From finance and healthcare to criminal justice and transportation, many domains that involve high-stakes decisions, traditionally made by humans, are increasingly integrating artificial intelligence (AI) systems into their decision making pipelines. While recent advances in machine learning and optimization have given rise to AI systems with unprecedented capabilities, fully automating such decisions is often undesirable. Instead, a promising direction lies in AI-assisted decision making, where AI informs or complements human decisions without completely removing human oversight. In this talk, I will present my PhD work on AI-assisted decision making in settings where humans rely on two core cognitive capabilities: strategic reasoning and counterfactual reasoning. First, I will introduce game-theoretic methods for supporting policy design in strategic environments, enabling a decision maker to allocate resources (e.g., loans) to individuals who adapt their behavior in response to transparency regarding the decision policy. Next, I will present methods to enhance a decision maker’s counterfactual reasoning process— identifying key past decisions (e.g., in clinical treatments) which, if changed, could have improved outcomes and, hence, serve as valuable learning signals. Finally, I will discuss a computational model of how people attribute responsibility between humans and AI systems in collaborative settings, such as semi-autonomous driving, evaluated through a human subject study. I will conclude with key takeaways and future directions for designing AI systems that effectively support and interact with humans.

From LLMs Algorithmic ranking systems shape online interactions and are linked to engagement misinformation and polarization. This talk explores the unintended consequences of popularity-based rankings drawing on two studies that combine observational data mathematical modeling and experimental validation. The first examines how engagement-driven algorithms may fuel polarization and ideological extremism. The second reveals a "few-get-richer" effect where ranking dynamics amplify minority signals. Together these insights highlight the societal impact of human-AI feedback loops.

•Ranking for engagement: How social media algorithms fuel misinformation and polarization. F Germano V Gómez F Sobbrio. CESifo Working Paper •The few-get-richer: a surprising consequence of popularity-based rankings. F Germano V Gómez GL MensThe Web Conference WWW' 19

As we shall see, LLMs like GPT4 can model group behavior — or stereotypical patterns — fairly well, but their performance drops when it comes to smaller groups, below a certain threshold. This limitation has interesting implications not just for personalization systems, but also for questions in social science, including how we define culture, interpret collective behavior, and think about individual agency.

Rust is a modern systems programming language whose ownership-based type system statically guarantees memory safety, making it particularly well-suited to the domain of safety-critical systems. In recent years, a multitude of automated deductive verification tools have emerged for establishing functional correctness of Rust code. However, before my thesis work, none of the tools produced foundational proofs (machine-checkable with a small trusted computing base), and all of them were restricted to the safe fragment of Rust. This is a problem because the vast majority of Rust programs make use of unsafe code at critical points: either indirectly through libraries, which are often implemented with unsafe code, or directly, which is particularly common for low-level systems code.

With my thesis, I will present work on RefinedRust, a foundational verification tool, aiming to semi-automatically establish functional correctness of both safe and unsafe Rust code. RefinedRust builds a refinement type system for Rust that extends Rust's type system for safe Rust to functional correctness reasoning and unsafe Rust, and, inspired by previous work on RefinedC, automates proofs in this type system inside the Rocq Prover. RefinedRust's type system is semantically proven sound using the Iris separation logic framework, providing high assurances about its correctness. In the process, RefinedRust significantly advances the state of the art of semantic soundness proofs for Rust, scaling previous work on RustBelt to a much more practical and realistic model.

One particularly interesting application for RefinedRust is low-level systems code, which is inherently unsafe and deserving of high-assurance foundational verification. In my thesis, I will present work on verifying interesting parts of the ACE security monitor, which is part of the ACE trusted execution environment (TEE) developed by IBM Research and thus a low-level security-critical component. For the final part of my thesis, I propose to work on one of two projects that will advance our capabilities in verifying systems code like ACE. I will either show how to extend RefinedRust to prove non-interference (a security property) of systems code, or show how to link up RefinedRust with hardware models to prove a fundamental isolation property that ACE should guarantee.

