Selected work

Representative Papers

Why Stop at Words? Unveiling the Bigger Picture through Line-Level OCR

Aug 29, 2025

To address the high character/word segmentation errors and insufficient contextual modeling in traditional OCR, this paper proposes a line-level OCR paradigm that bypasses explicit character and word segmentation and performs end-to-end recognition directly on full text lines. Methodologically, we introduce a unified sequence-to-sequence framework integrating object detection with deep language modeling. We provide the first systematic empirical validation of the advantages of line-level modeling and release LineOCR, the first fine-grained annotation dataset specifically designed for line-level training and evaluation (251 pages of English documents). Experiments demonstrate that our approach achieves a 5.4% absolute improvement in end-to-end accuracy and a 4× speedup in inference latency, substantially alleviating bottlenecks inherent in conventional “segment-then-recognize” pipelines. This work advances OCR toward a unified perception-and-understanding paradigm.

0 citationsRead paper

Preserve Anything: Controllable Image Synthesis with Object Preservation

Jun 27, 2025

Existing text-to-image (T2I) methods suffer from significant limitations in multi-object fidelity, prompt semantic consistency, and scene controllability. To address these challenges, we propose an N-channel ControlNet framework that jointly models object preservation, background semantic alignment, and explicit layout/illumination control. We introduce the first multimodal benchmark integrating 240K natural images and 18K 3D-synthetic images for comprehensive evaluation. Our method incorporates high-resolution background guidance, illumination consistency constraints, and high-frequency detail superposition, augmented with dedicated modules for color fidelity, detail enhancement, and artifact suppression. Quantitative results demonstrate state-of-the-art performance, achieving an FID of 15.26 and a CLIP-Score of 32.85. User studies confirm substantial improvements—25% in prompt alignment, 19% in photorealism, 13% in artifact suppression, and 14% in aesthetic quality—over prior approaches.

0 citationsRead paper

FisherSFT: Data-Efficient Supervised Fine-Tuning of Language Models Using Information Gain

May 20, 2025

This study addresses the low sample efficiency of supervised fine-tuning (SFT) for large language models (LLMs), aiming to improve generalization to new domains under limited annotation budgets. We propose an information-gain-based data selection framework: (1) the first integration of Fisher information approximation with information-gain maximization for SFT; and (2) a tractable, low-overhead Hessian estimation method via last-layer linearization, coupled with multinomial logistic regression to enable efficient active sampling. Experiments across multiple tasks show that using only 30%–50% of the training data achieves comparable or superior performance to full-data SFT, as validated by both LLM-based automatic evaluation and human assessment—significantly outperforming random sampling and state-of-the-art baselines. Our core contribution is a theoretically grounded, computationally efficient, information-driven sample selection paradigm specifically designed for SFT.

0 citationsRead paper

An Efficient Plugin Method for Metric Optimization of Black-Box Models

Mar 03, 2025

This work addresses the challenge of adapting immutable black-box classifiers to target distributions without access to model internals, training data, or feature representations. We propose a lightweight, training-free, post-hoc method that optimizes arbitrary non-differentiable confusion matrix metrics—such as F1-score and Cohen’s Kappa—using only a small number of predicted class probabilities and corresponding ground-truth labels. Our approach integrates probabilistic calibration with gradient estimation of the confusion matrix, enabling end-to-end optimization via few-shot supervision and iterative black-box queries. The key contribution is the first query-driven framework for direct metric optimization in black-box multi-class classification, requiring no model retraining or architectural assumptions. Experiments across tabular and language tasks demonstrate performance competitive with state-of-the-art methods, with significant gains in F1-score and Precision@k, while maintaining inference latency below 0.1 seconds per thousand samples.

0 citationsRead paper

Logits are All We Need to Adapt Closed Models

Feb 03, 2025

Closed-source large language models (LLMs) cannot expose internal parameters or gradients, limiting adaptation to prompt engineering alone. Method: We propose a lightweight black-box task alignment framework that reformulates closed-source LLM adaptation as a label-noise correction problem. Theoretically, we prove that reliable task transfer can be achieved via probabilistic reweighting of output token logits—requiring no parameter access, gradient computation, or architectural modification. Our plug-and-play logits reweighting module integrates supervised classification–guided reweighting, label-noise-robust learning, and autoregressive probability calibration, enabling efficient fine-tuning with minimal task data. Contribution/Results: Experiments across multiple closed-source LLMs (e.g., GPT-4, Claude) and diverse benchmark datasets demonstrate significant improvements over prompt tuning baselines. Our approach validates that controllable, logits-level intervention substantially enhances the downstream utility and adaptability of closed-source models.

