Selected work

Representative Papers

Knowledge Distillation from Large Reasoning Models to Compact Student Models: A Case Study on the John O Bryan Mathematics Competition

Jun 29, 2026

This study addresses the challenge of efficiently transferring mathematical reasoning capabilities from large inference models to smaller student models while mitigating overfitting and preserving reasoning quality. Leveraging problems from the John O’Brien Mathematics Competition, the authors construct a chain-of-thought (CoT) training corpus generated by a dual-agent framework and perform knowledge distillation fine-tuning of Qwen2.5-7B using the MLX framework with LoRA adapters on Apple Silicon. The research reveals a strong positive correlation between reasoning response length and problem-solving accuracy and demonstrates that constraining training iterations effectively alleviates overfitting. The resulting model achieves an accuracy of 69.43% (+4.76%) on the competition dataset and generalizes to 73.1% accuracy on the MATH-500 benchmark, significantly enhancing complex reasoning performance in compact models.

0 citationsRead paper

AutoFormBench: Benchmark Dataset for Automating Form Understanding

Mar 31, 2026

This study addresses the challenge of automating the processing of structured documents—such as government forms, medical records, and corporate invoices—in real-world scenarios where form layouts exhibit high variability. To this end, the authors introduce a multi-domain benchmark dataset comprising 407 real-world forms and present the first systematic evaluation of both traditional OpenCV-based methods and a suite of YOLO models—including YOLOv8, YOLOv11, YOLOv26-s, and YOLOv26-l—for the task of detecting fillable elements. Experimental results demonstrate that YOLOv11 consistently achieves the highest F1 scores and Jaccard accuracy across all element categories and tolerance levels, confirming its superiority in form understanding tasks. The released dataset and comprehensive performance comparison establish a valuable benchmark for future research in document intelligence.

0 citationsRead paper

Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents

Mar 30, 2026

This work addresses a critical gap in the evaluation of large language model (LLM) agents, which has predominantly focused on single-task success rates while neglecting reliability over extended, repeated execution. The study proposes the first reliability assessment framework tailored for long-horizon tasks, introducing four novel metrics—reliability decay curves, variance amplification factor, graceful degradation score, and collapse onset point—to establish reliability as a core dimension distinct from raw capability. Through large-scale experiments encompassing 10 models, 23,392 task replays, and 396 cross-domain tasks, combined with hierarchical time-bucket design and statistical analysis, the authors reveal significant discrepancies between model capability and reliability rankings, find that state-of-the-art models exhibit collapse rates as high as 19%, demonstrate that memory scaffolds commonly impair long-term performance, and uncover domain-specific patterns of reliability decay.

0 citationsRead paper

Severe Domain Shift in Skeleton-Based Action Recognition:A Study of Uncertainty Failure in Real-World Gym Environments

Mar 16, 2026

This study addresses the performance degradation and safety risks in skeleton-based action recognition when transferring models from controlled multi-view 3D environments to real-world monocular 2D settings, where compound domain shifts induce severe model failure. The work reveals, for the first time, that high out-of-distribution (OOD) detection AUROC does not guarantee safety in selective classification. To evaluate zero-shot transfer, the authors introduce Gym2D and a real-world fitness dataset derived from UCF101, demonstrating catastrophic performance collapse (accuracy drops from 63.2% to 1.6%) and miscalibrated uncertainty. They propose a novel OOD detection signal combining energy scores and Mahalanobis distance, along with a lightweight fine-tuned gating mechanism for risk-aware rejection decisions. Experiments show the approach significantly reduces high-confidence erroneous predictions, thereby enhancing deployment safety in open-world scenarios.

0 citationsRead paper

Visual Instruction Pretraining for Domain-Specific Foundation Models

Sep 22, 2025

Existing vision foundation models lack a backward guidance mechanism whereby high-level reasoning informs low-level perceptual feature learning, resulting in an incomplete perception–reasoning–generation loop. Method: We propose Vision Instruction Pre-training (ViTP), a novel paradigm built upon Vision Transformers and trained end-to-end on large-scale, domain-specific vision–language instruction data via instruction following and cross-modal alignment. Crucially, we introduce Vision Robustness Learning (VRL), which explicitly encourages the model to extract domain-relevant, interference-resistant, discriminative features from sparse visual tokens. Contribution/Results: ViTP establishes the first end-to-end, domain-specialized foundation model capable of reasoning-driven perceptual optimization. Evaluated on 16 remote sensing and medical imaging benchmarks, it achieves state-of-the-art performance, significantly improving generalization and semantic understanding across downstream tasks.