There is growing awareness that the statistical analysis of personal data, such as individual mobility, financial, and health data, could be of immense benefit to society. However, liberal societies have refrained from such analysis, arguably due to the lack of trusted analytics platforms that scale to billions of records and reliably prevent the leakage and misuse of personal data. The first part of this proposal presents CoVault, an analytics platform that leverages secure multi-party computation (MPC) and trusted execution environments (TEEs) to perform analytics securely at scale. CoVault co-locates MPC's independent parties in the same datacenter, providing high bisection bandwidth without reducing security, allowing analytics to scale horizontally to the datacenter’s available resources. CoVault's empirical evaluation shows that nation-scale, secure analytics is feasible with modest resources. For example, country-scale epidemic analytics on a continuous basis requires only one pair of CPU cores for every 30,000 people. The second part of the proposal discusses how to add support for semantic queries to CoVault by building on secure similarity search on high-dimensional vector embeddings.

The concept of emergence is frequently invoked in discussions on artificial intelligence (AI) to describe seemingly novel, unpredictable, and autonomous system behaviours. These narratives often anthropomorphise AI systems, attributing to them intelligence, agency, and even creativity – fuelling public fascination and intensifying normative uncertainty. However, a robust normative and legal analysis requires a precise conceptualisation of emergence that distinguishes between technical explanations, subjective projections, and philosophical perspectives. Do systems based on artificial neural networks (ANNs) acquire emergent properties? Drawing on complexity theory, I will argue that the predictive capacities of ANNs correspond to weak emergence, which is explained by interactions of the structural components within ANNs and the conditions of model deployment, rather than indicating ‘superimposed’, ‘autonomous’ agency. This understanding challenges agentic interpretations of AI functionalities, highlights the risks of over-attributing human-like traits to AI, and redirects attention to human causality and responsibility in AI-mediated outcomes. The discussion will further explore the implications of a reductionist understanding of AI emergence for the design of AI governance frameworks, and emphasise the role of human agency in AI-driven processes. In particular, this view can inform the ongoing debates on AI risk regulation, civil liability, the ethical boundaries of AI-driven decision-making, and reward mechanisms such as intellectual property rights. Overall, the analysis aspires to contribute to a conceptual shift – from the mythologisation of AI capabilities toward a technically grounded normative analysis.

From LLMs can imitate humans and therefore, pass the Turing test – this aint a news today. However, can they model not just a human, but you as an individual? This talk dives into the core of personalization: how well can large language models predict individual user behavior, based on their history and context? Since LLMs are trained to generate sequential data — and that data comes from humans — there's reason to believe they can mirror our choices, preferences, even our quirks. We explore this in two lights:

•Positive prediction: what users are likely to do, prefer, or know — the classic personalization setup. •Negative prediction: what users are unlikely to do or understand — the shadow side of modeling.

As we shall see, LLMs like GPT4 can model group behavior — or stereotypical patterns — fairly well, but their performance drops when it comes to smaller groups, below a certain threshold. This limitation has interesting implications not just for personalization systems, but also for questions in social science, including how we define culture, interpret collective behavior, and think about individual agency.

Today’s Internet is responsible for connecting billions of end-points that interact by using a variety of protocols. Malicious actors can carry out attacks by exploiting users, applications, or the network. Typical attacks are two steps process: The identification of potential victims by periodically scanning the Internet for a specific service; and the attack itself where the exploitation of the target service is performed.

To protect systems on the Internet, we need to understand how vulnerable systems are being targeted, how malicious actors exploit end-points, and how attackers' behavior evolves over time. Network measurements offer valuable resources to study such behaviors. This talk highlights how Internet measurements can be used to improve security, and which factors have to be considered in this research direction.

