Selected work

Representative Papers

Dataset Identity, Not Novelty: The Source of an Inflated OOD Detection Gain

Oct 01, 2026

This study addresses the inflated out-of-distribution (OOD) detection performance in existing methods, which stems from models fitting dataset identity rather than genuine novelty. To resolve this, we introduce a novel whole-dataset hold-out protocol that decouples identity bias from novelty bias. By integrating posterior detection, feature-space directional fitting, and closed-form mathematical derivations, we quantify the inflation of reported predictive gains. Our analysis reveals that conventionally reported improvements are largely spurious and that model capacity is not the primary contributing factor. Furthermore, we establish that a single constant baseline serves as an upper bound for genuine performance gains. This baseline remains robust on a predefined validation set, effectively recalibrating the evaluation standards within the field.

0 citationsRead paper

Verdict Instability of OOD Scores under Reference Resampling

Sep 01, 2026

研究通过重采样参考集测量OOD分数判决不稳定性，提出一种无参数闭式方法，并分析了类不平衡下判决不稳定性的特性。

0 citationsRead paper

Towards Value-Constrained Credit Assignment in Fully Delegated AI Cooperatives

Jun 26, 2026

This work addresses the challenge of achieving fair reward allocation and precise contribution attribution in fully delegated AI collaborative systems under heterogeneous human value constraints. The authors propose a value-constrained credit assignment framework that leverages value-conditioned gradient filtering and traversal-based learning to enable fine-grained attribution while preserving explicit gradient pathways. By integrating online marginal contribution signals with cumulative payoff settlement, the method effectively supports efficient collaboration among agents with diverse value preferences. Experimental results demonstrate that the proposed framework significantly outperforms conventional federated averaging approaches, maintaining model performance while avoiding the quality degradation commonly associated with aggregated learning.

0 citationsRead paper

TL++: Accuracy and Privacy Preserving Traversal Learning for Distributed Intelligent Systems

Jun 24, 2026

This work addresses the high communication overhead in federated learning, the inability of conventional split learning to replicate centralized mini-batch gradient dynamics, and its associated privacy risks by proposing TL++, a dual-mode traversing learning framework. TL++ constructs virtual batches across nodes to faithfully reproduce centralized training dynamics. In its base mode, it exchanges only activations and gradients at the cut layer; in its secure mode, it further integrates secret sharing to provide activation-level privacy guarantees. TL++ achieves, for the first time in split learning, accuracy nearly matching that of centralized training—reaching 91.41% (base) and 90.93% (secure) on CIFAR-10—outperforming the strongest baseline by over 12 percentage points while reducing communication overhead by 13.1×, and demonstrates strong performance on the PubMedQA task as well.

0 citationsRead paper

Field Validation of a Multi-Resolution ConvLSTM Framework for Retaining Wall Deformation Prediction

Jun 03, 2026

This study addresses the challenge of accurately predicting retaining wall deformation during staged excavation of foundation pits. The authors propose a multi-resolution ConvLSTM framework that integrates Gaussian noise-augmented numerical simulation data with a stacked ensemble strategy to model temporal dynamics across multiple time scales. Notably, this approach achieves high-precision predictions of wall deformation under diverse engineering conditions without requiring any field-measured data for training—relying solely on simulated and augmented data. Validation against monitoring data from 34 points across 11 construction sites in Korea demonstrates an average absolute error of 1.4 mm and a coefficient of determination (R²) of 0.93. The model reliably forecasts deformations induced by subsequent 5.0-meter excavation stages, significantly enhancing predictive generalizability and practical applicability in real-world geotechnical engineering scenarios.

0 citationsRead paper

Recent publications

Latest Papers

Dataset Identity, Not Novelty: The Source of an Inflated OOD Detection Gain

Oct 01, 2026

This study addresses the inflated out-of-distribution (OOD) detection performance in existing methods, which stems from models fitting dataset identity rather than genuine novelty. To resolve this, we introduce a novel whole-dataset hold-out protocol that decouples identity bias from novelty bias. By integrating posterior detection, feature-space directional fitting, and closed-form mathematical derivations, we quantify the inflation of reported predictive gains. Our analysis reveals that conventionally reported improvements are largely spurious and that model capacity is not the primary contributing factor. Furthermore, we establish that a single constant baseline serves as an upper bound for genuine performance gains. This baseline remains robust on a predefined validation set, effectively recalibrating the evaluation standards within the field.

0 citationsRead paper

Verdict Instability of OOD Scores under Reference Resampling

Sep 01, 2026

研究通过重采样参考集测量OOD分数判决不稳定性，提出一种无参数闭式方法，并分析了类不平衡下判决不稳定性的特性。

0 citationsRead paper

Towards Value-Constrained Credit Assignment in Fully Delegated AI Cooperatives

Jun 26, 2026

This work addresses the challenge of achieving fair reward allocation and precise contribution attribution in fully delegated AI collaborative systems under heterogeneous human value constraints. The authors propose a value-constrained credit assignment framework that leverages value-conditioned gradient filtering and traversal-based learning to enable fine-grained attribution while preserving explicit gradient pathways. By integrating online marginal contribution signals with cumulative payoff settlement, the method effectively supports efficient collaboration among agents with diverse value preferences. Experimental results demonstrate that the proposed framework significantly outperforms conventional federated averaging approaches, maintaining model performance while avoiding the quality degradation commonly associated with aggregated learning.

0 citationsRead paper

TL++: Accuracy and Privacy Preserving Traversal Learning for Distributed Intelligent Systems

Jun 24, 2026

This work addresses the high communication overhead in federated learning, the inability of conventional split learning to replicate centralized mini-batch gradient dynamics, and its associated privacy risks by proposing TL++, a dual-mode traversing learning framework. TL++ constructs virtual batches across nodes to faithfully reproduce centralized training dynamics. In its base mode, it exchanges only activations and gradients at the cut layer; in its secure mode, it further integrates secret sharing to provide activation-level privacy guarantees. TL++ achieves, for the first time in split learning, accuracy nearly matching that of centralized training—reaching 91.41% (base) and 90.93% (secure) on CIFAR-10—outperforming the strongest baseline by over 12 percentage points while reducing communication overhead by 13.1×, and demonstrates strong performance on the PubMedQA task as well.

0 citationsRead paper

Field Validation of a Multi-Resolution ConvLSTM Framework for Retaining Wall Deformation Prediction

Jun 03, 2026

This study addresses the challenge of accurately predicting retaining wall deformation during staged excavation of foundation pits. The authors propose a multi-resolution ConvLSTM framework that integrates Gaussian noise-augmented numerical simulation data with a stacked ensemble strategy to model temporal dynamics across multiple time scales. Notably, this approach achieves high-precision predictions of wall deformation under diverse engineering conditions without requiring any field-measured data for training—relying solely on simulated and augmented data. Validation against monitoring data from 34 points across 11 construction sites in Korea demonstrates an average absolute error of 1.4 mm and a coefficient of determination (R²) of 0.93. The model reliably forecasts deformations induced by subsequent 5.0-meter excavation stages, significantly enhancing predictive generalizability and practical applicability in real-world geotechnical engineering scenarios.

0 citationsRead paper