2026

Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize

Chashi Mahiul Islam, Samuel Jacob Chacko, Mao Nishino, Canlin Zhang, Xiuwen Liu

Conference on Neural Information Processing Systems (NeurIPS 2026) · Main Track 2026 Accepted

We proposed Neural Expansion, a geometric account of generalization based on growing directions along which predictions remain unchanged. Growing regions of stable predictions offer a common geometric explanation for several learning behaviors; the mechanism is supported by empirical evidence and partial theory.

Neural Expansion: A Unified Mechanism for How Deep Neural Networks Generalize

Chashi Mahiul Islam, Samuel Jacob Chacko, Mao Nishino, Canlin Zhang, Xiuwen Liu

Conference on Neural Information Processing Systems (NeurIPS 2026) · Main Track 2026 Accepted

We proposed Neural Expansion, a geometric account of generalization based on growing directions along which predictions remain unchanged. Growing regions of stable predictions offer a common geometric explanation for several learning behaviors; the mechanism is supported by empirical evidence and partial theory.

Numerical Instability and Chaos: Quantifying the Unpredictability of Large Language Models

Chashi Mahiul Islam, Alan Villarreal, Mao Nishino, Shaeke Salman, Xiuwen Liu

IEEE World Congress on Computational Intelligence (WCCI 2026), IJCNN 2026 Oral

We discover that microscopic floating-point input perturbations cause up to 10-million-fold amplification through LLM layers, collapsing logit margins and triggering unpredictable token flips — revealing a fundamental numerical chaos regime in large language models. We quantify regimes of stability and rounding-induced chaos to characterize when LLM predictions remain stable or become unpredictable.

Numerical Instability and Chaos: Quantifying the Unpredictability of Large Language Models

Chashi Mahiul Islam, Alan Villarreal, Mao Nishino, Shaeke Salman, Xiuwen Liu

IEEE World Congress on Computational Intelligence (WCCI 2026), IJCNN 2026 Oral

We discover that microscopic floating-point input perturbations cause up to 10-million-fold amplification through LLM layers, collapsing logit margins and triggering unpredictable token flips — revealing a fundamental numerical chaos regime in large language models. We quantify regimes of stability and rounding-induced chaos to characterize when LLM predictions remain stable or become unpredictable.

Adversarial Attacks on Large Language Models Using Regularized Relaxation

Samuel Jacob Chacko*, Sajib Biswas*, Chashi Mahiul Islam, Fatema Tabassum Liza, Xiuwen Liu (* equal contribution)

Information Sciences 2026

We co-developed Regularized Relaxation, which continuously optimizes adversarial suffix embeddings with regularization and maps them into valid vocabulary tokens. The method improves attack success across five LLMs and runs roughly 100 times faster than the compared greedy coordinate gradient (GCG) search.

Adversarial Attacks on Large Language Models Using Regularized Relaxation

Samuel Jacob Chacko*, Sajib Biswas*, Chashi Mahiul Islam, Fatema Tabassum Liza, Xiuwen Liu (* equal contribution)

Information Sciences 2026

We co-developed Regularized Relaxation, which continuously optimizes adversarial suffix embeddings with regularization and maps them into valid vocabulary tokens. The method improves attack success across five LLMs and runs roughly 100 times faster than the compared greedy coordinate gradient (GCG) search.

When Skills Don’t Help: A Negative Result on Procedural Knowledge for Tool-Grounded Agents in Offensive Cybersecurity

Samuel Jacob Chacko*, James Hugglestone*, Chashi Mahiul Islam, Xiuwen Liu (* equal contribution)

ACM CAIS 2026 AgentSkills Workshop 2026 Poster

We evaluated instruction strategies for MCP-based, tool-grounded cybersecurity agents in 180 runs across four guidance levels. Increasing context from 591 to 36,001 tokens produced no significant success-rate gain. We reinterpreted the study as a four-level Skills comparison and proposed environment feedback as an explanation for limited gains: the value of additional instructions may depend on how much guidance the tool environment already supplies.