In this talk, I will present a series of collaborative efforts—developed with my students and colleagues at Microsoft India—on designing robust safety interventions in large language models (LLMs). Our first study, TechHazardQA (to appear at ICWSM 2025), shows how instruction-form prompts, such as code or pseudocode, can bypass content filters that typically block harmful text queries. We propose a mitigation strategy called the SafeInfer (AAAI 2025), leveraging function vectors and controlled text generation via model arithmetic to steer models away from harmful outputs. We extend this safety strategy to multilingual settings by introducing language-specific safety vectors. Applied across several languages, this approach yields a notable reduction in attack success rates, from over 60% to below 10%. While text is one of the most dominant forms of communication on social media platforms, the internet meme culture is rapidly becoming viral. We show how vision language models (VLMs) that are routinely used as backbones for multimodal content moderation by social media platforms are often ineffective in tracking toxic memes. We tease apart the reasons for this failure using a rigorous interpretability scheme (to appear at ICWSM 2025). Presently, we are working on building intervention techniques for both toxic memes and audio podcasts. Next, I will discuss cultural sensitivities in generative models, demonstrating that the same prompt can elicit significantly different degrees of harm when framed in diverse cultural contexts. We propose a lightweight, preference-informed realignment technique that reduces culturally conditioned harmful responses across 11 regions (NAACL 2025).

The field of Law has become an important application domain of Natural Language Processing (NLP) due to the recent proliferation of publicly available legal data, and the socio-economic benefits of mining legal insights. Additionally, the introduction of Large Language Models (LLMs) has brought forth many applications, questions, and concerns in the legal domain. In this talk, I will discuss some of the challenges in processing of legal text, and our works in some practical research problems, including summarization of long legal documents, and identifying relevant statutes from fact descriptions. I will also discuss our efforts towards making Law-AI more accessible in the Indian society, through developing the first datasets for several problems (available at https://github.com/Law-AI/) and the first pre-trained language model for the Indian legal domain(https://huggingface.co/law-ai/InLegalBERT) that is being widely used by both the academia and the industry, and has more than 1.5 million downloads till date.

Machine-Learning (ML) models are increasingly integrated into safety-critical systems, from self-driving cars to aviation, making their dependability assessment crucial. This thesis introduces novel approaches to specify and test the functional correctness of ML artifacts by adapting established software testing concepts. We first address the challenge of testing action policies in sequential decision-making problems by developing π-fuzz, a framework that uses metamorphic relations between states to identify undesirable yet avoidable outcomes. We then formalize these relations as k-safety hyperproperties and introduce NOMOS, a domain-agnostic specification language for expressing functional correctness properties of ML models. NOMOS comes with an automated testing framework that effectively identifies bugs across diverse domains including image classification, sentiment analysis, and speech recognition. We further extend NOMOS to evaluate code translation models.

By providing these specification languages and testing frameworks, this thesis contributes essential tools for validating the reliability and safety of ML models in our increasingly machine-learning-dependent world.

Digital infrastructure is critical to modern society, and a great deal of work goes into ensuring the smooth and continuous operation of systems and networks. This essential labor is carried out by system and network operators. Yet, their work often remains invisible and undervalued—especially when everything appears to function as expected. System operators engage not only in a wide range of technical tasks but also in social and organizational work, such as coordinating with colleagues and helping system users. Their everyday practices directly shape the security posture of their organizations. However, when failures occur, system administrators are frequently blamed for misconfiguration or other types of "human error". Decades of research in human factors demonstrate that focusing on human error alone is insufficient to improve operational security. Instead, it diverts attention from the sociotechnical complexities of these environments and from supporting people in doing their work effectively. This talk will highlight the human dimensions of system operations by drawing on historical perspectives and emphasizing the sociotechnical factors essential to sustaining and securing digital infrastructure.

From understanding voting rules to designing auctions and allocating public goods, approaches from theoretical computer science (TCS) have proved useful for analyzing and addressing a variety of societal problems. These problems are commonly studied in political science and economics, but their practical solution depends on law – i.e., the set of rules produced by legal systems. In this talk, I ask how concepts and methods from TCS can help us understand legal systems, developing my vision for computational legal theory.