0 citationsRead paper

Recent publications

Latest Papers

Why Stop at Words? Unveiling the Bigger Picture through Line-Level OCR

Aug 29, 2025

To address the high character/word segmentation errors and insufficient contextual modeling in traditional OCR, this paper proposes a line-level OCR paradigm that bypasses explicit character and word segmentation and performs end-to-end recognition directly on full text lines. Methodologically, we introduce a unified sequence-to-sequence framework integrating object detection with deep language modeling. We provide the first systematic empirical validation of the advantages of line-level modeling and release LineOCR, the first fine-grained annotation dataset specifically designed for line-level training and evaluation (251 pages of English documents). Experiments demonstrate that our approach achieves a 5.4% absolute improvement in end-to-end accuracy and a 4× speedup in inference latency, substantially alleviating bottlenecks inherent in conventional “segment-then-recognize” pipelines. This work advances OCR toward a unified perception-and-understanding paradigm.

0 citationsRead paper

Preserve Anything: Controllable Image Synthesis with Object Preservation

Jun 27, 2025

Existing text-to-image (T2I) methods suffer from significant limitations in multi-object fidelity, prompt semantic consistency, and scene controllability. To address these challenges, we propose an N-channel ControlNet framework that jointly models object preservation, background semantic alignment, and explicit layout/illumination control. We introduce the first multimodal benchmark integrating 240K natural images and 18K 3D-synthetic images for comprehensive evaluation. Our method incorporates high-resolution background guidance, illumination consistency constraints, and high-frequency detail superposition, augmented with dedicated modules for color fidelity, detail enhancement, and artifact suppression. Quantitative results demonstrate state-of-the-art performance, achieving an FID of 15.26 and a CLIP-Score of 32.85. User studies confirm substantial improvements—25% in prompt alignment, 19% in photorealism, 13% in artifact suppression, and 14% in aesthetic quality—over prior approaches.

0 citationsRead paper

FisherSFT: Data-Efficient Supervised Fine-Tuning of Language Models Using Information Gain

May 20, 2025

This study addresses the low sample efficiency of supervised fine-tuning (SFT) for large language models (LLMs), aiming to improve generalization to new domains under limited annotation budgets. We propose an information-gain-based data selection framework: (1) the first integration of Fisher information approximation with information-gain maximization for SFT; and (2) a tractable, low-overhead Hessian estimation method via last-layer linearization, coupled with multinomial logistic regression to enable efficient active sampling. Experiments across multiple tasks show that using only 30%–50% of the training data achieves comparable or superior performance to full-data SFT, as validated by both LLM-based automatic evaluation and human assessment—significantly outperforming random sampling and state-of-the-art baselines. Our core contribution is a theoretically grounded, computationally efficient, information-driven sample selection paradigm specifically designed for SFT.

0 citationsRead paper

An Efficient Plugin Method for Metric Optimization of Black-Box Models

Mar 03, 2025

This work addresses the challenge of adapting immutable black-box classifiers to target distributions without access to model internals, training data, or feature representations. We propose a lightweight, training-free, post-hoc method that optimizes arbitrary non-differentiable confusion matrix metrics—such as F1-score and Cohen’s Kappa—using only a small number of predicted class probabilities and corresponding ground-truth labels. Our approach integrates probabilistic calibration with gradient estimation of the confusion matrix, enabling end-to-end optimization via few-shot supervision and iterative black-box queries. The key contribution is the first query-driven framework for direct metric optimization in black-box multi-class classification, requiring no model retraining or architectural assumptions. Experiments across tabular and language tasks demonstrate performance competitive with state-of-the-art methods, with significant gains in F1-score and Precision@k, while maintaining inference latency below 0.1 seconds per thousand samples.

0 citationsRead paper

Logits are All We Need to Adapt Closed Models

Feb 03, 2025

Closed-source large language models (LLMs) cannot expose internal parameters or gradients, limiting adaptation to prompt engineering alone. Method: We propose a lightweight black-box task alignment framework that reformulates closed-source LLM adaptation as a label-noise correction problem. Theoretically, we prove that reliable task transfer can be achieved via probabilistic reweighting of output token logits—requiring no parameter access, gradient computation, or architectural modification. Our plug-and-play logits reweighting module integrates supervised classification–guided reweighting, label-noise-robust learning, and autoregressive probability calibration, enabling efficient fine-tuning with minimal task data. Contribution/Results: Experiments across multiple closed-source LLMs (e.g., GPT-4, Claude) and diverse benchmark datasets demonstrate significant improvements over prompt tuning baselines. Our approach validates that controllable, logits-level intervention substantially enhances the downstream utility and adaptability of closed-source models.

0 citationsRead paper