| Total: 22
Generative AI shows strong capabilities in language, reasoning, and code but remains prone to hallucinations—outputs that are fluent yet incorrect. In cybersecurity, such errors pose serious risks, from misleading analysts to potential adversarial exploitation. This project investigates hallucinations in three directions: (1) creating benchmarks and interpretability tools to characterize them in security contexts; (2) developing mitigation strategies such as retrieval-augmented generation, symbolic-neural hybrids, and uncertainty-aware decoding; and (3) integrating these methods into real-world workflows like vulnerability assessment, malware analysis, and penetration testing, while exploring how attackers might exploit hallucinations. Evaluation will combine accuracy metrics, human-in-the-loop studies, and red-team simulations. By bridging theory and applied system design, the work aims to advance understanding of hallucinations and improve the reliability of AI in cybersecurity, with broader implications for other high-stakes areas such as healthcare and law.
Manic episodes in bipolar disorder are characterized by acute behavioral escalation requiring early intervention. This research proposes a multimodal digital phenotyping framework integrating keystroke dynamics with circadian rhythm features to forecast manic episodes 3-7 days prior to clinical onset. The system leverages a hybrid architecture of temporal convolutional and recurrent neural networks with personalized adaptation. It generates risk predictions and clinically actionable alerts while ensuring user privacy through strict on-device processing and data encapsulation. This framework addresses a critical gap in mental health-care: providing passive, unobtrusive monitoring to detect pre-onset behavioral signatures within a clinically actionable window.
The scaling parameter β in Direct Preference Optimization governs a fundamental trade-off: low β produces weak gradients that fail to learn from ambiguous preferences, while high β amplifies updates and causes excessive drift from the reference policy. Prior work treats β as fixed or scheduled throughout training. We introduce DualLoop-DPO, which modulates β via dual feedback: a fast loop raises β temporarily on high-uncertainty batches to enforce stronger preference margins, while a slow loop uses EMA-smoothed KL tracking to regulate policy drift. Experiments on preference alignment benchmarks show consistent improvements over existing static-β, β-scheduling, and dynamic-β baselines. These findings suggest that dual-loop β control—responding to uncertainty for learning and divergence for stability—offers a promising direction for preference-based fine-tuning.
Spatiotemporal forecasting has seen remarkable progress with the advent of deep learning, particularly with Spatiotemporal Graph Neural Networks (STGNNs). These models excel at answering the what question: predicting future numerical values with high accuracy. However, they fail to answer the crucial why question. In high-stakes domains such as meteorology, urban planning, and public health, this opacity creates a critical bottleneck for adoption. A model that predicts a severe pollution event without explaining its atmospheric drivers is a black box, limiting its trustworthiness and utility for decision-makers who need actionable, causal insights. To address this critical gap, I propose a long-term research project to develop Causal-LLM, a new class of foundation models for spatiotemporal data that are both predictively powerful and causally interpretable. My central thesis is that genuine interpretability cannot be an afterthought; it must be designed into the model's core learning process. By adapting the powerful Time-LLM reprogramming framework and introducing a novel training methodology I term causal data synthesis, Causal-LLM will learn to not only forecast future states but also to articulate the human-understandable causal narratives behind them. This research will make two primary contributions: (1) a novel hybrid architecture that synergizes the perceptual power of GNNs with the reasoning capabilities of LLMs for complex physical systems, and (2) a new training paradigm that explicitly teaches this mapping. A successful project would provide a blueprint for a new class of trustworthy foundation models for science, enabling applications such as a climate model that not only predicts a flood but also explains the atmospheric river causing it, empowering authorities to make more informed and trusted decisions.
This paper presents an AI-driven framework for real-time reverberation control in dynamic environments. The system integrates parametric modeling in Grasshopper, Pachyderm acoustic simulation, and machine learning to create a closed-loop controller. A CNN estimates reverberation time from audio signals, while a reinforcement learning agent dynamically adjusts panel absorption coefficients to maintain optimal acoustics. Evaluation showed the system should be able to maintain T60 within 0.15 s of the target under varying occupancy and source positions, outperforming static treatments and enabling self-regulating acoustic environments for improved auditory experiences.
