FAS Theses and Dissertations

Life is full of stress-inducing activities, which have the potential to shape habitual behaviors that can increase anxiety and decrease overall well-being if not dealt with properly. While existing stress management technologies aim to mitigate these behaviors, such interventions are largely impersonal and have a high barrier to entry, offering generalized guidance and requiring users to manually trigger the intervention. This thesis presents a novel, smartwatch-based system for disrupting stress habits that automatically detects moments of physiological arousal and delivers personalized, in-the-moment reflection exercises.

A within-subjects user study examined 24 participants engaging with a prototype of this system, finding that 87.5% reported greater awareness of their stress habits after using the application. Participants also significantly preferred (p = 0.003) the automatic triggering over scheduled methods. Ultimately, this work contributes a novel integration of physiological sensing and behavior-change theory, offering insights into how wearables can support emotional regulation and habit-loop disruption within everyday routines.

Large language models hallucinate false content and produce sycophantic outputs, yet the dominant approaches to these failures, such as Reinforcement Learning from Human Feedback, Constitutional AI, and fine-tuning, share a common omission: they target symptoms without first asking what function an LLM serves. This thesis begins with that prior question. I argue that a central class of LLM outputs functions as assertions, propositions put forward as true, and should therefore be governed by assertoric norms. Operationalizing Timothy Williamson’s Knowledge Norm of Assertion for mechanistic systems, I define the Epistemic Governance Norm of Assertion (EGNA), which requires (i) that a system possess the capacity to partition epistemically available from unavailable content, and (ii) that this partition govern what is asserted. Drawing on recent mechanistic interpretability research, I show that modern LLMs already approximate the first condition. Current training objectives, however, structurally decouple epistemic registration from assertoric output, producing hallucination and sycophancy as predictable consequences. I implement a two-stage EGNA scaffold enforcing epistemic classification as a precondition of assertion and evaluate it across five frontier models. The scaffold reduces hallucination to zero or near-zero rates across all models and substantially suppresses sycophancy, while maintaining competitive response rates on legitimate queries. A control scaffold preserving the two-stage structure without the epistemic partition leaves hallucination largely unchecked, isolating the partition as the operative mechanism. The results also reveal that social inference enters epistemic classification itself, a failure mode that purely behavioral interventions would not have surfaced. Ultimately, getting the prior question right is not a philosophical exercise but a precondition for these systems to perform their function well.

LLMs have gained rapid popularity in recent years, with the ability to align outputs to human preferences and objectives. Eager end users have caused surging demand for LLM inference. This has motivated new research to improve efficiency, reduce costs, and better understand systems in inference serving, alongside decades of similar research in distributed training. Meanwhile, the wave of systems- and hardware-oriented reinforcement learning (RL) research has just begun. RL offers a unique paradigm that combines inference, preparation, and training phases. Although RL has been a popular arena for algorithmic inventions, there is still a need to better understand how RL policies and phase workloads affect underlying systems, beyond algorithms. This requires carefully designed experiments that vary key axes to test system effects.

To address this gap, this thesis conducts experiments along four axes: model family/policy, reward paradigm, task/dataset, and scaling. We build profiling extensions atop VeRL and collect metrics such as power, latency, energy, and other system-level metrics at the granularity of phases and subphases—a novel contribution in RL. We find that different RL policies induce distinct phase signatures, driven both by time spent in phases and by extra computational machinery. We quantify the costs of different reward-scoring paradigms—a learned reward model or an explicit reward function—in data movement and extra computation. We find that different tasks and datasets alter the response-length structure and rollout imbalance, contributing to latency differences. Finally, we outline the best use cases for different scaling configurations on H200 and A100 GPUs, which are strongly dependent on phase and stability requirements. We hope that our work may motivate new RL algorithms that address the less desirable attributes of phases, advance asynchronous RL methods that lower transition costs between phases, and inform data centers and cloud operators about the distinct workloads of RL. Overall, we hope that the profiling suite and metrics pipeline will inspire further energy-aware innovation in RL.

Research and development (R&D) competition generates innovation but can also produce socially wasteful duplication of effort. In our thesis, we model this dynamic through a stochastic all-pay auction, where firms make sunk-cost investments to increase their chances of obtaining a technological prize.

