Sycophancy is a key behavioral risk in LLMs, yet is often treated as an isolated failure mode that occurs via a single causal mechanism. We instead propose modeling it as geometric and causal compositions of psychometric traits such as emotionality, openness, and agreeableness - similar to factor decomposition in psychometrics. Using Contrastive Activation Addition (CAA), we map activation directions to these factors and study how different combinations may give rise to sycophancy (e.g., high extraversion combined with low conscientiousness). This perspective allows for interpretable and compositional vector-based interventions like addition, subtraction and projection; that may be used to mitigate safety-critical behaviors in LLMs.

Large Vision Language Models (LVLMs) have demonstrated remarkable capabilities across various downstream tasks, including biometric face recognition (FR) with description. However, demographic biases remain a critical concern in FR, as these foundation models often fail to perform equitably across diverse demographic groups, considering ethnicity/race, gender, and age. Therefore, through our work DemoBias, we conduct an empirical evaluation to investigate the extent of demographic biases in LVLMs for biometric FR with textual token generation tasks. We fine-tuned and evaluated three widely used pre-trained LVLMs: LLaVA, BLIP-2, and PaliGemma on our own generated demographic-balanced dataset. We utilize several evaluation metrics, like group-specific BERTScores and the Fairness Discrepancy Rate, to quantify and trace the performance disparities. The experimental results deliver compelling insights into the fairness and reliability of LVLMs across diverse demographic groups. Our empirical study uncovered demographic biases in LVLMs, with PaliGemma and LLaVA exhibiting higher disparities for Hispanic/Latino, Caucasian, and South Asian groups, whereas BLIP-2 demonstrated comparably consistent. Repository: https://github.com/Sufianlab/DemoBias.

--We present an unsupervised slide-quality assessment pipeline that combines seven expert-inspired visual-design metrics (whitespace, colorfulness, edge density, brightness contrast, text density, color harmony, layout balance) with CLIP-ViT embeddings, using Isolation Forest-based anomaly scoring to evaluted presentation slides. Trained on 12k professional lecture slides and evaluated on six academic talks (115 slides), our method achieved Pearson correlations up to 0.83 with human visual-quality ratings--1.79 to 3.23 stronger than scores from leading vision-language models (ChatGPT o4-mini-high, Chat-GPT o3, Claude Sonnet 4, Gemini 2.5 Pro). We demonstrate convergent validity with visual ratings, discriminant validity against speaker-delivery scores, and exploratory alignment with overall impressions. Our results show that augmenting low-level design cues with multimodal embeddings closely approximates audience perceptions of slide quality, enabling scalable, objective feedback in real time. Slideware such as PowerPoint, Keynote and Google Slides has become the primary visual channel in classrooms, boardrooms and pitch competitions.

Large Language Models (LLMs) are increasingly used to generate dynamic dialogue for game NPCs. However, their integration raises new security concerns. In this study, we examine whether adversarial prompt injection can cause LLM-based NPCs to reveal hidden background secrets that are meant to remain undisclosed.

Large Language Models (LLMs) are widely deployed in applications that accept user-submitted content, such as uploaded documents or pasted text, for tasks like summarization and question answering. In this paper, we identify a new class of attacks, prompt in content injection, where adversarial instructions are embedded in seemingly benign inputs. When processed by the LLM, these hidden prompts can manipulate outputs without user awareness or system compromise, leading to biased summaries, fabricated claims, or misleading suggestions. We demonstrate the feasibility of such attacks across popular platforms, analyze their root causes including prompt concatenation and insufficient input isolation, and discuss mitigation strategies. Our findings reveal a subtle yet practical threat in real-world LLM workflows.

The performance of modern machine learning systems depends on access to large, high-quality datasets, often sourced from user-generated content or proprietary, domain-specific corpora. However, these rich datasets inherently contain sensitive personal information, raising significant concerns about privacy, data security, and compliance with regulatory frameworks. While conventional anonymization techniques can remove explicit identifiers, such removal may result in performance drop in downstream machine learning tasks. More importantly, simple anonymization may not be effective against inference attacks that exploit implicit signals such as writing style, topical focus, or demographic cues, highlighting the need for more robust privacy safeguards during model training. To address the challenging issue of balancing user privacy and data utility, we propose a reinforcement learning framework that fine-tunes a large language model (LLM) using a composite reward function that jointly optimizes for explicit and implicit privacy, semantic fidelity, and output diversity. To effectively capture population level regularities, the privacy reward combines semantic cues with structural patterns derived from a minimum spanning tree (MST) over latent representations. By modeling these privacy-sensitive signals in their distributional context, the proposed approach guides the model to generate synthetic rewrites that preserve utility while mitigating privacy risks. Empirical results show that the proposed method significantly enhances author obfuscation and privacy metrics without degrading semantic quality, providing a scalable and model-agnostic solution for privacy preserving data generation in the era of large language models.

