Language models (LMs) are said to be exhibiting reasoning, but what does this entail? We assess definitions of reasoning and how key papers in the field of natural language processing (NLP) use the notion and argue that the definitions provided are not consistent with how LMs are trained, process information, and generate new tokens. To illustrate this incommensurability we assume the view that transformer-based LMs implement an \textit{implicit} finite-order Markov kernel mapping contexts to conditional token distributions. In this view, reasoning-like outputs correspond to statistical regularities and approximate statistical invariances in the learned kernel rather than the implementation of explicit logical mechanisms. This view is illustrative of the claim that LMs are "statistical pattern matchers"" and not genuine reasoners and provides a perspective that clarifies why reasoning-like outputs arise in LMs without any guarantees of logical consistency. This distinction is fundamental to how epistemic uncertainty is evaluated in LMs. We invite a discussion on the importance of how the computational processes of the systems we build and analyze in NLP research are described.

Large language models are increasingly being used in critical domains of politics, business, and education, but the nature of their normative ethical judgment remains opaque. Alignment research has, to date, not sufficiently utilized perspectives and insights from the field of moral psychology to inform training and evaluation of frontier models. I perform a synthetic experiment on a wide range of models from most major model providers using Jonathan Haidt's influential moral foundations theory (MFT) to elicit diverse value judgments from LLMs. Using multiple descriptive statistical approaches, I document the bias and variance of large language model responses relative to a human baseline in the original survey. My results suggest that models rely on different moral foundations from one another and from a nationally representative human baseline, and these differences increase as model capabilities increase. This work seeks to spur further analysis of LLMs using MFT, including finetuning of open-source models, and greater deliberation by policymakers on the importance of moral foundations for LLM alignment.

The optimal assignment of Large Language Models (LLMs) to specialized roles in multi-agent systems is a significant challenge, defined by a vast combinatorial search space, expensive black-box evaluations, and an inherent trade-off between performance and cost. Current optimization methods focus on single-agent settings and lack a principled framework for this multi-agent, multi-objective problem. This thesis introduces MALBO (Multi-Agent LLM Bayesian Optimization), a systematic framework designed to automate the efficient composition of LLM-based agent teams. We formalize the assignment challenge as a multi-objective optimization problem, aiming to identify the Pareto front of configurations between task accuracy and inference cost. The methodology employs multi-objective Bayesian Optimization (MOBO) with independent Gaussian Process surrogate models. By searching over a continuous feature-space representation of the LLMs, this approach performs a sample-efficient exploration guided by the expected hypervolume improvement. The primary contribution is a principled and automated methodology that yields a Pareto front of optimal team configurations. Our results demonstrate that the Bayesian optimization phase, compared to an initial random search, maintained a comparable average performance while reducing the average configuration cost by over 45%. Furthermore, MALBO identified specialized, heterogeneous teams that achieve cost reductions of up to 65.8% compared to homogeneous baselines, all while maintaining maximum performance. The framework thus provides a data-driven tool for deploying cost-effective and highly specialized multi-agent AI systems.

Recent advances in text-to-image generation have produced strong single-shot models, yet no individual system reliably executes the long, compositional prompts typical of creative workflows. We introduce Image-POSER, a reflective reinforcement learning framework that (i) orchestrates a diverse registry of pretrained text-to-image and image-to-image experts, (ii) handles long-form prompts end-to-end through dynamic task decomposition, and (iii) supervises alignment at each step via structured feedback from a vision-language model critic. By casting image synthesis and editing as a Markov Decision Process, we learn non-trivial expert pipelines that adaptively combine strengths across models. Experiments show that Image-POSER outperforms baselines, including frontier models, across industry-standard and custom benchmarks in alignment, fidelity, and aesthetics, and is consistently preferred in human evaluations. These results highlight that reinforcement learning can endow AI systems with the capacity to autonomously decompose, reorder, and combine visual models, moving towards general-purpose visual assistants.

Benchmarks such as ARC, Raven-inspired tests, and the Blackbird Task are widely used to evaluate the intelligence of large language models (LLMs). Yet, the concept of intelligence remains elusive- lacking a stable definition and failing to predict performance on practical tasks such as question answering, summarization, or coding. Optimizing for such benchmarks risks misaligning evaluation with real-world utility. Our perspective is that evaluation should be grounded in generality rather than abstract notions of intelligence. We identify three assumptions that often underpin intelligence-focused evaluation: generality, stability, and realism. Through conceptual and formal analysis, we show that only generality withstands conceptual and empirical scrutiny. Intelligence is not what enables generality; generality is best understood as a multitask learning problem that directly links evaluation to measurable performance breadth and reliability. This perspective reframes how progress in AI should be assessed and proposes generality as a more stable foundation for evaluating capability across diverse and evolving tasks.

