My research focuses on robustness and explainability of large multimodal models — understanding how microscopic input perturbations cascade through neural representations to cause failures in vision-language understanding, embodied navigation, and LLM reasoning.
Frozen LLM hidden states, logistic regression · 2026
Built a lightweight detector for indirect prompt injection in network-integrated LLM agents 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.
Co-developed regularized continuous optimization of adversarial suffix embeddings with mapping to valid vocabulary tokens. Improved attack success across five LLMs while running approximately 100 times faster than greedy coordinate gradient (GCG) search.
Paper CodebaseMCP, agent evaluation · ACM CAIS 2026 AgentSkills Workshop
Evaluated instruction strategies in 180 cybersecurity-agent runs across four guidance levels. Increasing context from 591 to 36,001 tokens produced no significant success-rate gain, suggesting that additional procedural guidance may have limited value when the tool environment already supplies useful feedback.
arXivAccepted at NeurIPS 2026
Proposed Neural Expansion, a geometric account of generalization based on Generalization Intervals, directions along which predictions remain unchanged. Stable regions grow during training; a theoretical bound connects interval size to classification margin and directional sensitivity.
Paper CodebaseInternational Joint Conference on Neural Networks (IJCNN 2026), part of IEEE WCCI 2026
Discovered that microscopic floating-point input perturbations cause up to 10-million-fold amplification through LLM layers, collapsing logit margins and triggering unpredictable token flips. Quantified regimes of stability and rounding-induced chaos to characterize when LLM predictions remain stable or become unpredictable.
arXiv CodebaseInternational Symposium on Visual Computing (ISVC 2025)
Examined how small input changes affect layer-wise representations in Vision Transformers (ViTs). Proposed NeuroShield-ViT to neutralize vulnerable neurons and enhance adversarial robustness without retraining, raising PGD-100 accuracy from 15.6% to 72.3%. Classification accuracy also improved by 71.6% against IGO attacks on ImageNet1K.
PaperIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024)
Developed algorithms that exploit semantic representation gaps in vision-language navigation (VLN) models to redirect robotic agents along malicious paths. Achieved 91.7% landmark-conditioned path redirection and proposed a noise-sensitivity-based detection mechanism improving navigation policy stability by 96.3%.
Paper CodebaseInternational Symposium on Visual Computing (ISVC 2025) · LNCS, Springer
Developed SpatialViLT, a vision-language model integrating depth maps, 3D coordinates, and edge representations via multi-task learning to improve spatial grounding for embodied intelligence. Improved the Visual Spatial Reasoning (VSR) benchmark by 2.1%.
PaperInternational Symposium on Visual Computing (ISVC 2025)
Analyzed how vision encoder embeddings propagate hallucination vulnerabilities through multimodal LLMs, identifying representation-level failure modes that cause systematic visual misperception in deployed systems such as DeepSeek.
Paper Related arXivXception, super-resolution · 2022
Evaluated state-of-the-art models for identifying low-resolution deep fakes in video data. Added super-resolution to improve test accuracy by 19% on the FaceShifter dataset using Xception Net.
Codebase