When Skills Don’t Help: A Negative Result on Procedural Knowledge for Tool-Grounded Agents in Offensive Cybersecurity

Samuel Jacob Chacko*, James Hugglestone*, Chashi Mahiul Islam, Xiuwen Liu (* equal contribution)

ACM CAIS 2026 AgentSkills Workshop 2026 Poster

We evaluated instruction strategies for MCP-based, tool-grounded cybersecurity agents in 180 runs across four guidance levels. Increasing context from 591 to 36,001 tokens produced no significant success-rate gain. We reinterpreted the study as a four-level Skills comparison and proposed environment feedback as an explanation for limited gains: the value of additional instructions may depend on how much guidance the tool environment already supplies.

ActivationGuard: Efficient Detection of Indirect Prompt Injection in Network-Integrated LLM Agents

Chashi Mahiul Islam, Samuel Jacob Chacko, Xiuwen Liu

Submitted to ICNC 2027 2026 Submitted

We developed a lightweight prompt-injection detector using logistic regression on frozen LLM hidden states. Expanded training coverage raised attack recall from 17.2% to 93.8% under a 5% false-positive-rate validation budget, achieving 0.979 AUROC without enlarging the inference architecture. Content detection alone does not establish end-to-end agent safety.

ActivationGuard: Efficient Detection of Indirect Prompt Injection in Network-Integrated LLM Agents

Chashi Mahiul Islam, Samuel Jacob Chacko, Xiuwen Liu

Submitted to ICNC 2027 2026 Submitted

We developed a lightweight prompt-injection detector using logistic regression on frozen LLM hidden states. Expanded training coverage raised attack recall from 17.2% to 93.8% under a 5% false-positive-rate validation budget, achieving 0.979 AUROC without enlarging the inference architecture. Content detection alone does not establish end-to-end agent safety.

On the Representation Vulnerability of Vision Transformers

Montasir Shams, Shaeke Salman, Chashi Mahiul Islam, Mao Nishino, Xiuwen Liu

Manuscript ready 2026 Manuscript ready

We combined mathematical analysis and experiments to characterize local representation structure and its consequences across downstream tasks. Similar embeddings do not necessarily imply similar meaning, which affects every application that relies on them.

On the Representation Vulnerability of Vision Transformers

Montasir Shams, Shaeke Salman, Chashi Mahiul Islam, Mao Nishino, Xiuwen Liu

Manuscript ready 2026 Manuscript ready

We combined mathematical analysis and experiments to characterize local representation structure and its consequences across downstream tasks. Similar embeddings do not necessarily imply similar meaning, which affects every application that relies on them.

2025

Vision Embeddings and Their Role in Hallucination Vulnerabilities of Multimodal Large Language Models

Chashi Mahiul Islam*, Samuel Jacob Chacko*, Preston Horne, Xiuwen Liu (* equal contribution)

International Symposium on Visual Computing (ISVC 2025) 2025 Poster

We analyze how vision encoder embeddings propagate hallucination vulnerabilities through multimodal LLMs, identifying representation-level failure modes that cause systematic visual misperception.

Vision Embeddings and Their Role in Hallucination Vulnerabilities of Multimodal Large Language Models

Chashi Mahiul Islam*, Samuel Jacob Chacko*, Preston Horne, Xiuwen Liu (* equal contribution)

International Symposium on Visual Computing (ISVC 2025) 2025 Poster

We analyze how vision encoder embeddings propagate hallucination vulnerabilities through multimodal LLMs, identifying representation-level failure modes that cause systematic visual misperception.

Spatial-ViLT: Enhancing Visual Spatial Reasoning Through Multi-Task Learning

Chashi Mahiul Islam*, Oteo Mamo*, Samuel Jacob Chacko, Xiuwen Liu, Weikuan Yu (* equal contribution)

International Symposium on Visual Computing (ISVC 2025) · Lecture Notes in Computer Science, Springer. 2025 Oral

SpatialViLT integrates depth maps, 3D coordinates, and edge representations into a vision-language model via multi-task learning, improving 3D spatial grounding for embodied intelligence and achieving a 2.1% gain on the Visual Spatial Reasoning (VSR) benchmark.

