Computer Science, 1987-2026

Rideshare platforms like Uber and Lyft attract millions of drivers with the appeal of flexible, autonomous work, yet research consistently shows that compensation falls below minimum wage once expenses are accounted for. This raises a fundamental question about how ridehail drivers evaluate whether their pay is fair. Prior research has defined fairness in ridehail compensation through frameworks—wage floor, non-discrimination, platform take rates, and consumer perceptions—but none center drivers’ own perspectives. To address this gap, we conducted a multi-step qualitative study consisting of an exploratory analysis of online driver discussions across Reddit communities and semi-structured interviews with nine ridehail drivers across the United States. Using simulated ride offer screens, we asked drivers to walk through their decision-making process for accepting or rejecting rides and to articulate what they believed the ride should pay. Our findings reveal that drivers evaluate fairness primarily through six categories: spatial positioning, passenger characteristics, unpaid labor and time uncertainty, vehicle and operating costs, time of day, and perceived platform control. They assemble these factors into an integrated evaluation framework, applying concrete decision rules like distance thresholds, experience-based reinterpretation of platform estimates, and avoidance strategies. Critically, these factors compound, and fairness cannot be reduced to a single metric like wage floor or take rate. We further identify a critical gap between what drivers consider fair and what they accept, driven by both algorithmic pressure to maintain acceptance rates and the economic reality of market scarcity. Our work contributes a worker-centered operationalization of fairness that connects the costs and constraints documented in prior literature with the fairness judgments drivers construct in practice. Crucially, observed acceptance rates cannot be interpreted as evidence of driver satisfaction with compensation.

In our research, we identify the drawbacks of currently applied outcome-based approaches to LLM agent evaluation and propose to adopt the log-based one. Due to the development and improvement of the agent itself from a predictor into a highly sophisticated system capable of multiple-step reasoning, tool usage, and interacting with the environment, more complex metrics should be applied to such a model’s assessment. In addition, the outcome-based approaches may not provide insight regarding why the AI system succeeds/fails, which means that the problems related to the agent scaffold, the environment, or other determiners might not be identified. This paper develops an extensive framework for logs’ analysis in the context of the HAL leaderboard. The current experiments are performed under carefully designed settings, which include changes in model parameters, scaffolds, and reasoning modes and SciCode dataset for code generation tasks. Log-based approach enables the creation of the classification taxonomy of errors, covering such classes as model errors, scaffold errors, infrastructure errors, and error propagation. As demonstrated by the experiment results, the impact of the scaffold complexity on the agent’s accuracy is rather high. Higher scaffold complexity leads to the increased costs due to excessive planning, a higher number of steps needed and context accumulation, thus reducing the agent’s accuracy despite its strong model. In addition, a significant reasoning effort in combination with poorly designed scaffolds may have an adverse effect on performance. Therefore, the advantages of log-based analysis become evident.

Advanced LLM reasoning structures, specifically Graph of Thoughts (GoT), enhance Large Language Model capabilities by modeling complex problems as a directed graph. However, deploying GoT on open-weight Small Language Models (SLMs) in the 7B-14B parameter range on local GPUs introduces accuracy and speed bottlenecks. In particular, executing GoT locally results in graph nodes executing sequentially rather than in parallel, leading to unmitigated generation of redundant, erroneous reasoning branches that waste computation. Furthermore, SLMs tend towards "Diversity Collapse" during reasoning, repeatedly generating the same incorrect outputs, which corrupts the graph downstream. To resolve these limitations, this thesis proposes a novel Proactive Mitigation Framework designed to optimize GoT execution on local SLMs. Specifically, we reengineer the architecture by introducing a mid-generation intervention pipeline that uses an O(1) similarity heuristic to detect low thought diversity and halt the generation of redundant branches. After Diversity Collapse is detected, our pipeline routes the redundant thought to an LLM Judge for verification of output accuracy, followed by prompting through a set of zero-shot Mixture-of-Experts (MoE) personas. We evaluate the performance of our architecture by ablation of its components. Our results across GoT tasks (Sorting, Set Intersection, and Keyword Counting) highlight that:

• The similarity heuristic reduces execution time by up to 50% without degrading task performance.
• The LLM Judge introduces a time trade-off and demonstrates high capability to identify correct outputs, and struggles to correctly identify incorrect outputs.
• MoE persona prompting increases thought diversity compared to a generic retry prompt and doubles node recovery rates across all tasks. Furthermore, it achieves noticeably higher overall performance on certain tasks, improving accuracy in Set Intersection by up to 15%.