Equilibrium and welfare properties are then analyzed for deterministic all-pay auctions, proportional (Chinese) auctions, and a broader class of stochastic “share-of-score’’ mechanisms. The analysis derives equilibrium bidding behavior, welfare bounds, and approximate Price of Anarchy guarantees, and identifies conditions under which total social welfare asymptotically increases with greater competition and more firms.

Policy interventions are then studied from the perspective of a social planner seeking to reduce inefficient investment while preserving innovation incentives. Transfers, predictive information about firm valuations, and incentive-compatible reporting mechanisms are examined as coordination tools.

Finally, computational methods — including multi-agent reinforcement learning, numerical approximation, and differentiable mechanism design — are used to approximate equilibria and explore welfare-improving mechanisms in environments without closed-form solutions.

Taken together, these results provide the crucial theoretical and computational tools for understanding how institutional design shapes incentives and efficiency in R&D, providing hope for a more socially-optimal future allocation of scientific resources and investment.

Many networks contain multiple views of the data, which occur when you have one set of vertices with multiple sets of edges that connect them. These multiple sets of edges represent different sets of relationships between the vertices and can be a potent source of information about the underlying groupings of the nodes. Graph Neural Networks (GNNs) are often used to extract information from multi-view data, however, the mechanisms underlying the effectiveness of GNNs are not well understood. We analyzed the performance of two different frameworks for simple GNNs on multi-layer data generated using the multi-layer contextual stochastic block model, which allowed for fine-grained adjustment of graph properties. We found that the layer-wise probability of a community difference and the graph signal-to-noise ratio in each layer both affect model performance, and present intuition behind why this is the case.

APO is a two-stage reinforcement learning algorithm that is used to fine-tune large language models (LLM) on specific tasks. Stage 1 estimates the KL-regularized optimal value function V offline, and Stage~2 trains the policy online by regressing on optimal advantage targets. The log-mean-exp (LME) estimator of V* used in Stage 1 is downward-biased due to Jensen's inequality. This bias is significant, especially when the budget is small and the pass rate is low.

This thesis develops and evaluates two families of alternative estimators. Firstly, inspired by the replica trick from statistical physics, we develop two types of replica-based estimators. We prove that the single-replica plug-in estimator does not reduce bias under a fixed budget, and develop the multi-n slope estimator, which exploits the structure of the replica trick more directly. Secondly, for the binary reward, we show that LME is equivalent to plugging the maximum-likelihood pass-rate estimate into a known closed form, which causes endpoint snapping when the pass rate is low. Inspired by conjugate priors, we develop two different Beta-smooth estimators to solve this problem.

We evaluate these estimators in three experiments: (1) a scalar experiment where pass rate is fixed, (2) a contextual bandit experiment with 500 prompts of different pass rates, and (3) a full two-stage simulation of the APO pipeline. We discovered that replica-based estimators have a lower RMSE compared to LME when the pass rate is very low. However, they have a much larger advantage distortion that limits their practical value in Stage 2. The Beta-smooth estimators perform better compared to LME in all three experiments. In particular, the Jeffreys Beta-smooth estimator (p = (k+0.5)/(N+1)) achieves the best RMSE, reduces advantage distortion compared to LME, and carries over the improvement in Stage 1 into 25% lower parameter variance in the learned policy in Stage 2. We recommend it as a drop-in replacement for LME in APO's Stage 1.

Humanity is at a precipice: democracy is declining from a historical peak as backsliding occurs around the world, while artificial intelligence is emerging as a revolutionary technology that could permanently upend the socio-economic fabric underpinning modern society. This thesis presents four pieces of original research as part of a collective attempt to build upon our understanding of democratic processes, especially as democracies around the world adapt to the age of AI. In the following chapters, we (1) measure persistent non-response bias in a well-calibrated nationwide poll for the US Presidential Election, finding that the five largest states saw a >99% reduction in effective sample size, (2) extend a theoretical measure of polarization to participatory budgeting, finding that a newly proposed measure degenerates to a well-known measure ~99.4% of the time when applied to real elections, (3) adapt proportional veto core to the setting with infinite candidates (heralded by generative AI) as a measure of bridging, finding a strong pattern that some voting rules consistently perform better in bridging, and (4) demonstrate that innocuous-looking datasets that are easy to produce can poison AI models to have partisan leanings even after heavy semantic filtering, and develop attribution techniques that reduce subliminal effects by 100% when used for filtering.

