language models, representational anisotropy is frequently mischaracterized as a deficiency, leaving its functional specialization and interaction mechanisms with supervised fine-tuning (SFT) and reinforcement learning (RL) poorly understood. This work reveals that a small subset of channels constitutes the foundation for linguistic coherence. Accordingly, we propose SphereGate, which safeguards these core channels through a constrained gain mechanism while achieving efficient parameter-efficient fine-tuning via residual anomaly detection and activation-weighted gradient constraints. Introducing merely 0.1M additional parameters, our method surpasses baselines by 2 to 7.3 points on MATH-500, yielding performance comparable to full-model GRPO. Ultimately, this study establishes a novel paradigm for understanding and leveraging representational anisotropy in large model post-training.

0 citationsRead paper

MEVL-STP: Multi-Encoder and Vision Language Model for Arbitrarily Shaped Scene Text Spotting

Sep 23, 2026

This study addresses the recognition failure of arbitrary-shaped text in natural scenes caused by detection and localization errors. To this end, we propose a two-stage end-to-end recognition framework that integrates multi-encoder segmentation with vision-language models. Methodologically, six visual encoders, including CLIP, are kept frozen and combined with FPN and PSENet, leveraging feature orthogonality to prevent representation homogenization. Furthermore, Qwen3-VL is incorporated and fine-tuned via LoRA to eliminate background interference. Evaluated on the CTW1500 dataset without synthetic data, the proposed framework achieves state-of-the-art performance, yielding a detection F-measure of 91.99% and an end-to-end H-mean of 85.86%.

0 citationsRead paper

Recent publications

Latest Papers

On Function-Correcting Lee Metric Codes with Data Protection

Oct 08, 2026

This study addresses the joint protection of data and function values under the Lee metric by proposing a framework of function-correcting codes with data protection. Methodologically, it integrates algebraic coding theory, Plotkin bound analysis, and combinatorial construction techniques to design an asymmetric error-correction mechanism that provides stronger protection for functions than for data. The main contributions include establishing theoretical lower and upper bounds on optimal redundancy, deriving explicit redundancy upper bounds for specific function classes, and demonstrating that the proposed framework naturally extends to the Hamming metric. Ultimately, this work achieves a unified coding scheme that simultaneously accommodates data and function error correction, thereby providing both a theoretical foundation and a constructive paradigm for asymmetric protection coding.

0 citationsRead paper

Structured Visual Target Learning For Cross-Subject eeg-to-image retrieval

Sep 29, 2026

This study addresses the challenge in cross-subject EEG-based image retrieval where source-domain neural representations struggle to align with the visual space of unseen subjects, proposing a novel framework optimized from the visual target side. The method transforms perceptual encoder patch grids into compact visual views, which are aggregated via a block-structured content-dependent router and jointly trained with an EEG encoder using contrastive learning and MMD regularization. Additionally, a training-free representation refinement strategy is designed to achieve cross-domain alignment of frozen embeddings without updating the encoders. Evaluated on the THINGS-EEG2 dataset, the proposed approach attains a Top-1 accuracy of 48.1%, outperforming the strongest baseline by 18.5% and significantly improving retrieval performance across all held-out subjects.

0 citationsRead paper

Masking Frequent Tokens Sharpens Direct Preference Optimization

Sep 26, 2026

This study addresses the issue that high-frequency symmetric tokens dilute preference signals in Direct Preference Optimization (DPO), leading to gradient entanglement and degraded optimization efficacy. To mitigate this, we propose Anisotropic DPO, which introduces a vocabulary frequency-based hard masking mechanism coupled with a non-uniform weighting strategy. This approach suppresses the reward contributions of high-frequency shared tokens, thereby amplifying discriminative preference signals. Notably, the method incurs no additional parameters or computational overhead, as it solely adjusts token-level objective weights to alleviate gradient interference. Experimental evaluations on benchmarks such as AlpacaEval demonstrate that Anisotropic DPO significantly outperforms standard DPO, effectively enhancing model robustness against noise. Ultimately, this work provides a zero-cost, plug-and-play solution for preference optimization.

0 citationsRead paper

Understanding and Exploiting Anisotropy in Post-Training

Sep 26, 2026

In the post-training of large language models, representational anisotropy is frequently mischaracterized as a deficiency, leaving its functional specialization and interaction mechanisms with supervised fine-tuning (SFT) and reinforcement learning (RL) poorly understood. This work reveals that a small subset of channels constitutes the foundation for linguistic coherence. Accordingly, we propose SphereGate, which safeguards these core channels through a constrained gain mechanism while achieving efficient parameter-efficient fine-tuning via residual anomaly detection and activation-weighted gradient constraints. Introducing merely 0.1M additional parameters, our method surpasses baselines by 2 to 7.3 points on MATH-500, yielding performance comparable to full-model GRPO. Ultimately, this study establishes a novel paradigm for understanding and leveraging representational anisotropy in large model post-training.

0 citationsRead paper

MEVL-STP: Multi-Encoder and Vision Language Model for Arbitrarily Shaped Scene Text Spotting

Sep 23, 2026

This study addresses the recognition failure of arbitrary-shaped text in natural scenes caused by detection and localization errors. To this end, we propose a two-stage end-to-end recognition framework that integrates multi-encoder segmentation with vision-language models. Methodologically, six visual encoders, including CLIP, are kept frozen and combined with FPN and PSENet, leveraging feature orthogonality to prevent representation homogenization. Furthermore, Qwen3-VL is incorporated and fine-tuned via LoRA to eliminate background interference. Evaluated on the CTW1500 dataset without synthetic data, the proposed framework achieves state-of-the-art performance, yielding a detection F-measure of 91.99% and an end-to-end H-mean of 85.86%.

0 citationsRead paper