Spatial-ViLT: Enhancing Visual Spatial Reasoning Through Multi-Task Learning

Chashi Mahiul Islam*, Oteo Mamo*, Samuel Jacob Chacko, Xiuwen Liu, Weikuan Yu (* equal contribution)

International Symposium on Visual Computing (ISVC 2025) · Lecture Notes in Computer Science, Springer. 2025 Oral

SpatialViLT integrates depth maps, 3D coordinates, and edge representations into a vision-language model via multi-task learning, improving 3D spatial grounding for embodied intelligence and achieving a 2.1% gain on the Visual Spatial Reasoning (VSR) benchmark.

NeuroShield-ViT: Mechanistic Understandings of Representation Vulnerabilities and Engineering Robust Vision Transformers

Chashi Mahiul Islam, Samuel Jacob Chacko, Mao Nishino, Xiuwen Liu

International Symposium on Visual Computing (ISVC 2025) 2025 Oral

We examine how small input changes affect layer-wise representations in Vision Transformers (ViTs) and propose NeuroShield-ViT, which neutralizes vulnerable neurons without retraining. PGD-100 accuracy improves from 15.6% to 72.3%; classification accuracy also improves by 71.6% against IGO attacks on ImageNet1K.

NeuroShield-ViT: Mechanistic Understandings of Representation Vulnerabilities and Engineering Robust Vision Transformers

Chashi Mahiul Islam, Samuel Jacob Chacko, Mao Nishino, Xiuwen Liu

International Symposium on Visual Computing (ISVC 2025) 2025 Oral

We examine how small input changes affect layer-wise representations in Vision Transformers (ViTs) and propose NeuroShield-ViT, which neutralizes vulnerable neurons without retraining. PGD-100 accuracy improves from 15.6% to 72.3%; classification accuracy also improves by 71.6% against IGO attacks on ImageNet1K.

Statistical Time Series Modeling of Quantum Systems Using Rough Transformers

Yuanhang Shao, Chashi Mahiul Islam, Samuel Jacob Chacko, Ryan Carmichael, Jake Scally, Xiuwen Liu, Bayaner Arigong

IEEE International Conference on Quantum Computing and Engineering (QCE 2025) 2025

We apply rough transformer architectures to model statistical time series of quantum systems, demonstrating improved forecasting of quantum state evolution over classical baselines.

Statistical Time Series Modeling of Quantum Systems Using Rough Transformers

Yuanhang Shao, Chashi Mahiul Islam, Samuel Jacob Chacko, Ryan Carmichael, Jake Scally, Xiuwen Liu, Bayaner Arigong

IEEE International Conference on Quantum Computing and Engineering (QCE 2025) 2025

We apply rough transformer architectures to model statistical time series of quantum systems, demonstrating improved forecasting of quantum state evolution over classical baselines.

Are Vision Transformer Representations Semantically Meaningful? A Case Study in Medical Imaging

Montasir Shams, Chashi Mahiul Islam, Shaeke Salman, Phat Tran, Xiuwen Liu

arXiv preprint 2025

We investigate whether Vision Transformer intermediate representations encode semantically meaningful features in the medical imaging domain, finding systematic gaps between representation geometry and clinical semantics.

Are Vision Transformer Representations Semantically Meaningful? A Case Study in Medical Imaging

Montasir Shams, Chashi Mahiul Islam, Shaeke Salman, Phat Tran, Xiuwen Liu

arXiv preprint 2025

We investigate whether Vision Transformer intermediate representations encode semantically meaningful features in the medical imaging domain, finding systematic gaps between representation geometry and clinical semantics.

DeepSeek on a Trip: Inducing Targeted Visual Hallucinations via Representation Vulnerabilities

Chashi Mahiul Islam, S. Jacob Chacko, P. Horne, Xiuwen Liu

arXiv preprint 2025

We demonstrate that DeepSeek and other multimodal LLMs can be induced to generate targeted hallucinations by exploiting representation vulnerabilities in the vision encoder, revealing a critical security surface in deployed vision-language systems.

DeepSeek on a Trip: Inducing Targeted Visual Hallucinations via Representation Vulnerabilities

Chashi Mahiul Islam, S. Jacob Chacko, P. Horne, Xiuwen Liu

arXiv preprint 2025

We demonstrate that DeepSeek and other multimodal LLMs can be induced to generate targeted hallucinations by exploiting representation vulnerabilities in the vision encoder, revealing a critical security surface in deployed vision-language systems.