Ultimately, this thesis provides an in-depth study of the reasoning capabilities and limitations of SLMs with the GoT architecture and the potential of our Proactive Mitigation Framework for more efficient computation and improved task performance. All code can be found at the following Github repository: https://github.com/arab7716/slm-graph-of-thoughts

Rank-balanced trees [2] are binary search trees whose nodes carry integer ranks constrained by local parent–child rank-difference rules. The weak AVL (WAVL) tree relaxes AVL by permitting (2, 2)-nodes, which simplifies deletion rebalancing while maintaining low update costs. Following the framework of Haeupler, Sen, and Tarjan [2], we investigate the possible height of WAVL trees in terms of the current size n. The rank rule implies an upper bound of h ≤ log√2 n, but it is not a priori clear whether this worst-case behavior can actually occur. We show that it can. We first construct, via a sequence of insertions and deletions, arbitrarily large WAVL trees in which every internal node is of type (2, 2), and show that these trees still have height only Θ(log2 n). We then use these trees as building blocks in a more unbalanced construction whose right spine consists entirely of 1-children and whose left subtrees are largely composed of (2, 2)-nodes. The resulting family has height Θ(log√2 n), showing that the structural upper bound is asymptotically tight. We then propose a modification of the deletion rebalancing rules that suppresses our extremal configuration. Under this modified rule, we conjecture that the height improves to O(logb n) for some b > √2, while the total number of rebalancing steps remains O(m + d). As a secondary contribution, we also attempt a history-independent argument of the classical bound h ≤ logϕ m, by slightly modifying the history-dependent counting argument used in [2].

This work analyzed the relationship between persuasive language and its observable effect, agreement, in structured dyadic dialogue. In particular, we studied the presence and distribution of linguistic features and Cialdini-inspired persuasion techniques regarding conversational behavior. We leveraged large language models (LLMs) to annotate for persuasion principles across multiple dialogue datasets, while also examining the reliability and consistency of LLM-based evaluation. We analyzed these datasets using a combination of Ordinary Least Squares (OLS) linear regression modeling, Multivariate Analyses of Variance (MANOVAs), K-Means clustering, and descriptive comparisons to evaluate whether these features were indicative of conversational outcomes and contextual conditions, specifically in the form of pre-conversational prompting. Our findings showed that while these features provided limited predictive power for agreement outcomes, the discourse markers, structural properties, and persuasion principles varied across different prompting conditions: persuasion, compromise, and general dyadic dialogue. These results suggest that conversational intent played a significant role in shaping feature usage, while also emphasizing the importance of context in computational analyses of dialogue.

Vector-indexed systems provide the foundation for many modern machine learning features including semantic search and retrieval-augmented generation, where top-K approximate nearest neighbor (ANN) search is the standard for vector retrieval. While caching is a natural optimization, existing caching approaches are either approximate or limited to the top-1 case [7, 5, 4, 13, 12], leaving top-K exact caching unaddressed. This paper formalizes the exact caching problem for top-K ANN search and introduces two solutions we dub the Circular Inclusion Guarantee (CIG) and Half-Gap Guarantee (HGG) that certify which cached response vectors, if any, are among the true nearest-neighbors for a given query. We derive five caching algorithms from these guarantees, prove their correctness under metric distances, and extend them to cosine distance. We evaluate these algorithms across synthetically generated datasets isolating the effects of different workload characteristics such as inter-query distance, cache-to-test response set size (K/N) ratio, and cache size, as well as on SIFT-1M and ESCI. All algorithms achieve 100% accuracy under Euclidean and angular distances on synthetic data. We find hit rate is primarily decided by K/N ratio and inter-query distance in synthetic data, with hits observed in workload distributions satisfying both K/N > 1 and angular inter-query distance of < 3 degrees. Evaluation on SIFT-1M and ESCI yields zero hits, which data characterization suggests is primarily due to their inter-query distances lying outside of the ranges in which these algorithms produce hits under synthetic conditions. We also provide evidence of an inherent cache utilization ceiling for our algorithms arising from the conservativeness of the guarantees.