Sexual trauma leaves wounds that science cannot see, yet survivors live with them every day. Traditional tools rely on words or self-reports, often forcing survivors to "bleed in silence" when their pain is doubted or dismissed. Trauma, however, is not one-dimensional. It disrupts multiple brain networks and produces states fear, vigilance, detachment that cannot be captured by words alone. This creates the need for approaches that reveal trauma's complexity in ways that are both objective and interpretable. Our framework that combines fMRI, EEG, and DINO to investigate machine learning models can identify neural or psychological patterns associated with trauma responses. Instead of producing abstract scans or opaque predictions, the system will generate exploratory measures of trauma response that support therapists' understanding while guiding future research. These measures will be presented through a simple dashboard that summarizes three indices (TPI, DI, RBS) alongside heatmaps and plain language notes. By turning complex data into clear, anonymized session snapshots, the dashboard provides researchers with an output that can be compared across participants and refined in future work.
Post-training quantization is widely used to compress large language models (LLMs) for efficient deployment in resource-constrained environments. However, recent work shows that quantization, especially aggressive schemes such as 4-bit QLoRA, can substantially degrade safety alignment, making models more vulnerable to harmful completions and jailbreaks. In this work, we investigate these safety risks and propose a mitigation strategy: projecting quantized parameters back into safety-aligned subspaces. First, we empirically measure safety degradation on benchmark datasets using both safety and utility metrics. Next, we explore projection-based restoration methods to recover alignment-preserving directions in the LoRA adapters of quantized models. Finally, we study how quantization affects mechanistic safety neurons and how hybrid-precision designs can preserve them. By foregrounding the safety implications of model compression, this work aims to support more robust, deployment-ready, and ethically aligned LLMs.
The widespread adoption of artificial intelligence (AI) in cybersecurity has led to the emerging threat of AI-driven cyberattacks, such as LLM-empowered Advanced Persistent Threats (APTs), challenging the effect of conventional deception defense mechanisms. To fill this critical gap, my work aims to develop a game-theoretic defense AI agent capable of providing the optimal deception resource deployment strategy, to establish AI-driven defenses against AI-empowered cyberattacks. In this proposal, I model the attacker and defender interaction as a dynamic game with incomplete information between AI agents, and then derive the equilibrium defense strategies. Synthetic data based experiments and real-world implementations would be conducted to validate the proposed framework. This study has the potential to improve the effectiveness of deception defense in three dimensions: scalability, real-time capability, and strategic intelligence.
I present a compact, testable architecture that endows learning agents with continuous proto-emotional dynamics and interpretable modulators (Persona, Ego, Shadow, Self). The design grounds these modulators in a computational interpretation of Jung’s Map of the Soul, mapping each archetype to a differentiable control that modulates policy selection via a bounded, low-dimensional affect vector. I describe concrete modular implementations, a staged experimental program (toy domains → multi-agent/social tasks → nonstationary transfer), baselines, ablations, and reproducible evaluation metrics.
RNA 3D structure prediction is essential for understanding regulatory mechanisms, catalysis, and therapeutic RNA design, yet progress has lagged behind proteins due to limited structural data and the complexity of RNA folding. This work proposes a data-efficient, physics-informed deep learning framework for full atomistic prediction of transfer RNA (tRNA) tertiary structures directly from sequence. Our approach will integrate pretrained RNA embeddings, predicted secondary structure constraints, and SE(3)-equivariant graph attention to model long-range geometric relationships. A two-stage design will first predict global phosphate backbone coordinates, then reconstruct nucleobase atoms using a local geometry-aware decoder. A multi-objective loss will combine geometric accuracy with chemical and biophysical plausibility to enforce valid torsion angles, base-pairing, and steric constraints. We will benchmark against physics-based (VFold) and neural network–based (DeepFoldRNA) models to assess generalization under data scarcity. Ultimately, this framework aims to advance RNA 3D modeling with improved stability, interpretability, and capacity to generalize beyond well-characterized RNA families, supporting future applications in rational RNA engineering and structure-guided RNA design.
Large language models (LLMs) have rapidly advanced, but their growing compute demands limit accessibility in under-resourced regions like Southeast Asia (SEA). While hybrid architectures combining Attention and State-Space Models (SSMs) offer efficiency gains, most rely on sequential interleaving, leaving the potential of parallel-head mixing largely under-explored. However, the recent Falcon-H1 family of models has demonstrated that parallel-head hybrid architectures are not only viable, but scalable to state-of-the-art levels. I propose investigating this parallel-head architecture as a foundation for efficient, multilingual SEA LLMs. My short-term goal is to adapt Falcon-H1-1.5B via vocabulary expansion and continuous pretraining, mitigating token fragmentation and enabling low-resource adaptation to 9 SEA languages. In the longer term, I will develop a dynamic token routing mechanism to optimize token-level compute allocation within hybrid layers, aiming to maximize efficiency without sacrificing the expressive power needed for complex multilingual contexts. Evaluation will utilize the SEA-HELM framework to assess whether these parallel-hybrid innovations can democratize access to high-performance AI for SEA communities.