High-precision robotic tasks, such as peg insertion, pose a challenge for learned world models. The critical states that determine success occur within an extremely narrow portion of each task trajectory, causing uniform sampling to underrepresent them. This thesis investigates whether targeted data selection improves the generation quality and downstream success of a 4D latent world model on the ManiSkill3 PegInsertionSide-v1 benchmark. We propose \textit{Pre-insertion Keyframe Sampling}, an algorithm that identifies and injects a single geometrically aligned pre-insertion frame into training data, and introduce a \textit{Peg-Hole Alignment Metric} to evaluate 3D generation quality beyond standard image metrics. Our results demonstrate that while the intervention yields statistically significant improvements in translational alignment, these gains do not translate to higher task success rates. A ground-truth subgoal ablation reveals a 56-percentage-point performance gap, identifying the structural fidelity of 3D generations as the primary bottleneck. This suggests that while data coverage improves spatial precision, high-precision control requires increasing the geometric resolution and structural integrity of the world model's output.

Humans have long designed representational tools---charts, diagrams, maps---to offload cognitive work and make complex information tractable. Conversational AI is becoming a powerful new information intermediary, one that increasingly determines how people access and interpret data. How AI systems process information, what they attend to, and how they treat the people querying them all shape the data understanding processes central to modern life---yet these interfaces are poorly understood. This thesis examines AI as an information intermediary across three studies, each probing a different stage in the pipeline from data to human insight.

The first study examines what goes in: how input representation shapes what AI can extract from data. Across three representative analysis tasks, the two systems describe synthetic datasets more precisely and accurately when raw data is accompanied by a scatterplot, especially as datasets grow in complexity. Comparison with two baselines---providing a blank chart and a chart with mismatched data---shows that the improved performance is due to the content of the charts. Our results are initial evidence that AI systems, like humans, can benefit from visualization.

The second study examines what happens inside: whether AI models' internal computations during chart question-answering are similar to human visual processing. These models process images through ``attention''---an internal mechanism that assigns more weight to particular regions of the input. Individual AI attention components on inputs charts correlate with human eye-tracking data. Also, a linear combination of attention components predicts human gaze on par with than prior methods, and generalizes better out-of-domain. Ablation studies indicate the attention components most associated with human fixation patterns are implicated in AI chart understanding.

The third study examines what comes out: specifically, what information reaches the user at all. Guardrail behavior---whether an AI model refuses or complies with a sensitive query---underscores that AI systems model the people querying them. Refusal rates vary systematically with declared user demographics, political affiliation, and even NFL team fandom, as models infer ideological identity even from incidental contextual details.

Together, these findings show that representational format, internal structure, and inferred user identity each shape what people can ultimately learn from AI-mediated information systems, with implications for the design of AI systems that are accurate, explainable, and equitable.

While protein language models have revolutionized our understanding of molecular structures, they generally model proteins in isolation, ignoring the dynamic cellular environments that dictate their biological function. Consequently, these models struggle to capture how a protein's role shifts depending on its cellular context, a critical factor in treating complex diseases. In this thesis, we introduce SEMPER (SystEm-scale Multimodal ProtEin Representations), a unified representation learning framework that anchors context-agnostic protein sequences in context-aware cellular networks. Leveraging a dual-encoder architecture and a contrastive objective, our approach systematically aligns the representations of context-agnostic amino acid sequences and 442 cell type-specific protein interaction networks derived from a single-cell transcriptomic atlas of the adult human brain. Using this multimodal approach, we demonstrate that SEMPER systematically outperforms state-of-the-art protein language models in predicting protein function and nominating therapeutic targets across five neurodegenerative diseases in a zero-shot manner. Finally, we show that integrating SEMPER into large language models improves the reliability of de novo target discovery, laying the groundwork for context-aware reasoning of protein function.