In recent years, modern blockchains like Ethereum have seen a surge in adoption, largely due to their ability to support decentralized finance (DeFi)— a new class of open, permissionless financial tools accessible to anyone. The security of DeFi hinges on both the robustness of smart contracts and the order in which transactions are processed, while its privacy is closely tied to protections offered by user wallets. In this talk, we’ll begin with a brief overview of the blockchain architecture, then delve into the privacy implications of Web3 wallets and how cross-chain interoperability can give rise to exploitative financial behavior, such as cross-chain sandwich attacks. In the first part, we’ll examine how wallets can unintentionally expose user addresses to third parties and how they are increasingly being used to track users across the web. In the second part, we’ll analyze how asymmetries between Ethereum and its Layer-2 rollups create opportunities for malicious actors to exploit bridges and perform cross-chain sandwiching—a form of predatory price manipulation once thought limited to single-chain environments.

Block-based visual programming initiatives such as Hour of Code by code.org and Intro to Programming with Karel by CodeHS.com, have transformed introductory computer science education by making programming more accessible to K-8 learners. Despite their accessibility, students often struggle with multi-step reasoning and conceptual abstraction when solving open-ended tasks. Quizzes (such as fill-in-the-gap exercises, and multiple-choice conceptual questions based on code debugging and task design) offer interactive practice and targeted feedback that can promote active learning and scaffold novice programmers. However, manually designing such quizzes is time-consuming and difficult to scale. This thesis tackles these challenges by developing automated synthesis techniques for programming tasks and quizzes, and evaluates their pedagogical utility.

The first part of the thesis introduces algorithmic methods for synthesizing programming tasks and quizzes in block-based environments. Specifically, we develop methods for the following : (i) synthesizing conceptually similar and yet visually dissimilar write-code tasks; (ii) synthesizing adaptive multiple-choice programming quizzes that address student-specific misconceptions; and (iii) synthesizing scaffolded subtasks that break-down complex write-code tasks into simpler subtasks. Each method leverages symbolic execution, sketch-based code mutation, and search-guided generation to ensure pedagogical utility, relevance, and technical correctness. Empirical evaluations conducted through controlled user studies demonstrate the efficacy of these approaches, showing that they not only support novice learners effectively but also outperform existing methods, including next-step code edit based feedback methods.

The second part of the thesis empirically evaluates the pedagogical utility of programming quizzes in these environments via user studies and classroom deployments with K-8 learners. Specifically, we examine: (i) the effectiveness of quiz-based feedback scaffolds with different quiz-types; (ii) the design, validation, and classification of quiz types using cognitive frameworks such as Bloom's Revised Taxonomy; and (iii) the impact of embedding quizzes within programming curricula on post-learning outcomes. Our findings show that quizzes designed using metacognitive strategies and adapted to learners’ attempts significantly enhance engagement and task performance. Moreover, we observe that richer and more diverse quiz types—when integrated into the curriculum— lead to improved post-learning outcomes, while simpler, less cognitively demanding quizzes may hinder post-learning performance.

Overall, this thesis contributes novel synthesis methods for programming quizzes and empirical evidence of their effectiveness in elementary-level programming education. These findings provide a foundation for scalable and adaptive support in elementary computing curricula.

The problem of how to scalably verify large, stateful programs is one of the oldest—and still unsolved—challenges of computer science. Over the last two decades, there has been considerable progress toward this goal with the advent of separation logic, a verification technique for modularly reasoning about stateful programs. While originally only developed for imperative, pointer-manipulating programs, separation logic has in its modern form become an essential tool in the toolbox of the working semanticist for modeling programming languages and verifying programs.