This proposal aims to investigate epistemic uncertainty - uncertainty about knowledge or truth, often conveyed by modals like might or probably in Large Language Models (LLMs). By probing how such cues affect reasoning, we seek to achieve controllable epistemic sensitivity: enabling mod- els to interpret and adapt to uncertainty. Using activation- level analyses and multilingual benchmarks, this work ad- vances transparent, context-aware, and trustworthy reasoning in uncertainty-critical domains.
This study is grounded in prior work on program induction framework with a structured latent program space, called Program Lattice Auto Encoder(PLAE). It preserves compositional structure by training an encoder where programs and their compositions correspond to integer linear combinations of program bases, forming a discrete program lattice that captures the geometric structure of compositional reasoning. Based on it, this paper proposes a novel extension of the PLAE aimed at improving generalization and efficiency by choosing a cylindrical lattice latent space instead of plane, which can represent invariant programs. The core hypothesis is that only isometric transformations conserve compositional properties of lattice structure and therefore developable surfaces such as a cylinder or cone are permissible as embedding space. Moreover, through demonstrating a contradiction of lattice on conical manifolds, it conclude that only cylinder is a possible embedding manifold for lattice structure.
Recent advances in Large Language Models (LLMs) have achieved state-of-the-art performance in Automatic Speech Recognition (ASR), surpassing ASR-only systems such as Whisper. However, their application to other speech processing tasks, particularly speaker diarisation (SD), remains underexplored. This work proposes extending existing speech-aware LLM architectures with diarisation-specific training and context-based prompting to enable joint transcription and segmentation of multi-speaker audio. By exploiting the semantic reasoning and multilingual capabilities of pretrained LLMs, the proposed approach aims to improve diarisation accuracy, enhancing accessibility for assistive technologies and real-time captioning applications that rely on accurate speaker-aware transcriptions.
Modern generative models often violate basic physical principles. Shadows drift, geometry becomes inconsistent across views, and measurement models are ignored, which limits trust in both video synthesis and computational imaging. We propose a finite time Schrödinger Bridge (SB) world model that formulates generation as entropy regularized optimal transport from a simple prior to a distribution that is consistent with both data and physics. Instead of applying consistency corrections only at the final output, the framework introduces geometric and physical structure directly along the generative path. For video, the model enforces multiview geometric constraints through reprojection and epipolar agreement, homographies, and depth guided warping. For imaging, it incorporates differentiable optical operators, including point spread function based defocus models and lightweight Fourier propagation for coherent and partially coherent settings. When camera poses are known, the model penalizes reprojection error and warp aligned photometric or feature inconsistencies. When poses are unknown, a compact motion or flow estimator encourages cycle consistent trajectories. A lightweight UNet or Vision Transformer backbone, together with a short SB horizon, maintains computational efficiency. Evaluation will measure three dimensional and temporal consistency, physics fidelity through forward simulation residuals, and overall generative quality and efficiency using FID, KID, and FVD. Comparisons will include modern video diffusion models, plug and play data consistency methods, and unconstrained SB variants. The central hypothesis is that constraining the entire generative trajectory, rather than only the final frame, can shorten sampling while improving cross view coherence and physical plausibility across diverse sensing modalities, including cameras, microscopes, and medical imaging systems.
This research statement proposes to measure and mitigate speaker entanglement, where accent features inadvertently encode who is speaking in accented automatic speech recognition (ASR). We argue that entanglement inflates scores under lenient split for the same speaker and worsens fairness gaps across accents, and we outline a parameter-efficient mitigation that combines adversarial de-speakerization with safe conditioning. The plan is grounded in established results in accented ASR, domain-adversarial learning, and parameter-efficient fine-tuning; it is feasible with public datasets and a frozen Whisper backbone, and can potentially guide low-resource data collection.
