Selected work

Representative Papers

Predicting When Random Low-Dimensional Reparameterizations Train Neural Networks

Aug 12, 2026

This work addresses the problem of selecting the latent space dimension in randomized low-dimensional reparameterization to ensure neural networks can be efficiently trained into low-loss regions. By characterizing accessibility phase transitions through conic geometry, the authors propose a directionally resolved quadratic theoretical framework that accurately predicts residual errors in random slices. Integrating structured random projections—such as Hadamard or recycled Gaussian mappings—with matrix-free curvature approximations and optimizer state compression, they develop a memory-efficient training framework. The method automatically determines the optimal dimensionality without exhaustive scanning, and empirical results on both vision and language models reveal training phase transitions that align closely with theoretical predictions, substantially outperforming existing approximation strategies that neglect directional information.

0 citationsRead paper

LoRAScan: Detecting Backdoor Prompts in Low-Rank Adapters for Large Language Models via Down-Projection Activation Spikes

Aug 07, 2026

This work addresses the latent backdoor threats in Low-Rank Adaptation (LoRA) by proposing an adapter-aware defense mechanism that operates during inference without modifying adapter parameters. The method innovatively identifies backdoor triggers through sparse spike patterns in the down-projection activations of LoRA, integrating low-rank structural analysis, selection of low-variance insertion points, and real-time anomaly detection to achieve high-precision discrimination at approximately 5% stable insertion points. Experimental results demonstrate that the approach effectively rejects 98.49% of malicious inputs on standard backdoor benchmarks while maintaining an extremely low false positive rate on clean samples, significantly outperforming existing defense strategies.

0 citationsRead paper

Active Learning for Data-Efficient Calibration of Stochastic Simulation Models

Jul 30, 2026

This work addresses the challenge of efficiently calibrating unknown parameters in computationally expensive stochastic simulation models while optimally allocating resources between exploring new input configurations and replicating simulations. To this end, the authors propose a Bayesian active learning framework tailored for posterior density estimation. The approach introduces an uncertainty-aware acquisition criterion and derives two distinct acquisition functions—one for exploration and one for replication—dynamically balancing these strategies to enhance surrogate model construction. Empirical evaluations on both synthetic benchmarks and a real-world epidemiological model demonstrate that the proposed method significantly improves the efficiency of learning the parameter posterior distribution, achieving comparable or superior accuracy with substantially fewer simulation runs.

0 citationsRead paper

Explainability Framework for Policy-Aware Autonomous Agents

Jul 23, 2026

This work proposes an explainability framework grounded in principles from social science to enhance the trustworthiness and transparency of rule-based agents in everyday applications. The framework innovatively integrates a policy-violation penalty mechanism with counterfactual reasoning to systematically generate contrastive natural language explanations—such as “Had this action not been taken, adverse event X would have occurred.” The core reasoning logic is implemented using Answer Set Programming, complemented by Python-based modules for information extraction and natural language generation. User studies demonstrate that the generated explanations significantly improve human understanding of agent decisions, thereby validating both the effectiveness and comprehensibility of the proposed approach.

0 citationsRead paper

The Token Tax of Epistemic Accuracy: Comparing RAG and Long-Context Architectures for Document-Grounded Generative AI Applications

Jun 18, 2026

This study addresses the critical challenge of balancing generation accuracy against computational cost in high-stakes, knowledge-intensive tasks. It systematically compares two document-grounding approaches—Retrieval-Augmented Generation (RAG) and long-context prompting—within the context of manufacturing safety training, evaluating their performance trade-offs. The work introduces two metrics, “cognitive accuracy” and “token tax,” to quantify the benefits and costs of broader evidence access. Experimental results across expert-validated benchmarks, multiple language models, and diverse environments show that long-context prompting achieves a cognitive accuracy of 73.1%, significantly outperforming semantic RAG at 65.4%, yet incurs a 26-fold increase in token consumption per query. These findings reveal a pronounced tension between accuracy and computational efficiency in grounded generation.