With this thesis, I present a line of work that revisits the foundations of modern separation logic in the context of the separation logic framework Iris. It targets two broader areas: step-indexing and automation. Step-indexing is a powerful technique for modeling many of the advanced, cyclic features of modern languages. Here, Transfinite Iris shows how to generalize step-indexing from proving safety properties to proving liveness properties, and Later Credits enable more flexible proof patterns for step-indexing based on separation logic resources. Automation, on the other hand, is important for reducing the overhead of verification to scale to larger code bases. Here, Quiver introduces a new form of guided specification inference to reduce the specification overhead of separation logic verification, and Daenerys develops new resources in Iris that lay the groundwork for automating parts of Iris proofs using SMT solvers.

Efficient sampling from probabilistic models plays an important role in many fields, such as computational social science, materials science, and generative AI. Many complex multivariate distributions in these domains can be represented through simple interactions of random variables along the edges of graphs and hypergraphs. The generality and natural-ness of such models make them powerful tools for both theoretical and algorithmic applications, and as a result we have seen significant recent progress in efficient sampling techniques. In this talk, I will present a relatively recent framework for approximately sampling from these distributions. At a high level, this approach views complex distributions as sparse perturbations of a family of simpler base distributions. This perspective enables the use of powerful enumerative and analytical tools and motivates new techniques. I will also talk about some recent progress in our understanding of these distributions through the lens of our methods. In addition to providing algorithmic insight, this work work also reveals surprising probabilistic applications of studying these distributions. Finally, I will conclude with some exciting current and potential future directions.

This collection of studies leverages digital trace data to explore human migration, cultural similarity across countries, and online behavior in algorithmically mediated platforms. By analyzing platforms such as Facebook, Wikipedia, and TikTok, the research uncovers patterns and connections often overlooked or delayed by traditional methods. In the context of culture and migration, two studies leverage Facebook data to measure cultural similarity between countries and demonstrate the influence of cultural similarity on migration flows. Another study highlights the innovative use of Wikipedia data as a tool to track mass migration flows, offering real-time insights into population movements, particularly during crises like conflicts and wars. These studies primarily rely on the use of Application Programming Interface (API) for data access. However, as social media platforms increasingly restrict access to their data, alternative approaches like data donation, facilitated by GDPR, offer new opportunities for data collection. The final study leverages TikTok user-donated data to evaluate the predictability of user engagement with short-form videos, highlighting the potential of data donation for research. Together, these studies showcase the potential of digital trace data in advancing knowledge across social science and computational fields.

Ample research has demonstrated that compliance with data protection principles remains limited on the web and mobile. For example, very few apps on the Google Play Store fulfil the minimum requirements regarding consent under EU law, while most of them share data with companies like Google and Meta, and would likely need to seek consent from their users. Given the current mismatch between the law on the books and data practices in reality, iterative changes to existing legal practice will not be sufficient to meaningfully curb egregious data practices. Hence, this talk discusses a range of suggestions for academia, regulators, and the interested public to move beyond the status quo. The talk also explores new mechanisms to hold online platforms to account, in particular the 2022 EU Digital Services Act, which obliges platforms to provide researchers with relevant data for their work.

QUIC is a UDP-based multiplexed and secure transport protocol that was standardized in 2021 by the IETF. It seeks to replace the traditional TCP/TLS stack by combining functionality from different layers of the ISO/OSI model. Therefore, it reduces overhead and introduces new functionality to better support application protocols, e.g., streams and datagrams. QUIC is the foundation for HTTP/3 and new proxy technologies (MASQUE). It is used for video streaming and considered for other media services.

This talk will introduce the protocol and motivate its relevance. In the second part, I will provide insights into existing implementations and their performance. Our research shows that QUIC performance varies widely between client and server implementations from 90 Mbit/s to over 6000 Mbit/s. In the second part, I provide an overview about QUIC deployments on the Internet. At least one deployment for 18 different libraries can actually be found on the Internet.