0 citationsRead paper

Recent publications

Latest Papers

Knowledge Distillation from Large Reasoning Models to Compact Student Models: A Case Study on the John O Bryan Mathematics Competition

Jun 29, 2026

This study addresses the challenge of efficiently transferring mathematical reasoning capabilities from large inference models to smaller student models while mitigating overfitting and preserving reasoning quality. Leveraging problems from the John O’Brien Mathematics Competition, the authors construct a chain-of-thought (CoT) training corpus generated by a dual-agent framework and perform knowledge distillation fine-tuning of Qwen2.5-7B using the MLX framework with LoRA adapters on Apple Silicon. The research reveals a strong positive correlation between reasoning response length and problem-solving accuracy and demonstrates that constraining training iterations effectively alleviates overfitting. The resulting model achieves an accuracy of 69.43% (+4.76%) on the competition dataset and generalizes to 73.1% accuracy on the MATH-500 benchmark, significantly enhancing complex reasoning performance in compact models.

0 citationsRead paper

AutoFormBench: Benchmark Dataset for Automating Form Understanding

Mar 31, 2026

This study addresses the challenge of automating the processing of structured documents—such as government forms, medical records, and corporate invoices—in real-world scenarios where form layouts exhibit high variability. To this end, the authors introduce a multi-domain benchmark dataset comprising 407 real-world forms and present the first systematic evaluation of both traditional OpenCV-based methods and a suite of YOLO models—including YOLOv8, YOLOv11, YOLOv26-s, and YOLOv26-l—for the task of detecting fillable elements. Experimental results demonstrate that YOLOv11 consistently achieves the highest F1 scores and Jaccard accuracy across all element categories and tolerance levels, confirming its superiority in form understanding tasks. The released dataset and comprehensive performance comparison establish a valuable benchmark for future research in document intelligence.

0 citationsRead paper

Beyond pass@1: A Reliability Science Framework for Long-Horizon LLM Agents

Mar 30, 2026

This work addresses a critical gap in the evaluation of large language model (LLM) agents, which has predominantly focused on single-task success rates while neglecting reliability over extended, repeated execution. The study proposes the first reliability assessment framework tailored for long-horizon tasks, introducing four novel metrics—reliability decay curves, variance amplification factor, graceful degradation score, and collapse onset point—to establish reliability as a core dimension distinct from raw capability. Through large-scale experiments encompassing 10 models, 23,392 task replays, and 396 cross-domain tasks, combined with hierarchical time-bucket design and statistical analysis, the authors reveal significant discrepancies between model capability and reliability rankings, find that state-of-the-art models exhibit collapse rates as high as 19%, demonstrate that memory scaffolds commonly impair long-term performance, and uncover domain-specific patterns of reliability decay.

0 citationsRead paper

Severe Domain Shift in Skeleton-Based Action Recognition:A Study of Uncertainty Failure in Real-World Gym Environments

Mar 16, 2026

This study addresses the performance degradation and safety risks in skeleton-based action recognition when transferring models from controlled multi-view 3D environments to real-world monocular 2D settings, where compound domain shifts induce severe model failure. The work reveals, for the first time, that high out-of-distribution (OOD) detection AUROC does not guarantee safety in selective classification. To evaluate zero-shot transfer, the authors introduce Gym2D and a real-world fitness dataset derived from UCF101, demonstrating catastrophic performance collapse (accuracy drops from 63.2% to 1.6%) and miscalibrated uncertainty. They propose a novel OOD detection signal combining energy scores and Mahalanobis distance, along with a lightweight fine-tuned gating mechanism for risk-aware rejection decisions. Experiments show the approach significantly reduces high-confidence erroneous predictions, thereby enhancing deployment safety in open-world scenarios.

0 citationsRead paper

Visual Instruction Pretraining for Domain-Specific Foundation Models

Sep 22, 2025

Existing vision foundation models lack a backward guidance mechanism whereby high-level reasoning informs low-level perceptual feature learning, resulting in an incomplete perception–reasoning–generation loop. Method: We propose Vision Instruction Pre-training (ViTP), a novel paradigm built upon Vision Transformers and trained end-to-end on large-scale, domain-specific vision–language instruction data via instruction following and cross-modal alignment. Crucially, we introduce Vision Robustness Learning (VRL), which explicitly encourages the model to extract domain-relevant, interference-resistant, discriminative features from sparse visual tokens. Contribution/Results: ViTP establishes the first end-to-end, domain-specialized foundation model capable of reasoning-driven perceptual optimization. Evaluated on 16 remote sensing and medical imaging benchmarks, it achieves state-of-the-art performance, significantly improving generalization and semantic understanding across downstream tasks.

0 citationsRead paper