0 citationsRead paper

Recent publications

Latest Papers

Predicting When Random Low-Dimensional Reparameterizations Train Neural Networks

Aug 12, 2026

This work addresses the problem of selecting the latent space dimension in randomized low-dimensional reparameterization to ensure neural networks can be efficiently trained into low-loss regions. By characterizing accessibility phase transitions through conic geometry, the authors propose a directionally resolved quadratic theoretical framework that accurately predicts residual errors in random slices. Integrating structured random projections—such as Hadamard or recycled Gaussian mappings—with matrix-free curvature approximations and optimizer state compression, they develop a memory-efficient training framework. The method automatically determines the optimal dimensionality without exhaustive scanning, and empirical results on both vision and language models reveal training phase transitions that align closely with theoretical predictions, substantially outperforming existing approximation strategies that neglect directional information.

0 citationsRead paper

LoRAScan: Detecting Backdoor Prompts in Low-Rank Adapters for Large Language Models via Down-Projection Activation Spikes

Aug 07, 2026

This work addresses the latent backdoor threats in Low-Rank Adaptation (LoRA) by proposing an adapter-aware defense mechanism that operates during inference without modifying adapter parameters. The method innovatively identifies backdoor triggers through sparse spike patterns in the down-projection activations of LoRA, integrating low-rank structural analysis, selection of low-variance insertion points, and real-time anomaly detection to achieve high-precision discrimination at approximately 5% stable insertion points. Experimental results demonstrate that the approach effectively rejects 98.49% of malicious inputs on standard backdoor benchmarks while maintaining an extremely low false positive rate on clean samples, significantly outperforming existing defense strategies.

0 citationsRead paper

Active Learning for Data-Efficient Calibration of Stochastic Simulation Models

Jul 30, 2026

This work addresses the challenge of efficiently calibrating unknown parameters in computationally expensive stochastic simulation models while optimally allocating resources between exploring new input configurations and replicating simulations. To this end, the authors propose a Bayesian active learning framework tailored for posterior density estimation. The approach introduces an uncertainty-aware acquisition criterion and derives two distinct acquisition functions—one for exploration and one for replication—dynamically balancing these strategies to enhance surrogate model construction. Empirical evaluations on both synthetic benchmarks and a real-world epidemiological model demonstrate that the proposed method significantly improves the efficiency of learning the parameter posterior distribution, achieving comparable or superior accuracy with substantially fewer simulation runs.

0 citationsRead paper

Explainability Framework for Policy-Aware Autonomous Agents

Jul 23, 2026

This work proposes an explainability framework grounded in principles from social science to enhance the trustworthiness and transparency of rule-based agents in everyday applications. The framework innovatively integrates a policy-violation penalty mechanism with counterfactual reasoning to systematically generate contrastive natural language explanations—such as “Had this action not been taken, adverse event X would have occurred.” The core reasoning logic is implemented using Answer Set Programming, complemented by Python-based modules for information extraction and natural language generation. User studies demonstrate that the generated explanations significantly improve human understanding of agent decisions, thereby validating both the effectiveness and comprehensibility of the proposed approach.

0 citationsRead paper

The Token Tax of Epistemic Accuracy: Comparing RAG and Long-Context Architectures for Document-Grounded Generative AI Applications

Jun 18, 2026

This study addresses the critical challenge of balancing generation accuracy against computational cost in high-stakes, knowledge-intensive tasks. It systematically compares two document-grounding approaches—Retrieval-Augmented Generation (RAG) and long-context prompting—within the context of manufacturing safety training, evaluating their performance trade-offs. The work introduces two metrics, “cognitive accuracy” and “token tax,” to quantify the benefits and costs of broader evidence access. Experimental results across expert-validated benchmarks, multiple language models, and diverse environments show that long-context prompting achieves a cognitive accuracy of 73.1%, significantly outperforming semantic RAG at 65.4%, yet incurs a 26-fold increase in token consumption per query. These findings reveal a pronounced tension between accuracy and computational efficiency in grounded generation.

0 citationsRead paper