Time series Forecasting with large languagemodels (LLMs) requires bridging numericalpatterns and natural language. Effective fore-casting on LLM often relies on extensive pre-processing and fine-tuning.Recent studiesshow that a frozen LLM can rival specializedforecasters when supplied with a carefully en-gineered natural-language prompt, but craft-ing such a prompt for each task is itself oner-ous and ad-hoc. We introduce FLAIRR-TS, atest-time prompt optimization framework thatutilizes an agentic system: a Forecaster-agentgenerates forecasts using an initial prompt,which is then refined by a refiner agent, in-formed by past outputs and retrieved analogs.This adaptive prompting generalizes across do-mains using creative prompt templates andgenerates high-quality forecasts without inter-mediate code generation.Experiments onbenchmark datasets show improved accuracyover static prompting and retrieval-augmentedbaselines, approaching the performance ofspecialized prompts.FLAIRR-TS providesa practical alternative to tuning, achievingstrong performance via its agentic approach toadaptive prompt refinement and retrieval.

Recent advances in Chain-of-Thought (CoT) prompting have substantially enhanced the reasoning capabilities of large language models (LLMs), enabling sophisticated problem-solving through explicit multi-step reasoning traces. However, these enhanced reasoning processes introduce novel attack surfaces, particularly vulnerabilities to computational inefficiency through unnecessarily verbose reasoning chains that consume excessive resources without corresponding performance gains. Prior overthinking attacks typically require restrictive conditions including access to external knowledge sources for data poisoning, reliance on retrievable poisoned content, and structurally obvious templates that limit practical applicability in real-world scenarios. To address these limitations, we propose POT (Prompt-Only OverThinking), a novel black-box attack framework that employs LLM-based iterative optimization to generate covert and semantically natural adversarial prompts, eliminating dependence on external data access and model retrieval. Extensive experiments across diverse model architectures and datasets demonstrate that POT achieves superior performance compared to other methods.

In countries without civil registration and vital statistics, verbal autopsy (VA) is a critical tool for estimating cause of death (COD) and inform policy priorities. In VA, interviewers ask proximal informants for details on the circumstances preceding a death, in the form of unstructured narratives and structured questions. Existing automated VA cause classification algorithms only use the questions and ignore the information in the narratives. In this thesis, we investigate how the VA narrative can be used for automated COD classification using pretrained language models (PLMs) and machine learning (ML) techniques. Using empirical data from South Africa, we demonstrate that with the narrative alone, transformer-based PLMs with task-specific fine-tuning outperform leading question-only algorithms at both the individual and population levels, particularly in identifying non-communicable diseases. We explore various multimodal fusion strategies combining narratives and questions in unified frameworks. Multimodal approaches further improve performance in COD classification, confirming that each modality has unique contributions and may capture valuable information that is not present in the other modality. We also characterize physician-perceived information sufficiency in VA. We describe variations in sufficiency levels by age and COD and demonstrate that classification accuracy is affected by sufficiency for both physicians and models. Overall, this thesis advances the growing body of knowledge at the intersection of natural language processing, epidemiology, and global health. It demonstrates the value of narrative in enhancing COD classification. Our findings underscore the need for more high-quality data from more diverse settings to use in training and fine-tuning PLM/ML methods, and offer valuable insights to guide the rethinking and redesign of the VA instrument and interview.

Retrieval Augmented Generation (RAG) is an important aspect of conversing with Large Language Models (LLMs) when factually correct information is important. LLMs may provide answers that appear correct, but could contain hallucinated information. Thus, building benchmarks that can evaluate LLMs on multi-turn RAG conversations has become an increasingly important task. Simulating real-world conversations is vital for producing high quality evaluation benchmarks. We present RAGAPHENE, a chat-based annotation platform that enables annotators to simulate real-world conversations for benchmarking and evaluating LLMs. RAGAPHENE has been successfully used by approximately 40 annotators to build thousands of real-world conversations.