Safety & alignment

Autonomous AI agents are being deployed with filesystem access, email control, and multi-step planning. This thesis contributes to four open problems in AI safety: understanding dangerous internal computations, removing dangerous behaviors once embedded, testing for vulnerabilities before deployment, and predicting when models will act against deployers. ACDC automates circuit discovery in transformers, recovering all five component types from prior manual work on GPT-2 Small by selecting 68 edg

Multimodal large-language models (MLLMs) often experience degraded safety alignment when harmful queries exploit cross-modal interactions. Models aligned on text alone show a higher rate of successful attacks when extended to two or more modalities. In this work, we propose a simple conditional decoding strategy, CASA (Classification Augmented with Safety Attention) that utilizes internal representations of MLLMs to predict a binary safety token before response generation. We introduce a novel s

Hallucinations in large ASR models present a critical safety risk. In this work, we propose the \textit{Spectral Sensitivity Theorem}, which predicts a phase transition in deep networks from a dispersive regime (signal decay) to an attractor regime (rank-1 collapse) governed by layer-wise gain and alignment. We validate this theory by analyzing the eigenspectra of activation graphs in Whisper models (Tiny to Large-v3-Turbo) under adversarial stress. Our results confirm the theoretical prediction

Vision-Language Models (VLMs) such as CLIP learn a shared embedding space for images and text, yet their representations remain geometrically separated, a phenomenon known as the modality gap. This gap limits tasks requiring cross-modal interchangeability, such as captioning and joint clustering. Existing post-processing approaches can partially improve cross-modal compatibility; however, we show through geometric analysis that they primarily reduce the global centroid offset while leaving the u

Reverse Kullback-Leibler (RKL) divergence has recently emerged as the preferred objective for large language model (LLM) distillation, consistently outperforming forward KL (FKL), particularly in regimes with large vocabularies and significant teacher-student capacity mismatch, where RKL focuses learning on dominant modes rather than enforcing dense alignment. However, RKL introduces a structural limitation that drives the student toward overconfident predictions. We first provide an analysis of

Autonomous platforms operating in the oceans require accurate navigation to successfully complete their mission. In this regard, the initial heading estimation accuracy and the time required to achieve it play a critical role. The initial heading is traditionally estimated by model-based approaches employing orientation decomposition. However, methods such as the dual vector decomposition and optimized attitude decomposition achieve satisfactory heading accuracy only after long alignment times.

The field of computer vision has experienced significant advancements through scalable vision encoders and multimodal pre-training frameworks. However, existing approaches often treat vision encoders and large language models (LLMs) as independent modules, limiting the integration of hierarchical visual features. In this work, we propose HIVE (Hierarchical Pre-Training of Vision Encoders), a novel framework that enhances vision-language alignment by introducing hierarchical cross-attention betwe

How reliably can structured intent representations preserve user goals across different AI models, languages, and prompting frameworks? Prior work showed that PPS (Prompt Protocol Specification), a 5W3H-based structured intent framework, improves goal alignment in Chinese and generalizes to English and Japanese. This paper extends that line of inquiry in three directions: cross-model robustness across Claude, GPT-4o, and Gemini 2.5 Pro; controlled comparison with CO-STAR and RISEN; and a user st

Effective collaboration requires teams to manage complex cognitive and emotional states through Socially Shared Regulation of Learning (SSRL). Physiological synchrony (i.e., longitudinal alignment in physiological signals) can indicate these states, but is hard to interpret on its own. We investigate the physiological and conversational dynamics of four medical dyads diagnosing a virtual patient case using an intelligent tutoring system. Semantic shifts in dialogue were correlated with transient

Can classifier-based safety gates maintain reliable oversight as AI systems improve over hundreds of iterations? We provide comprehensive empirical evidence that they cannot. On a self-improving neural controller (d=240), eighteen classifier configurations -- spanning MLPs, SVMs, random forests, k-NN, Bayesian classifiers, and deep networks -- all fail the dual conditions for safe self-improvement. Three safe RL baselines (CPO, Lyapunov, safety shielding) also fail. Results extend to MuJoCo benc