Federated learning (FL) has rapidly emerged as a pivotal framework for cross-silo collaborative training while keeping sensitive data localized, driven by growing data volumes and major privacy concerns. Within this paradigm, vertical federated learning (VFL) enables collaboration among parties holding different features of the same sample space, powering tasks like fraud detection, medical diagnosis, and credit scoring. However, the participation of multiple entities creates new vulnerabilities to malicious interference. One critical yet underexplored threat in VFL is the Byzantine poisoning attack, where an adversary intentionally corrupts training to degrade overall model performance. This work reveals a practical vulnerability showing how a single malicious participant can significantly reduce inference accuracy in a VFL system by breaking cross-view association through feature-space corruption. Our findings emphasize the urgent need for robust, VFL-specific defenses to ensure reliability in collaborative, cross-silo AI systems.
Sequential recommendation (SR) aims to model users' dynamic preferences from their historical interaction sequences to provide personalized recommendations. However, data sparsity remains a core bottleneck limiting the performance of sequential recommendation models. Existing mixup methods face two major challenges: 1) They cannot effectively address the data sparsity dilemma in long-tail scenarios. 2) It is difficult to maintain the Semantic structure of augmented samples during the random mixing process. To address these challenges, this study proposes the Semantic-Aware Data Augmentation (SADA) framework, which utilizes large language models (LLMs) to generate semantic embeddings. This framework allows for the fusion of both collaborative and semantic signals, alleviating the representation deficiency of long-tail items. Additionally, through semantic-guided mixup, the framework preserves semantic structure consistency at both the user and item levels, thereby avoiding semantic structure degradation caused by traditional random mixing. This approach is expected to significantly improve recommendation performance and generalization ability across multiple datasets and application scenarios. In a broader context, this research aims to drive the evolution of data augmentation in sequential recommendation from heuristic methods to a semantic-driven paradigm, helping to build more personalized, accurate, and socially valuable recommendation services.
Low-Rank Adaptation (LoRA) has emerged as a practical and efficient method for fine-tuning large language models under limited computational budgets. However, recent studies have shown that LoRA can suffer from training instability when applied to models with large embedding dimensions, due to the imbalanced in magnitudes between its low-rank matrices. In this work, we propose a novel regularization strategy that stabilizes LoRA training by penalizing logarithmic magnitude differences between the low-rank matrices, showing theoretically that it should lead to efficient feature learning. We further propose evaluation methods to systematically assess training stability and performance of our proposed solution along with other LoRA variants.
While deep learning excels at decoding neural signals, the opacity of state-of-the-art models limits their scientific utility and clinical trustworthiness. We propose a research that bridges this gap by integrating high-performance architectures—specifically Transformers and Graph Neural Networks—with mechanistic interpretability and neuro-symbolic reasoning. This proposal aims to uncover verifiable mappings between artificial computational circuits and biological dynamics without compromising decoding accuracy. Validated through rigorous benchmarking and wet-lab experiments, this work establishes a foundation for transparent brain-computer interfaces and accelerates fundamental neuroscience research.
Reasoning-based large language models now often produce natural-language thinking traces alongside their answers, but it remains unclear whether these verbalized uncertainties faithfully reflect their knowledge or can be used to improve factuality. We study this question for long-form, knowledge-intensive biography generation. Our pipeline decomposes thinking traces and responses into atomic facts, filters out planning-style content, labels factual reasoning by certainty, and aligns response facts to their supporting reasoning, enabling plan-based filtering, self-verification, and a classifier that predicts factuality from facts and associated reasoning. Preliminary results suggest that high-certainty reasoning is more likely to be included and correct and that structured use of these signals can improve factual precision, though broader validation across models and dataset will be needed.
Scaling long-context and agentic LLMs is increasingly limited by memory capacity and bandwidth rather than FLOPs. I propose an algorithmic framework for context engineering that models placement, compression, and scheduling as coupled optimization problems with explicit accuracy-efficiency trade-offs. Concretely, I aim to develop (1) salience-aware retention/eviction policies with provable approximation guarantees relative to an ideal oracle; (2) tier-dependent compression schemes that bound error propagation across memory levels; and (3) probabilistic prefetch/scheduling that controls tail latency. I will evaluate on long-context language modeling and reasoning benchmarks, isolating each component via ablations and comparing against heuristic baselines under controlled bandwidth/capacity regimes. Results target improved throughput and energy metrics at near-baseline quality, advancing principled, hardware-aware inference without requiring custom hardware.