Distributed networks rely on cooperation between autonomous agents to achieve global outcomes, but adversarial agents can disrupt these outcomes by exploiting shared responsibility for selfish gain while hiding their malicious actions among the actions of others. This creates a detection gap, where adversarial agents are known to exist in the network yet remain indistinguishable from honest agents. Prior work by Dani et al. [7] established a worst-case framework for cooperation verification via symmetric random walks on general graphs, but was restricted to a single accusation. We generalize this framework to the multi-accusation setting on path graphs, where a strategy may accuse up to k agents simultaneously and succeeds if any accused agent is adversarial. We construct an accusation set Sk(p) with appropriately chosen p and prove that either cover time is achieved or an adversary is successfully accused in O(n^2(log n+ log δ−1) + nb^2(log n/k + 1/k log δ−1)) rounds with probability at least 1−δ. Under strict error conditions, this yields a factor-of-k speedup over the state-of-the-art single-accusation result. We further derive an optimal budget size given cost-aware applications.

This thesis is an analysis of how Julius Caesar builds his political authority, civic identity, and legitimacy through his Commentarii, specifically the De Bello Gallico and De Bello Civili. Although the two works describe two dramatically different historical contexts, one revolving around an external war and the other around internal conflict, this study argues that Caesar maintains a highly stable rhetorical and semantic framework across the two texts. To substantiate this claim, the thesis combines three computational methods: Latent Dirichlet Allocation topic modeling, semantic vector analysis, and evaluative framing analysis. Through the topic modeling method, we see that the two texts occupy distinct thematic environments. The De Bello Gallico is dominated by geographic, ethnographic, and military vocabulary, while the De Bello Civili puts more emphasis on Roman political institutions and civic conflict. Notwithstanding these differences, semantic vector analysis shows a high degree of consistency in the use of political, civic, and military terms across both texts. Evaluative framing analysis similarly reveals a stable justificatory tone, with Roman actors consistently receiving more legitimizing language than enemies. Taken together, these findings suggest that Caesar uses linguistic consistency as a rhetorical strategy through which he presents his authority as rational, necessary, and continuous across different historical circumstances.

Training a transformer at scale is a memory bandwidth problem as much as a compute problem. Libraries like QuACK provide hand-optimised GPU kernels for every operation in a transformer block, but using them at full efficiency requires non-obvious wiring decisions: which primitives to call, in which order, and with which arguments to eliminate intermediate HBM round-trips. Those decisions require hardware knowledge that does not appear at the Python level, and making them correctly across an entire transformer block is what separates a fast implementation from an optimal one. This thesis asks whether a large language model can make those decisions automatically. The task is kernel composition: given the QuACK API and a reference PyTorch implementation of a LLaMA-3 8B transformer block, synthesise a fused torch.autograd.Function using only existing library primitives, with no new kernel code written. We propose a two-stage approach that separates fusion planning from code generation. In Stage 1, the LLM produces a kernel breakdown table showing every operation with its HBM reads, writes, and register contents. In Stage 2, it implements the function from the agreed table. This separation makes fusion decisions inspectable before any code is written and gives the human a natural point to intervene at the planning level. The resulting implementation achieves a 1.14× speedup over PyTorch eager on a single H100 at LLaMA-3 training shapes, with the advantage growing linearly with layer count. The central finding is that fusion decisions determinable from the API surface are handled automatically by the LLM; decisions that depend on runtime performance consequences require human steering. The approach reduces the barrier to building hardware-optimal fused kernels, but does not eliminate the need for someone who can interpret what the profiler is telling them.

Strict Serializability is a consistency model that ensures a single global order for transactions that is consistent with real time. Regular Sequential Serializability (RSS) weakens this guarantee for read-only transactions, instead enforcing a causal constraint. This improves the tail latency of read-only transactions while rarely returning stale values in practice. However, RSS does not have any guarantee against network reordering. Under RSS, the order in which a client invokes its concurrent transactions is not necessarily reflected in the system responses. I introduce Invocation-Ordered Regular Sequential Serializability (IO-RSS), which is a new consistency model that provides observable invocation ordering in addition to the guarantees of RSS. Alongside this, I introduce IO-Spanner-RSS, the protocol that enforces IO-RSS. IO-Spanner-RSS has two variants, Die-Die (both lock holder and requester abort on contention) and OPST (Oldest Predecessor Start Time), which define different locking protocols for handling contention. The performance of IO-Spanner-RSS is measured against Spanner-RSS configured with manual invocation ordering as a baseline. In my evaluations, IO-Spanner-RSS consistently delivers higher throughput in low-to-medium contention settings, though it degrades significantly under extreme contention.