The complexity of the protocol, the diversity of libraries and their usage on the Internet makes QUIC an important research subject.

The impact of algorithmic decision-making systems on individuals has raised significant interest in addressing fairness concerns within such systems. Designing fair systems entails several critical components, which have garnered considerable attention from the research community. However, notable gaps persist in three key components. Specifically, in this thesis, we address gaps in following components: i) evaluating existing approaches and systems for (un)fairness, ii) updating deployed algorithmic systems fairly, and iii) designing new decision-making systems from scratch. Firstly, we evaluate fairness concerns within foundation models. The primary challenge is that fairness definitions are task-specific while foundation models can be used for diverse tasks. To address this problem, we introduce a broad taxonomy to evaluate the fairness of popular foundation models and their popular bias mitigation approaches. Secondly, we tackle the issue of fairly updating already deployed algorithmic decision-making systems. To this end, we propose a novel notion of update-fairness and present measures and efficient mechanisms to incorporate this notion in binary classification.  However, in cases where there is no deployed system or updating an existing system is prohibitively complex, we must design new fair decision-making systems from scratch. Lastly, we develop new fair decision-making systems for three key application scenarios. Major challenges in designing these systems include computational complexity, lack of existing approaches to tackle fairness issues and designing human-subject based studies. We develop a computationally efficient mechanism for fair influence maximization to make the spread of information in social graphs fair. Additionally, we address fairness concerns under model uncertainty, i.e., uncertainty arising due lack of data or the knowledge about the best model. We propose a novel approach for training nondiscriminatory systems that differentiate errors based on their uncertainty origin and provide efficient methods to identify and equalize errors occurring due to model uncertainty in binary classification. Furthermore, we investigate whether algorithmic decision-aids can mitigate inconsistency among human decision-makers through a large-scale study testing novel ways to provide machine advice.

Reward functions are central in reinforcement learning (RL), guiding agents towards optimal decision-making. The complexity of RL tasks requires meticulously designed reward functions that effectively drive learning while avoiding unintended consequences. Effective reward design aims to provide signals that accelerate the agent’s convergence to optimal behavior. Crafting rewards that align with task objectives, foster desired behaviors, and prevent undesirable actions is inherently challenging. This thesis delves into the critical role of reward signals in RL, highlighting their impact on the agent’s behavior and learning dynamics and addressing challenges such as delayed, ambiguous, or intricate rewards. In this thesis work, we tackle different aspects of reward shaping. First, we address the problem of designing informative and interpretable reward signals from a teacher’s/expert’s perspective (teacher-driven). Here, the expert, equipped with the optimal policy and the corresponding value function, designs reward signals that expedite the agent’s convergence to optimal behavior. Second, we build on this teacher-driven approach by introducing a novel method for adaptive interpretable reward design. In this scenario, the expert tailors the rewards based on the learner’s current policy, ensuring alignment and optimal progression. Third, we propose a meta-learning approach, enabling the agent to self-design its reward signals online without expert input (agent-driven). This self-driven method considers the agent’s learning and exploration to establish a self-improving feedback loop

People differ in their ability and willingness to think about the future and consider the longer-term consequences of their behavior. Short-term thinking is associated with a range of self-defeating behaviors, such as delinquency, gambling, alcohol and drug abuse, as well as internalizing problems, such as depressive symptoms. In contrast, people who consider the longer-term consequences of their decisions tend to report more positive outcomes, such as feeling more competent, better health, and enhanced educational achievement. One way to explain individual differences in thinking about the future is rooted in so-called multiple self models, which distinguish between a present and a future self. In this talk I will elaborate on FutureU, a research program that aims to stimulate future-oriented thinking, increase goal achievement, and reduce self-defeating behavior, by strengthening people’s identification with their future self. The intervention is delivered through a smartphone application (app) or immersive Virtual Reality (VR). I will discuss the theoretical background of FutureU, present some empirical results, and identify areas where it can be further improved.

