Selected work

Representative Papers

Explanation Stability of Test-Time Adaptation in Computational Pathology: A Large-Scale Benchmark

Aug 07, 2026

This study addresses the underexplored trade-off between accuracy and explanation stability in test-time adaptation (TTA) for computational pathology. While TTA can enhance model accuracy, it may compromise the stability of model explanations, thereby undermining clinical trustworthiness. The authors present the first large-scale systematic evaluation of 17 TTA methods across the Camelyon17 and NCT-CRC-HE datasets, examining their impact on explanation stability using four attribution techniques and encompassing convolutional networks, Vision Transformers, and foundation models in pathology—totaling 2,958 experiments. Their findings reveal that explanation stability is decoupled from predictive accuracy and should be treated as an independent reliability metric for TTA. Notably, TTA strategies that freeze the backbone yield the most stable explanations, whereas continual adaptation approaches like CoTTA induce significant explanation drift. Convolutional architectures exhibit greater sensitivity to TTA-induced instability than Transformers. The complete benchmark and evaluation protocol are publicly released.

0 citationsRead paper

RECON: Benchmarking Agent Memory for Compositional Reasoning over Long Contexts

Jul 18, 2026

This work addresses the challenge that current AI agents struggle to effectively memorize, track, and reason about the evolution of facts across extended interactive contexts, particularly failing to accurately assess the causal impact of fact revisions on downstream conclusions. To this end, the paper introduces RECON, a novel benchmark comprising 24 cross-domain narrative documents (each 50k–100k words), which for the first time focuses on causal tracing and counterfactual reasoning under factual changes. RECON systematically evaluates agents’ compositional reasoning capabilities across six task types, including reconstructing multi-hop evidence chains, handling cascading failures, and resolving source conflicts. Using an integrated retrieval-and-reasoning evaluation framework, experiments reveal that even the strongest non-oracle systems achieve only 22.4% accuracy, highlighting significant limitations of existing approaches in scenarios involving dynamic knowledge evolution.

0 citationsRead paper

Eyes All Around: Design and Analysis of 360-Degree LiDAR Perception Using Equivariant Feature Learning in Unstructured Traffic

May 30, 2026

This work addresses the challenges posed by unstructured urban traffic—such as heterogeneous road users, frequent occlusions, complex motion patterns, and non-standardized road layouts—by proposing a 360-degree LiDAR-based panoramic perception framework. The method integrates azimuth-aware sectorization with rotation-equivariant sparse convolutions to achieve robust 3D object detection in complex urban environments. As the first 360-degree perception system validated on real-world unstructured urban traffic data from India, it demonstrates strong performance, achieving AP scores of 92.02 and 90.51 for cars. While detection performance for pedestrians and cyclists is comparatively lower due to their small scale and high shape variability, the results remain practically viable for real-world deployment.

0 citationsRead paper

Two Steps Are All You Need: Efficient 3D Point Cloud Anomaly Detection with Consistency Models

May 06, 2026

This work addresses the high computational cost and inference latency of existing diffusion-based 3D point cloud anomaly detection methods, which hinder deployment on edge devices. The authors propose a novel paradigm based on consistency models that directly predicts anomaly-free geometric structures in just one or two forward passes. A reconstruction-oriented hybrid loss function is introduced to guide the model toward learning normal patterns effectively. The proposed approach substantially reduces both inference steps and computational overhead, achieving up to an 80-fold speedup over the current state-of-the-art method under CPU-only conditions. It attains competitive detection performance with I-AUROC scores of 76.20% on Anomaly-ShapeNet and 72.80% on Real3DAD, demonstrating a strong balance between efficiency and accuracy.

0 citationsRead paper

MANOJAVAM: A Scalable, Unified FPGA Accelerator for Matrix Multiplication and Singular Value Decomposition in Principal Component Analysis

May 02, 2026

This work addresses the computational bottleneck in principal component analysis (PCA)—particularly in matrix multiplication and singular value decomposition (SVD)—which hinders performance in applications such as hyperspectral imaging and genomics due to limitations in existing hardware regarding uniformity, scalability, and energy efficiency. To overcome this, the authors propose MANOJAVAM, a unified and scalable FPGA-based accelerator architecture that, for the first time, integrates both matrix multiplication and SVD within a single structure. Leveraging tiled streaming processing, a parallel Jacobi algorithm, a T×T systolic array, CORDIC-based pipelined rotation units, and a mode-aware two-level cache, MANOJAVAM supports arbitrary input dimensions. Implemented on a Virtex Ultrascale+ FPGA at 434 MHz, it achieves up to 22.75× lower SVD latency and 42.14× better energy efficiency compared to an NVIDIA A6000 GPU, making it suitable for both high-performance and edge computing scenarios.