Generative driving simulators have emerged as a promising alternative to real-world data collection for autonomous vehicle (AV) development, but they typically offer limited controllability at inference time. Recent work on language-conditioned scenario generation addresses this by allowing users to specify desired scene attributes through text. However, capturing how road structure and agent interactions evolve over time in driving sequences remains a limited area of research. This thesis develops a BEV scenario captioning approach tailored to the nuPlan dataset. We introduce a rule-based labeling pipeline that converts per-frame scene-graph data into structured target captions. We then use the resulting dataset of BEV sequences and temporally-grounded reference descriptions to fine-tune a VLM for scenario captioning.

Kinetic plasma physics simulations are essential for advancing semiconductor manufacturing, but reaching quasi-steady state convergence can require millions to hundreds of millions of time steps, taking days to weeks on modern supercomputers. This thesis investigates whether Dynamic Mode Decomposition (DMD), a lightweight data-driven technique, can accelerate these simulations by learning patterns from early simulation data and using them to leap the solution forward in time. We evaluate DMD on a series of test problems of increasing complexity: a linear advection-diffusion equation, a linear diffusion equation with boundary conditions and source terms, and finally production-scale capacitively coupled plasma (CCP) simulation data from both Particle-in-Cell (PIC) and Vlasov solvers. On clean data, DMD achieves near-zero prediction error over rollout horizons at least as long as the training window, while noisy data degrades performance significantly. We tackle the noisy data challenge through temporal smoothing, data decimation, and the Hankel time-delay embedding which mitigates its effects in varying degrees. Hankel DMD shows promise for electron density in the PIC simulations, maintaining stable errors of 1–2% with no upward drift under certain hyperparameters. Most importantly, we integrate DMD directly into a running Vlasov simulation using an online simulate-learn-leap framework, achieving a 1.95x speedup to reach 1% error relative to the ground-truth steady state, with the entire DMD training and inference step completing in under one second. These results provide a proof of concept that lightweight, online data-driven methods can meaningfully reduce the cost of expensive kinetic plasma simulations without sacrificing final accuracy, offering a path toward making these simulations practical for industrial use.

Deep Neural Networks (DNNs) are increasingly used in high-stakes settings, such as healthcare, education, hiring, lending, and autonomous vehicles. These models consume massive amounts of environmental and economic resources, and are largely inaccessible to researchers and users without access to expensive high-performance compute. Therefore, it has become essential to compress DNNs to reduce costs in storage, compute, and energy. However, compression comes with a hidden cost: existing methods degrade performance for minority subgroups and out-of-distribution data, creating unfair and non-robust models. Our novel framework, Concept-aware Network Pruning (CNP), mitigates this shortcut learning by identifying and removing human-understandable concepts within a model during pruning. CNP augments a pretrained DNN with a virtual concept layer in which nodes represent semantic concepts, ablates human-identified target concepts, and then applies node-level pruning so that nodes serving the ablated concepts are naturally irrelevant. We demonstrate on binary image classification tasks that CNP can mitigate model reliance on targeted concepts, leading to improved performance over vanilla pruning techniques for under-represented subgroups while maintaining overall accuracy. We evaluate applications in robustness and fairness in compression.

Artificial Intelligence (AI) and Large Language Models (LLMs) have entrenched themselves among the most important tools in contemporary life. AI and LLMs wield tremendous power and can be harnessed to perform the duties typically restricted to highly trained individuals. While the extent to which these tools should be incorporated into our lives is up for debate, many professionals and students now rely on these tools in their day-to-day lives. One particular area where AI stands to play a critical role is in the automation of scientific research. Skills that are required to conduct research at the highest standard, like surveying all the relevant literature, designing effective experiments, and conducting rigorous data analysis, all fall within the wheelhouse of LLMs. Although questions can be raised as to whether or not AI has gotten to the point where it can rival human researchers, human researchers can leverage AI to improve the rigor of their own experiments. This thesis aims to construct a VSCode extension that enables human researchers to invoke an AI agent to generate test cases on their code. The agent uses retrieval-augmented generation (RAG) to aid in generative test suits designed for implementations of bespoke statistical tests. Acceptance or rejection of the proposed test suit is made simple using the tab complete mechanism common to AI coding aids. By designing a user-friendly extension to generate test cases, this thesis aims to promote the adoption of thorough testing suites by ensuring researchers can easily delegate the task to agentic AI.