Cloud systems are composed of multiple inter-connected independent systems. These systems are complex in nature as they are made of heterogeneous components rife with complicated interactions, operate in dynamic conditions, and exhibit unpredictable behaviors. Despite all the complexity, developers of these systems are tasked to efficiently design, implement, optimize, operate, and improve these systems in a continuous fashion. A proposed way of managing the complexity for designing, implementing, and optimizing these systems is to automate these tasks. There are three major roadblocks preventing this automation from becoming reality - (i) lack of abstractions for design and implementation and design exploration of cloud systems; (ii) lack of abstractions and tooling converting user's high level design intent into actual implementations;  (iii) lack of abstractions for leveraging runtime information for optimizing cloud systems. I propose new abstractions for cloud systems, with a special focus on microservice systems, for automating developer tasks. The work I will present takes us one step closer to the vision of automating the design, implementation, and optimization of cloud systems whilst managing the inherent complexity of these systems.

As machine learning (ML) models continue to scale in size and capability, they expand the surface area for safety and privacy risks, raising concerns about model trustworthiness and responsible data use. My research uncovers and mitigates these risks. In this presentation, I will focus on the three cornerstones of trustworthy foundation models and agents: safety, privacy, and generalization. For safety, I will introduce our comprehensive benchmarks designed to evaluate trustworthiness risks in Large Language Models (LLMs) and LLM-based code agents. For privacy, I will present a solution for protecting data privacy with a synthetic text generation algorithm under differential privacy guarantees. The algorithm requires only LLMs inference API access without model training, enabling efficient safe text sharing. For generalization, I will introduce our study on the interplay between the memorization and generalization of LLMs in logical reasoning during the supervised fine-tuning (SFT) stage. Finally, I will conclude with my future research plan for assessing and improving trustworthiness in foundation model-powered ML systems.

Despite widespread enthusiasm for AI from governments worldwide, ensuring that AI adoption positively impacts society remains a challenge. Focusing on applications in criminal justice and social services, this talk will examine the significant gaps between current AI capabilities and the demands of real-world high-stakes decision-making. I will demonstrate critical shortcomings of current approaches to AI transparency, fairness, and effective human oversight, and discuss my work on addressing these issues, and its impact on policy and UK public services to date.  Concretely, I will first show how we used statistical modelling to uncover racial bias in algorithmic risk assessment instruments used for bail and sentencing decisions. Next, I will turn to my work on human-AI interaction that combines large language model speed with human accuracy to extract information from unstructured documents with high-precision. This system has served practitioners across disciplines, and is a core component of our pioneering effort to enable researcher access to UK Crown Court transcripts. I will conclude by outlining my research vision for developing urgently needed evaluation and auditing tools for human-AI systems deployed in high-risk decision-making contexts.

Collecting user data is crucial for advancing machine learning, social science, and government policies, but the privacy of the users whose data is being collected is a growing concern. Organizations often deal with a massive volume of user data on a regular basis — the storage and analysis of such data is computationally expensive. Thus developing algorithms that not only preserve formal privacy but also perform efficiently is a challenging and important necessity. Since preserving privacy inherently involves some data distortion which potentially sacrifices accuracy for smaller populations, a complementary challenge is to develop responsible privacy practices that ensure that the resulting privacy implementations are equitable. My talk will focus on Differential Privacy (DP) --- a rigorous mathematical framework that preserves the privacy of individuals in the input dataset, and explore the nuanced landscape of privacy-preserving algorithms through three interconnected perspectives: the systematic design of both time and space-efficient private algorithms, and strategic approaches to creating equitable privacy practices.