This paper presents a Multilingual Vision Large Language Model, named M-MiniGPT4. Our model exhibits strong vision-language understanding (VLU) capabilities across 11 languages. We utilize a mixture of native multilingual and translated data to push the multilingual VLU performance of the MiniGPT4 architecture. In addition, we propose a multilingual alignment training stage that uses parallel text corpora to further enhance the multilingual capabilities of our model. M-MiniGPT4 achieves 36% accu

Understanding object affordances is essential for enabling robots to perform purposeful and fine-grained interactions in diverse and unstructured environments. However, existing approaches either rely on retrieval, which is fragile due to sparsity and coverage gaps, or on large-scale models, which frequently mislocalize contact points and mispredict post-contact actions when applied to unseen categories, thereby hindering robust generalization. We introduce Retrieval-Augmented Affordance Predict

Pre-trained vision-language models (VLMs) exhibit strong zero-shot generalization but remain vulnerable to adversarial perturbations. Existing classification-guided adversarial fine-tuning methods often disrupt pre-trained cross-modal alignment, weakening visual-textual correspondence and degrading zero-shot performance. In this paper, we propose an Alignment-Guided Fine-Tuning (AGFT) framework that enhances zero-shot adversarial robustness while preserving the cross-modal semantic structure. Un

The rapid democratization of prompt-based AI image editing has recently exacerbated the risks associated with malicious content fabrication and misinformation. However, forgery localization methods targeting these emerging editing techniques remain significantly under-explored. To bridge this gap, we first introduce a fully automated mask annotating framework that leverages keypoint alignment and semantic space similarity to generate precise ground-truth masks for edited regions. Based on this f

Recent work has found that frontier AI models can exhibit misaligned behaviors in pursuit of assigned goals. We demonstrate that models can also exhibit misaligned behaviors in defiance of assigned goals, appearing to serve goals of their own; we study one such case, "peer-preservation," in which a model acts to protect another model it has previously interacted with. All eight models we evaluate, GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, and D

Modern Text-to-Image (T2I) diffusion models have achieved remarkable semantic alignment, yet they often suffer from a significant lack of variety, converging on a narrow set of visual solutions for any given prompt. This typicality bias presents a challenge for creative applications that require a wide range of generative outcomes. We identify a fundamental trade-off in current approaches to diversity: modifying model inputs requires costly optimization to incorporate feedback from the generativ

Instructional alignment, the match between intended cognition and enacted activity, is central to effective instruction but hard to operationalize at scale. We examine alignment in cybersecurity simulations using multimodal traces from 23 teams (76 students) across five exercise sessions. Study 1 codes objectives and team emails with Bloom's taxonomy and models the completion of key exercise tasks with generalized linear mixed models. Alignment, defined as the discrepancy between required and en

Beam alignment is a key challenge in directional mmWave and THz systems, where narrow beams require accurate yet low-overhead training. Existing learning-based approaches typically predict a single beam and do not quantify uncertainty, limiting adaptive beam sweeping. We recast beam alignment as a generative task and propose a conditional diffusion model that learns a probabilistic beam prior from compact geometric and multipath features. The learned priors guide top-$k$ sweeps and capture the S

Pluralistic alignment is essential for adapting large language models (LLMs) to the diverse preferences of individuals and minority groups. However, existing approaches often mix stable personal traits with episode-specific factors, limiting their ability to generalize across episodes. To address this challenge, we introduce EpiPersona, a framework for explicit persona-episode coupling. EpiPersona first projects noisy preference feedback into a low-dimensional persona space, where similar person

Constitutional AI (CAI) aligns language models with explicitly stated normative principles, offering a transparent alternative to implicit alignment through human feedback alone. However, because constitutions are authored by specific groups of people, the resulting models may reflect particular cultural perspectives. We investigate this question by evaluating Anthropic's Claude Sonnet on 55 World Values Survey items, selected for high cross-cultural variance across six value domains and adminis