2024

Malicious Path Manipulations via Exploitation of Representation Vulnerabilities of Vision-Language Navigation Systems

Chashi Mahiul Islam*, Shaeke Salman*, Montasir Shams, Xiuwen Liu, Piyush Kumar (* equal contribution)

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024) 2024 Oral

We develop adversarial algorithms that exploit semantic representation gaps in vision-language navigation models to redirect robotic agents along malicious paths, achieving 91.7% landmark-conditioned path redirection and proposing a noise-sensitivity-based detection mechanism that improves navigation policy stability by 96.3%.

Malicious Path Manipulations via Exploitation of Representation Vulnerabilities of Vision-Language Navigation Systems

Chashi Mahiul Islam*, Shaeke Salman*, Montasir Shams, Xiuwen Liu, Piyush Kumar (* equal contribution)

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024) 2024 Oral

We develop adversarial algorithms that exploit semantic representation gaps in vision-language navigation models to redirect robotic agents along malicious paths, achieving 91.7% landmark-conditioned path redirection and proposing a noise-sensitivity-based detection mechanism that improves navigation policy stability by 96.3%.

2023

Uncovering the Interplay of Demographics and Healthcare Provider Availability on CMS HCC Risk Scores for Disabled Beneficiaries

Mohammad Ishtiaque Rahman, Forhan Bin Emdad, Chashi Mahiul Islam, Zhe He

International Conference on e-Health and Bioengineering (EHB 2023) · Springer Nature Switzerland. 2023

We investigate how demographic factors and healthcare provider availability jointly influence CMS Hierarchical Condition Category (HCC) risk scores for disabled Medicare beneficiaries, revealing systemic disparities in risk assessment.

Uncovering the Interplay of Demographics and Healthcare Provider Availability on CMS HCC Risk Scores for Disabled Beneficiaries

Mohammad Ishtiaque Rahman, Forhan Bin Emdad, Chashi Mahiul Islam, Zhe He

International Conference on e-Health and Bioengineering (EHB 2023) · Springer Nature Switzerland. 2023

We investigate how demographic factors and healthcare provider availability jointly influence CMS Hierarchical Condition Category (HCC) risk scores for disabled Medicare beneficiaries, revealing systemic disparities in risk assessment.

2022

InnerEye: A Tale on Images Filtered Using Instagram Filters — How Do We Interact with Them and How Can We Automatically Identify the Extent of Filtering?

Gazi Abdur Rakib, Rudaiba Adnin, Shekh Ahammed Adnan Bashir, Chashi Mahiul Islam, Abir Mohammad Turza, Saad Manzur, Monowar Anjum Rashik, Abdus Salam Azad, Tusher Chakraborty, Sydur Rahaman, Muhammad Rayhan Shikder, Syed Ishtiaque Ahmed, A. B. M. Alim Al Islam

International Conference on Mobile and Ubiquitous Systems: Computing, Networking, and Services (MobiQuitous 2022) · Springer Nature Switzerland. 2022

We study how Instagram filters alter user interaction patterns and build automated classifiers to identify the extent of image filtering, providing insight into social media image perception.

InnerEye: A Tale on Images Filtered Using Instagram Filters — How Do We Interact with Them and How Can We Automatically Identify the Extent of Filtering?

Gazi Abdur Rakib, Rudaiba Adnin, Shekh Ahammed Adnan Bashir, Chashi Mahiul Islam, Abir Mohammad Turza, Saad Manzur, Monowar Anjum Rashik, Abdus Salam Azad, Tusher Chakraborty, Sydur Rahaman, Muhammad Rayhan Shikder, Syed Ishtiaque Ahmed, A. B. M. Alim Al Islam

International Conference on Mobile and Ubiquitous Systems: Computing, Networking, and Services (MobiQuitous 2022) · Springer Nature Switzerland. 2022

We study how Instagram filters alter user interaction patterns and build automated classifiers to identify the extent of image filtering, providing insight into social media image perception.