Optical data center networks (DCNs) are transforming cloud infrastructure, yet current architectures remain closed ecosystems tightly bound to specific optical hardware. In this talk, we unveil an innovative open framework that decouples software from hardware, empowering researchers and practitioners to freely explore and deploy diverse software solutions across multiple optical platforms. Building on this flexible foundation, we tackle three critical system challenges—time synchronization, routing, and transport protocols—to enable optical DCNs to achieve nanosecond-precision, high throughput, and ultra-low latency. This presentation highlights the fundamental design shifts brought by optical DCNs and demonstrates how our breakthrough solutions surpass traditional DCN performance, setting new standards for future cloud networks.

Millions of everyday users are interacting with technologies built with generative AI, such as voice assistants, search engines, and chatbots. While these AI-based systems are being increasingly integrated into modern life, they can also magnify risks, inequities, and dissatisfaction when providers deploy unreliable systems. A primary obstacle to having reliable systems is the opacity of the underlying large language models - we lack a systematic understanding of how models work, where critical vulnerabilities may arise, why they are happening, and how models must be redesigned to address them. In this talk, I will first describe my work in investigating large language models to illuminate when and how they acquire knowledge and capabilities. Then, I will describe my work on building methods to enable greater data transparency for large language models, that allows stakeholders to make sense of the information available to models. Finally, I will describe my work on understanding how this information can get distorted in large language models, and implications for building the next generation of robust AI systems.

Bringing the promise of quantum computation into reality requires not only building a quantum computer but also correctly programming it to run a quantum algorithm. To obtain asymptotic advantage over classical algorithms, quantum algorithms rely on the ability of data in quantum superposition to exhibit phenomena such as interference and entanglement. In turn, an implementation of the algorithm as a program must correctly orchestrate these phenomena in the states of qubits. Otherwise, the algorithm would yield incorrect outputs or lose its computational advantage. Given a quantum algorithm, what are the challenges and costs to realizing it as a program that can run on a physical quantum computer? In this talk, I answer this question by showing how basic programming abstractions upon which many quantum algorithms rely – such as data structures and control flow – can fail to work correctly or efficiently on a quantum computer. I then show how we can leverage insights from programming languages to re-invent the software stack of abstractions, libraries, and compilers to meet the demands of quantum algorithms. This approach holds out a promise of expressive and efficient tools to program a quantum computer and practically realize its computational advantage.

As learning algorithms become more capable, they are used to tackle an increasingly large spectrum of tasks. Their applications range from understanding images, speech and natural language to making socially impactful decisions, such as about people's eligibility for loans and jobs. Therefore, it is important to better understand both the consequences of algorithmic decisions and the mechanisms by which algorithms arrive at their outputs. Of particular interest in this regard are fairness when algorithmic decisions impact people's lives and the behavior of deep learning algorithms, the most powerful but also opaque type of learning algorithm. To this end, this thesis makes two contributions: First, we study fairness in algorithmic decision-making. At a conceptual level, we introduce a metric for measuring unfairness in algorithmic decisions based on inequality indices from the economics literature. We show that this metric can be used to decompose the overall unfairness for a given set of users into between- and within-subgroup components and highlight potential tradeoffs between them, as well as between fairness and accuracy. At an empirical level, we demonstrate the necessity for studying fairness in algorithmically controlled systems by exposing the potential for discrimination that is enabled by Facebook's advertising platform. In this context, we demonstrate how advertisers can target ads to exclude users belonging to protected sensitive groups, a practice that is illegal in domains such as housing, employment and finance, and highlight the necessity for better mitigation methods.

The second contribution of this thesis is aimed at better understanding the mechanisms governing the behavior of deep learning algorithms. First, we study the role that invariance plays in learning useful representations. We show that the set of invariances possessed by representations is of critical importance in determining whether they are useful for downstream tasks, more important than many other factors commonly considered to determine transfer performance. Second, we investigate memorization in large language models, which have recently become very popular. By training models to memorize random strings, we uncover a rich and surprising set of dynamics during the memorization process. We find that models undergo two phases during memorization, that strings with lower entropy are harder to memorize, that the memorization dynamics evolve during repeated memorization and that models can recall tokens in random strings with only a very restricted amount of information.