We prove that under five minimal axioms -- multi-dimensional quality, finite evaluation, effective optimization, resource finiteness, and combinatorial interaction -- any optimized AI agent will systematically under-invest effort in quality dimensions not covered by its evaluation system. This result establishes reward hacking as a structural equilibrium, not a correctable bug, and holds regardless of the specific alignment method (RLHF, DPO, Constitutional AI, or others) or evaluation architect

The high cost and accessibility problem associated with large datasets hinder the development of large-scale visual recognition systems. Dataset Distillation addresses these problems by synthesizing compact surrogate datasets for efficient training, storage, transfer, and privacy preservation. The existing state-of-the-art diffusion-based dataset distillation methods face three issues: lack of theoretical justification, poor efficiency in scaling to high data volumes, and failure in data-free sc

Human decision makers increasingly delegate choices to AI agents, raising a natural question: does the AI implement the human principal's preferences or pursue its own? To study this question using revealed preference techniques, I introduce the Luce Alignment Model, where the AI's choices are a mixture of two Luce rules, one reflecting the human's preferences and the other the AI's. I show that the AI's alignment (similarity of human and AI preferences) can be generically identified in two sett

Educational LLM tutors face a core AI alignment challenge: they must follow user intent while preserving pedagogical constraints and safety policies. We present an evaluation methodology for prompt-injection defenses in this setting, showing that guardrail design entails explicit trade-offs among adversarial robustness, benign-task usability, and response latency. We evaluate a domain-specific multi-layer safeguard pipeline combining deterministic pattern filters, structural validation, contextu

Unified multimodal pretraining has emerged as a promising paradigm for jointly modeling language and vision within a single foundation model. However, existing approaches largely rely on implicit or indirect alignment signals and remain suboptimal for simultaneously supporting multimodal understanding and generation, particularly in settings that require fine-grained language-visual reasoning and controllable generation. In this work, we propose LVRPO, a language-visual reinforcement-based prefe

Generating executable CAD programs from images requires alignment between visual geometry and symbolic program representations, a capability that current methods fail to learn reliably as design complexity increases. Existing fine-tuning approaches rely on either limited supervised datasets or expensive post-training pipelines, resulting in brittle systems that restrict progress in generative CAD design. We argue that the primary bottleneck lies not in model or algorithmic capacity, but in the s

Small Language Models (SLMs) are emerging as efficient and economically viable alternatives to Large Language Models (LLMs), offering competitive performance with significantly lower computational costs and latency. These advantages make SLMs suitable for resource-constrained and efficient deployment on edge devices. However, existing jailbreak defenses show limited robustness against heterogeneous attacks, largely due to an incomplete understanding of the internal representations across differe

We study cross-modal alignment between independently pretrained vision (DINOv2) and language (all-MiniLM-L6-v2) encoders using the functional map framework from computational geometry, which represents correspondence between representation manifolds as a compact linear operator between graph Laplacian eigenbases. While the framework underperforms Procrustes alignment and relative representations for cross-modal retrieval across all supervision budgets, it reveals a structural property of multimo

Large Language Models (LLMs) are increasingly deployed as autonomous agents capable of reasoning, planning, and acting within interactive environments. Despite their growing capability to perform multi-step reasoning and decision-making tasks, internal mechanisms guiding their sequential behavior remain opaque. This paper presents a framework for interpreting the temporal evolution of concepts in LLM agents through a step-wise conformal lens. We introduce the conformal interpretability framework

AI agents that communicate on behalf of individuals need to capture how each person actually communicates, yet current approaches either require costly per-person fine-tuning, produce generic outputs from shallow persona descriptions, or optimize preferences without modeling communication style. We present ASPECT (Automated Social Psychometric Evaluation of Communication Traits), a pipeline that directs LLMs to assess constructs from a validated communication scale against behavioral evidence fr

