Selected work

Representative Papers

NeuRoute: Logit-Guided Neural Routing for Billion-Scale Vector Search with Sub-Hour Index Construction

Aug 15, 2026

This study addresses the prohibitive indexing time associated with billion-scale vector search by proposing NeuRoute, a learned hashing index framework. The method employs logit-guided neural routing integrated with adaptive multi-bucket probing and centroid-gated early stopping mechanisms. Furthermore, it utilizes a lightweight encoder optimized via a selective similarity preservation objective to enhance retrieval efficiency. Experimental evaluations on the BigANN-1B dataset demonstrate that NeuRoute achieves a Recall@10 of 90.3% and a throughput of 2,414 QPS. Notably, the end-to-end index construction time is reduced to under one hour. These results indicate that NeuRoute significantly accelerates both index building and query performance for large-scale vector retrieval systems, offering a practical solution to the scalability bottlenecks inherent in current high-dimensional indexing approaches.

0 citationsRead paper

Formal Verification of Learned Multi-Agent Communication Policies via Decision Tree Distillation

Jun 17, 2026

This work addresses the lack of formal safety guarantees in neural network policies for multi-agent reinforcement learning, which hinders their deployment in safety-critical applications such as drone swarms. The paper proposes the first end-to-end framework that distills learned communication policies into interpretable decision trees and formally verifies them using probabilistic model checking, ensuring that verified safety properties transfer back to the original policy. Key innovations include a compositional verification approach based on pairwise decomposition and union-bound aggregation, and the introduction of discrete VQ-VIB messages to enhance distillation fidelity and verification efficiency. Evaluated on cooperative tasks with 5–7 drones, the framework satisfies all 18 PCTL temporal logic specifications (collision probability 0.3% < 1%), achieves a distillation fidelity of 97.9% ± 1.2%, and incurs a verification bias of at most 0.6 percentage points.

0 citationsRead paper

When is 3D Worth It? A Resource-Performance Frontier for CNNs and Transformers in Lung CT

Jun 05, 2026

This study systematically evaluates whether the additional computational cost of 3D models is justified over 2D or 2.5D approaches in pulmonary CT analysis. Under a unified training protocol, the authors conduct controlled experiments comparing convolutional neural networks (CNNs) and Vision Transformers (ViTs) across 2D, 2.5D, and 3D input representations using the NLST (n=1,977) and LIDC-IDRI datasets, assessing performance, stability, and resource consumption. The work introduces the first joint dimension–architecture evaluation framework tailored for lung cancer screening, revealing that 3D CNNs suffer from threshold instability and ViTs are prone to degenerate predictions such as all-positive outputs. Results demonstrate that 2.5D CNNs achieve the optimal trade-off between discriminative capability and stability (ROC-AUC 0.682), highlighting the practical advantages of lower-dimensional models in real-world clinical deployment.

0 citationsRead paper

Quantifying Evidential Rigor in Meta-Analytic Corpora: A Simulation-Characterized, Bias-Robust Bayesian Workflow with a Nutrition Case Study

May 31, 2026

Traditional meta-analyses struggle to quantify the strength of evidence for the presence or absence of an effect and cannot adequately assess the sensitivity of conclusions to publication bias or small-study effects. This work proposes a Bayesian evidence auditing framework tailored to meta-analytic corpora, integrating bias-aware models with unbiased baseline specifications through Bayesian model averaging and Bayes factors. It introduces “rigor” as a composite metric that jointly evaluates the strength of evidence for an effect and robustness to bias—allowing null effects to achieve high rigor scores when supported by strong evidence. Built upon Bayesian random-effects models, the approach employs simulation and resampling strategies within the ADEMP framework, including synthetic data generation, registered-report resampling, and contour-enhanced funnel weighting. Applied to nutritional intervention studies, the method frequently attenuates conventional effect estimates, revealing that many nominally significant findings lack robust evidential support. Full reproducible resources are publicly released.

0 citationsRead paper

Decomposing Discrimination: Causal Mediation Analysis for AI-Driven Credit Decisions

Mar 29, 2026

This study addresses the conflation in existing AI-based credit decision systems between direct discrimination and structural inequality mediated through financial features. Building on Pearl’s causal framework, the authors propose a method to identify natural direct and indirect effects under weaker assumptions in the presence of treatment-induced confounding, and derive conservative bounds for the otherwise non-identifiable natural effects. To account for residual confounding along the direct pathway, they integrate E-value sensitivity analysis. A doubly robust augmented inverse probability weighting (AIPW) estimator with cross-fitting is employed to achieve semiparametric efficiency, and the accompanying CausalFair toolkit is publicly released. Applied to 89,465 mortgage applications from New York State in 2022, the analysis reveals that of the 7.9-percentage-point racial denial gap, at least 77% stems from structural inequality, while direct discrimination accounts for no more than a conservative lower bound of 23%.