0 citationsRead paper

Recent publications

Latest Papers

Explanation Stability of Test-Time Adaptation in Computational Pathology: A Large-Scale Benchmark

Aug 07, 2026

This study addresses the underexplored trade-off between accuracy and explanation stability in test-time adaptation (TTA) for computational pathology. While TTA can enhance model accuracy, it may compromise the stability of model explanations, thereby undermining clinical trustworthiness. The authors present the first large-scale systematic evaluation of 17 TTA methods across the Camelyon17 and NCT-CRC-HE datasets, examining their impact on explanation stability using four attribution techniques and encompassing convolutional networks, Vision Transformers, and foundation models in pathology—totaling 2,958 experiments. Their findings reveal that explanation stability is decoupled from predictive accuracy and should be treated as an independent reliability metric for TTA. Notably, TTA strategies that freeze the backbone yield the most stable explanations, whereas continual adaptation approaches like CoTTA induce significant explanation drift. Convolutional architectures exhibit greater sensitivity to TTA-induced instability than Transformers. The complete benchmark and evaluation protocol are publicly released.

0 citationsRead paper

RECON: Benchmarking Agent Memory for Compositional Reasoning over Long Contexts

Jul 18, 2026

This work addresses the challenge that current AI agents struggle to effectively memorize, track, and reason about the evolution of facts across extended interactive contexts, particularly failing to accurately assess the causal impact of fact revisions on downstream conclusions. To this end, the paper introduces RECON, a novel benchmark comprising 24 cross-domain narrative documents (each 50k–100k words), which for the first time focuses on causal tracing and counterfactual reasoning under factual changes. RECON systematically evaluates agents’ compositional reasoning capabilities across six task types, including reconstructing multi-hop evidence chains, handling cascading failures, and resolving source conflicts. Using an integrated retrieval-and-reasoning evaluation framework, experiments reveal that even the strongest non-oracle systems achieve only 22.4% accuracy, highlighting significant limitations of existing approaches in scenarios involving dynamic knowledge evolution.

0 citationsRead paper

Eyes All Around: Design and Analysis of 360-Degree LiDAR Perception Using Equivariant Feature Learning in Unstructured Traffic

May 30, 2026

This work addresses the challenges posed by unstructured urban traffic—such as heterogeneous road users, frequent occlusions, complex motion patterns, and non-standardized road layouts—by proposing a 360-degree LiDAR-based panoramic perception framework. The method integrates azimuth-aware sectorization with rotation-equivariant sparse convolutions to achieve robust 3D object detection in complex urban environments. As the first 360-degree perception system validated on real-world unstructured urban traffic data from India, it demonstrates strong performance, achieving AP scores of 92.02 and 90.51 for cars. While detection performance for pedestrians and cyclists is comparatively lower due to their small scale and high shape variability, the results remain practically viable for real-world deployment.

0 citationsRead paper

Two Steps Are All You Need: Efficient 3D Point Cloud Anomaly Detection with Consistency Models

May 06, 2026

This work addresses the high computational cost and inference latency of existing diffusion-based 3D point cloud anomaly detection methods, which hinder deployment on edge devices. The authors propose a novel paradigm based on consistency models that directly predicts anomaly-free geometric structures in just one or two forward passes. A reconstruction-oriented hybrid loss function is introduced to guide the model toward learning normal patterns effectively. The proposed approach substantially reduces both inference steps and computational overhead, achieving up to an 80-fold speedup over the current state-of-the-art method under CPU-only conditions. It attains competitive detection performance with I-AUROC scores of 76.20% on Anomaly-ShapeNet and 72.80% on Real3DAD, demonstrating a strong balance between efficiency and accuracy.

0 citationsRead paper

MANOJAVAM: A Scalable, Unified FPGA Accelerator for Matrix Multiplication and Singular Value Decomposition in Principal Component Analysis

May 02, 2026

This work addresses the computational bottleneck in principal component analysis (PCA)—particularly in matrix multiplication and singular value decomposition (SVD)—which hinders performance in applications such as hyperspectral imaging and genomics due to limitations in existing hardware regarding uniformity, scalability, and energy efficiency. To overcome this, the authors propose MANOJAVAM, a unified and scalable FPGA-based accelerator architecture that, for the first time, integrates both matrix multiplication and SVD within a single structure. Leveraging tiled streaming processing, a parallel Jacobi algorithm, a T×T systolic array, CORDIC-based pipelined rotation units, and a mode-aware two-level cache, MANOJAVAM supports arbitrary input dimensions. Implemented on a Virtex Ultrascale+ FPGA at 434 MHz, it achieves up to 22.75× lower SVD latency and 42.14× better energy efficiency compared to an NVIDIA A6000 GPU, making it suitable for both high-performance and edge computing scenarios.

0 citationsRead paper