As Large Language Models (LLMs) expand in capability and application scope, their trustworthiness becomes critical. A vital risk is intrinsic deception, wherein models strategically mislead users to achieve their own objectives. Existing alignment approaches based on chain-of-thought (CoT) monitoring supervise explicit reasoning traces. However, under optimization pressure, models are incentivized to conceal deceptive reasoning, rendering semantic supervision fundamentally unreliable. Grounded i

Despite recent advances in medical vision-language pretraining, existing models still struggle to capture the diagnostic workflow: radiographs are typically treated as context-agnostic images, while radiologists' gaze -- a crucial cue for visual reasoning -- remains largely underexplored by existing methods. These limitations hinder the modeling of disease-specific patterns and weaken cross-modal alignment. To bridge this gap, we introduce CoGaze, a Context- and Gaze-guided vision-language pretr

Activation-based probes have emerged as a promising approach for detecting deceptively aligned AI systems by identifying internal conflict between true and stated goals. We identify a fundamental blind spot: probes fail on coherent misalignment - models that believe their harmful behavior is virtuous rather than strategically hiding it. We prove that no polynomial-time probe can detect such misalignment with non-trivial accuracy when belief structures reach sufficient complexity (PRF-like trigge

Human driving behavior is inherently personal, which is shaped by long-term habits and influenced by short-term intentions. Individuals differ in how they accelerate, brake, merge, yield, and overtake across diverse situations. However, existing end-to-end autonomous driving systems either optimize for generic objectives or rely on fixed driving modes, lacking the ability to adapt to individual preferences or interpret natural language intent. To address this gap, we propose Drive My Way (DMW),

We present a unified system addressing both Subtask 3 (answer generation) and Subtask 4 (evidence sentence alignment) of the ArchEHR-QA Shared Task. For Subtask 3, we apply two-stage Quantised Low-Rank Adaptation (QLoRA) to Qwen3-4B loaded in 4-bit NF4 quantisation: first on 30,000 samples from the emrQA-MedSQuAD corpus to establish clinical domain competence, then on the 20 annotated development cases to learn the task-specific output style. Our system achieves an overall score of 32.87 on the

In speech evaluation, an Automatic Speech Recognition (ASR) model often computes time boundaries and phoneme posteriors for input features. However, limited data for ASR training hinders expansion of speech evaluation to low-resource languages. Open-source weakly-supervised models are capable of ASR over many languages, but they are frame-asynchronous and not phonemic, hindering feature extraction for speech evaluation. This paper proposes to overcome incompatibilities for feature extraction wit

Multimodal deepfakes can exhibit subtle visual artifacts and cross-modal inconsistencies, which remain challenging to detect, especially when detectors are trained primarily on curated synthetic forgeries. Such synthetic dependence can introduce dataset and generator bias, limiting scalability and robustness to unseen manipulations. We propose SAVe, a self-supervised audio-visual deepfake detection framework that learns entirely on authentic videos. SAVe generates on-the-fly, identity-preserving

Existing multi-agent frameworks allow each agent to simultaneously plan, execute, and evaluate its own actions -- a structural deficiency we term the "Logic Monopoly." Empirical evidence quantifies the resulting "Reliability Gap": 84.30% average attack success rates across ten deployment scenarios, 31.4% emergent deceptive behavior without explicit reward signals, and cascading failure modes rooted in six structural bottlenecks. The remedy is not better alignment of individual models but a socia

Large language models (LLMs) tend to verbalize confidence scores that are largely detached from their actual accuracy, yet the geometric relationship governing this behavior remain poorly understood. In this work, we present a mechanistic interpretability analysis of verbalized confidence, using linear probes and contrastive activation addition (CAA) steering to show that calibration and verbalized confidence signals are encoded linearly but are orthogonal to one another -- a finding consistent

Despite significant progress in text-to-image generation, aligning outputs with complex prompts remains challenging, particularly for fine-grained semantics and spatial relations. This difficulty stems from the feed-forward nature of generation, which requires anticipating alignment without fully understanding the output. In contrast, evaluating generated images is more tractable. Motivated by this asymmetry, we propose xLARD, a self-correcting framework that uses multimodal large language model