Generating complex, logically-sound SPARQL queries for multi-hop questions remains a critical bottleneck for Knowledge Graph Question Answering, as the brittle nature of one-shot generation by Large Language Models (LLMs) hinders reliable interaction with structured data. Current methods lack the adaptive policies needed to dynamically debug queries based on real-time execution feedback. This paper introduces a novel agentic framework where an LLM learns a resilient policy for the sequential process of iterative SPARQL construction. We show that a compact 3B-parameter model, trained exclusively via outcome-driven Reinforcement Learning (GRPO) without supervised fine-tuning, can learn effective policies for this task, discovering how to systematically recover from execution errors and refine its queries toward a correct answer. On a curated, executable single-answer subset of LC-QuAD 2.0, our agent achieves 49.7\% accuracy post-entity-linking, a significant 17.5 percentage point improvement over the strongest iterative zero-shot baseline. Further analysis reveals that while the agent's capability is driven by RL, its performance is enhanced by an explicit deliberative reasoning step that acts as a cognitive scaffold to improve policy precision. This work presents a generalizable blueprint for teaching agents to master formal, symbolic tools through interaction, bridging the gap between probabilistic LLMs and the structured world of Knowledge Graphs.

Molecular property prediction is fundamental to chemical engineering applications such as solvent screening. We present Socrates-Mol, a framework that transforms language models into empirical Bayesian reasoners through context engineering, addressing cold start problems without model fine-tuning. The system implements a reflective-prediction cycle where initial outputs serve as priors, retrieved molecular cases provide evidence, and refined predictions form posteriors, extracting reusable chemical rules from sparse data. We introduce ranking tasks aligned with industrial screening priorities and employ cross-model self-consistency across five language models to reduce variance. Experiments on amine solvent LogP prediction reveal task-dependent patterns: regression achieves 72% MAE reduction and 112% R-squared improvement through self-consistency, while ranking tasks show limited gains due to systematic multi-model biases. The framework reduces deployment costs by over 70% compared to full fine-tuning, providing a scalable solution for molecular property prediction while elucidating the task-adaptive nature of self-consistency mechanisms.

We transitioned our post-CS1 course that introduces various subfields of computer science so that it integrates Large Language Models (LLMs) in a structured, critical, and practical manner. It aims to help students develop the skills needed to engage meaningfully and responsibly with AI. The course now includes explicit instruction on how LLMs work, exposure to current tools, ethical issues, and activities that encourage student reflection on personal use of LLMs as well as the larger evolving landscape of AI-assisted programming. In class, we demonstrate the use and verification of LLM outputs, guide students in the use of LLMs as an ingredient in a larger problem-solving loop, and require students to disclose and acknowledge the nature and extent of LLM assistance. Throughout the course, we discuss risks and benefits of LLMs across CS subfields. In our first iteration of the course, we collected and analyzed data from students pre and post surveys. Student understanding of how LLMs work became more technical, and their verification and use of LLMs shifted to be more discerning and collaborative. These strategies can be used in other courses to prepare students for the AI-integrated future.

We present an end-to-end demonstration of how attackers can exploit AI safety failures to harm vulnerable populations: from jailbreaking LLMs to generate phishing content, to deploying those messages against real targets, to successfully compromising elderly victims. We systematically evaluated safety guardrails across six frontier LLMs spanning four attack categories, revealing critical failures where several models exhibited near-complete susceptibility to certain attack vectors. In a human validation study with 108 senior volunteers, AI-generated phishing emails successfully compromised 11\% of participants. Our work uniquely demonstrates the complete attack pipeline targeting elderly populations, highlighting that current AI safety measures fail to protect those most vulnerable to fraud. Beyond generating phishing content, LLMs enable attackers to overcome language barriers and conduct multi-turn trust-building conversations at scale, fundamentally transforming fraud economics. While some providers report voluntary counter-abuse efforts, we argue these remain insufficient.

Artificial intelligence (AI) is used in numerous fields of science, yet the initial research questions and targets are still almost always provided by human researchers. AI-generated creative ideas in science are rare and often vague, so that it remains a human task to execute them. Automating idea generation and implementation in one coherent system would significantly shift the role of humans in the scientific process. Here we present AI-Mandel, an LLM agent that can generate and implement ideas in quantum physics. AI-Mandel formulates ideas from the literature and uses a domain-specific AI tool to turn them into concrete experiment designs that can readily be implemented in laboratories. The generated ideas by AI-Mandel are often scientifically interesting - for two of them we have already written independent scientific follow-up papers. The ideas include new variations of quantum teleportation, primitives of quantum networks in indefinite causal orders, and new concepts of geometric phases based on closed loops of quantum information transfer. AI-Mandel is a prototypical demonstration of an AI physicist that can generate and implement concrete, actionable ideas. Building such a system is not only useful to accelerate science, but it also reveals concrete open challenges on the path to human-level artificial scientists.