Why do LLMs answer questions incorrectly, even those that seem trivial to the human mind? This thesis explores an approach to understanding Machine Learning success and failure that is centered around using Mechanistic Interpretability (MI) techniques to peer inside of LLMs, in order to shed light on their internal decision-making processes. Employing a curated dataset of simple grammatical and mathematical questions, as well as Logit Lens and Residual-Stream and Attention-Head Activation Patching, the intermediary outputs of the components of openai-community/gpt2, Qwen/Qwen2.5-1.5B, and meta-llama/Llama-3.1-8B are analyzed in-depth. Our results combine accuracy, top-1 confidence, entropy, and patched logit difference to explicate how our models generate tokens, as well as the roles of different model components in overall model performance.

High application latency has a significant effect on user traffic [23]. However, reducing latency while guaranteeing consistency is difficult because storage systems are located far from end users. While recent work improves latency for strong consistent applications by running closer to users, it imposes high cost for storage and compute [19]. In this project, we propose Radical-Local in order to minimize cost overhead while providing low latency for consistent applications. Radical-Local speculatively executes applications on end-user machines. It guarantees linearizability by simultaneously syncing with the primary datacenter. When speculative execution succeeds, Radical-Local avoids the cost of using cloud compute. It is also faster than execution in a distant datacenter. For real application deployments, Radical-Local reduces cost by 9 − 29% and median latency by 7% compared to previous literature [19]. This represents a latency improvement of 48% compared to execution in the datacenter.

Training large language models to predict human behavior is a promising alternative to traditional cognitive models, but it remains unclear whether behavioral fine-tuning induces genuine strategic understanding or surface-level alignment. I tested twelve LLMs on a dataset of more than 2,400 two-player 2 × 2 matrix games, asking models to predict aggregate human choice distributions from natural language game descriptions. Larger models generally aligned better with human behavior, but none matched traditional cognitive models such as the level-k quantal response model. To improve alignment, I fine-tuned four instruction-tuned LLMs and evaluated them on both a primary task (predicting human choice distributions) and an out-of-distribution task (choosing actions in the games themselves rather than predicting human frequencies). Fine-tuning improved performance on both: models became more accurate at predicting human choice distributions, and their own choice probabilities became more aligned with Nash equilibrium play and empirical human behavior. However, in the out-of-distribution task, models reduced overconfidence uniformly rather than adapting to each game’s payoff structure. This demonstrates that behavioral alignment and context-dependent understanding are dissociable and that standard performance metrics may not distinguish between them.

Molecular property prediction is central to computational drug discovery and quantum chemistry, yet standard graph neural networks are limited by their reliance on pairwise atomic interactions. Existing hypergraph approaches extend this framework to multi-atom structures but rely on fixed, fingerprint-derived motif vocabularies that cannot adapt to the prediction task. This thesis presents AdaptiveHON, a framework that generates candidate multi-atom subgraphs via depth-first search, represents each candidate by a 36-dimensional structural feature vector, and uses a learned scoring network to assign continuous relevance weights that gate higher-order message passing. AdaptiveHON improves on a pairwise MPNN baseline on all 19 QM9 targets and outperforms the fixed hypergraph baseline HyperMol on 6 of 19 targets, with the most consistent gains on electronic structure properties including the HOMO energy (5.1% reduction in MAE), the LUMO energy (17.0%), the HOMO–LUMO gap (14.1%) and ZPVE (74.9%) compared to the simple MPNN. On the αxx polarizability tensor component, Adaptive- HON achieves a 55.0% reduction in MAE relative to the MPNN and is the only condition to outperform geometry-aware SchNet on any target. Ablation studies show that learned scoring contributes up to 35% of the gain on targets where the signal is concentrated in a subset of distinct subgraphs. These results suggest that adaptive higher-order representations can capture structure that pairwise methods miss, and that which substructures to model matters as much as the message-passing architecture itself.

This paper presents a real-time emotion-conditioned generative music system that uses facial recognition to continuously adapt musical output to a user’s detected affective state. Emotion is represented using the valence-arousal framework derived from Russell’s circumplex model of affect. A Conditional Variational AutoEncoder (cVAE) combined with a Gated Recurrent Unit (GRU) serves as the generative backbone. This paper will focus on the construction of the real-time adaptive model as well as the literature that led to its conversation.