0 citationsRead paper

Recent publications

Latest Papers

NeuRoute: Logit-Guided Neural Routing for Billion-Scale Vector Search with Sub-Hour Index Construction

Aug 15, 2026

This study addresses the prohibitive indexing time associated with billion-scale vector search by proposing NeuRoute, a learned hashing index framework. The method employs logit-guided neural routing integrated with adaptive multi-bucket probing and centroid-gated early stopping mechanisms. Furthermore, it utilizes a lightweight encoder optimized via a selective similarity preservation objective to enhance retrieval efficiency. Experimental evaluations on the BigANN-1B dataset demonstrate that NeuRoute achieves a Recall@10 of 90.3% and a throughput of 2,414 QPS. Notably, the end-to-end index construction time is reduced to under one hour. These results indicate that NeuRoute significantly accelerates both index building and query performance for large-scale vector retrieval systems, offering a practical solution to the scalability bottlenecks inherent in current high-dimensional indexing approaches.

0 citationsRead paper

Formal Verification of Learned Multi-Agent Communication Policies via Decision Tree Distillation

Jun 17, 2026

This work addresses the lack of formal safety guarantees in neural network policies for multi-agent reinforcement learning, which hinders their deployment in safety-critical applications such as drone swarms. The paper proposes the first end-to-end framework that distills learned communication policies into interpretable decision trees and formally verifies them using probabilistic model checking, ensuring that verified safety properties transfer back to the original policy. Key innovations include a compositional verification approach based on pairwise decomposition and union-bound aggregation, and the introduction of discrete VQ-VIB messages to enhance distillation fidelity and verification efficiency. Evaluated on cooperative tasks with 5–7 drones, the framework satisfies all 18 PCTL temporal logic specifications (collision probability 0.3% < 1%), achieves a distillation fidelity of 97.9% ± 1.2%, and incurs a verification bias of at most 0.6 percentage points.

0 citationsRead paper

When is 3D Worth It? A Resource-Performance Frontier for CNNs and Transformers in Lung CT

Jun 05, 2026

This study systematically evaluates whether the additional computational cost of 3D models is justified over 2D or 2.5D approaches in pulmonary CT analysis. Under a unified training protocol, the authors conduct controlled experiments comparing convolutional neural networks (CNNs) and Vision Transformers (ViTs) across 2D, 2.5D, and 3D input representations using the NLST (n=1,977) and LIDC-IDRI datasets, assessing performance, stability, and resource consumption. The work introduces the first joint dimension–architecture evaluation framework tailored for lung cancer screening, revealing that 3D CNNs suffer from threshold instability and ViTs are prone to degenerate predictions such as all-positive outputs. Results demonstrate that 2.5D CNNs achieve the optimal trade-off between discriminative capability and stability (ROC-AUC 0.682), highlighting the practical advantages of lower-dimensional models in real-world clinical deployment.

0 citationsRead paper

Quantifying Evidential Rigor in Meta-Analytic Corpora: A Simulation-Characterized, Bias-Robust Bayesian Workflow with a Nutrition Case Study

May 31, 2026

Traditional meta-analyses struggle to quantify the strength of evidence for the presence or absence of an effect and cannot adequately assess the sensitivity of conclusions to publication bias or small-study effects. This work proposes a Bayesian evidence auditing framework tailored to meta-analytic corpora, integrating bias-aware models with unbiased baseline specifications through Bayesian model averaging and Bayes factors. It introduces “rigor” as a composite metric that jointly evaluates the strength of evidence for an effect and robustness to bias—allowing null effects to achieve high rigor scores when supported by strong evidence. Built upon Bayesian random-effects models, the approach employs simulation and resampling strategies within the ADEMP framework, including synthetic data generation, registered-report resampling, and contour-enhanced funnel weighting. Applied to nutritional intervention studies, the method frequently attenuates conventional effect estimates, revealing that many nominally significant findings lack robust evidential support. Full reproducible resources are publicly released.

0 citationsRead paper

Decomposing Discrimination: Causal Mediation Analysis for AI-Driven Credit Decisions

Mar 29, 2026

This study addresses the conflation in existing AI-based credit decision systems between direct discrimination and structural inequality mediated through financial features. Building on Pearl’s causal framework, the authors propose a method to identify natural direct and indirect effects under weaker assumptions in the presence of treatment-induced confounding, and derive conservative bounds for the otherwise non-identifiable natural effects. To account for residual confounding along the direct pathway, they integrate E-value sensitivity analysis. A doubly robust augmented inverse probability weighting (AIPW) estimator with cross-fitting is employed to achieve semiparametric efficiency, and the accompanying CausalFair toolkit is publicly released. Applied to 89,465 mortgage applications from New York State in 2022, the analysis reveals that of the 7.9-percentage-point racial denial gap, at least 77% stems from structural inequality, while direct discrimination accounts for no more than a conservative lower bound of 23%.

0 citationsRead paper