Invited Speakers

As software engineers increasingly direct, review, and integrate AI-assisted code, organizations need better ways to decide which changes are safe to ship and which deserve scarce expert attention. This talk presents a line of work on Diff/Defect Risk Score (DRS), a just-in-time defect prediction system that estimates whether a code change is likely to cause a production fault at Meta.

The talk begins with the deployment setting at Meta, where DRS has been remarkably successful. By using LLM-based risk prediction to distinguish safer changes from riskier ones, DRS enabled release processes to move away from broad code freezes toward more targeted, risk-aware deployment decisions. This reduced unnecessary blocking of low-risk work while still helping teams control production risk.

The same risk signal can also improve code review. Change safety depends not only on whether a diff is risky, but also on whether the right people review it. The Contribution-aware Changeset Safety Ratio captures the combined expertise of the author and the reviewer team, and motivates using risk signals to focus expert reviewers on the riskiest diffs.

Together, these works show how defect prediction becomes most valuable when embedded into real engineering workflows: deciding what can safely ship, selecting the right reviewers for risky changes, and helping engineers focus attention where it matters most.

AI is being deployed faster than the controls around it, and security teams are the ones paying the bill. In this session, Vincent shows how Databricks gives you a single governed foundation for data, AI, and agents, so innovation doesn’t outrun oversight. The talk closes with a look at Lakewatch, Databricks’ open agentic SIEM built for machine-speed defense.

This talk explores the emerging paradigm of foundation models trained on transaction data: large-scale models that learn behavioural embeddings from billions of financial events, much like LLMs learn from text. We discuss how these models are built and used, on their own or in tandem with LLMs, and share how they unlock powerful downstream applications across credit decisioning, fraud detection, and personalized advisory, opening a new chapter in AI for financial services.

Shipping an AI reviewer into daily engineering use comes down to a series of decisions about scope, signal, filtering, and audience, and each decision has a real tradeoff. Using Mozilla’s Review Helper, an AI code reviewer running across the Firefox codebase, as a running example, this talk discusses how those decisions compound, where the non-obvious tradeoffs sit, and how the choices that look small at the start come to define what the agent can and cannot be.

We reflect on which decisions proved load-bearing in practice and on the open questions they raise for the design and evaluation of agentic systems in software engineering.

The deployment of AI systems in heavily regulated domains such as healthcare and finance presents unique challenges that extend beyond technical performance metrics to encompass fundamental questions of trust, accountability, and societal impact. Both sectors share critical characteristics: they involve high-stakes decisions with material consequences for individuals, operate under stringent regulatory oversight, and are fundamentally client-facing, requiring that end-users place significant trust in system outputs that may influence their health outcomes or financial wellbeing.

This talk examines the core pillars of trustworthy AI in these contexts, including transparency and explainability of model decisions, robustness to distribution shift and adversarial inputs, fairness across demographic groups, rigorous validation against domain-specific standards, and mechanisms for human oversight and intervention.

Trustworthiness is a practical necessity, not merely an ethical desideratum. Systems that fail to earn stakeholder trust face regulatory barriers, adoption resistance, and reputational risk. By drawing parallels between both heavily regulated domains, the talk identifies transferable lessons for building AI systems that meet the elevated standards these domains demand.

Robots are moving from hand-programmed behavior toward systems that learn from data. This talk introduces robot learning as the problem of teaching robots how to behave in physical environments using demonstrations, interaction data, learned representations, and optimization.

The session discusses why learning robot behavior is difficult: robots must act under physical constraints, generalize across objects and scenes, and remain robust when real-world conditions differ from the training data. It also connects recent progress in deep generative models, 3D robot learning, learning from video, and motion optimization to the long-term goal of building more general robot systems.

The talk gives SEMLA attendees a practical view of how data-driven methods are changing robotics, and what software engineering challenges emerge when learned robot policies must be evaluated, deployed, monitored, and trusted in the real world.

Artificial intelligence has fundamentally transformed a wide range of industries over the past decade and is expected to reshape nearly every aspect of society in unprecedented ways. Software developers are at the forefront of this development, both as creators and as professionals whose working lives are shaped by the very technology they build. Rather than viewing AI as a neutral set of tools, it is becoming clear that it is embedded in complex social, economic, and ecological systems. From dependence on global networks of labor, data, and natural resources to the diverse applications of AI systems, these technologies raise critical questions about ethics, responsibility, and what we as humans consider a decent life. The talk explores how higher education in science and engineering fields, particularly in software engineering, can respond to these pressing questions of our times.

Artificial Intelligence (AI) models are energy-hungry. As these models grow larger and become more complex, their energy consumption increases substantially. This leads to higher carbon emissions and rising operational costs, creating serious challenges for sustainable computing. This talk highlights the environmental impact of modern AI systems, introduces key metrics for measuring energy use and carbon footprint, and discusses available tools and frameworks for measuring and monitoring energy consumption.

The session summarizes current research and practical techniques for developing greener AI systems, including model pruning, quantization, efficient architecture design, and workload scheduling, and situates these within the wider movement toward responsible and sustainable AI development. The talk concludes by highlighting important open research challenges and calling on the community to treat energy efficiency not as an afterthought but as a first-class objective in the design and evaluation of AI systems.

Multi-agent AI systems are increasingly used to coordinate agents, tools, permissions, and delegation across complex workflows. However, current evaluations often focus mainly on final task success, which may not show whether the system’s internal coordination structure has actually been evaluated. This talk introduces a new evaluation approach for multi-agent systems, focusing on how to evaluate agent reachability, tool access, permission boundaries, and delegation paths. The goal is to move beyond simple outcome-based evaluation and provide better evidence that multi-agent systems are evaluated in a meaningful and systematic way.

Artificial intelligence is increasingly presented as a promising tool to support aging populations, especially in contexts of cognitive decline and aging in place. However, real-world clinical practice reveals a more complex reality. Autonomy is not limited to functional independence. Cognitive disorders can alter insight, judgment, and risk perception themselves. As a result, expressed wishes may not always align with underlying needs, creating major challenges for decision-making algorithms and AI-supported care.

Drawing from clinical geriatrics and ongoing real-world technology projects involving older adults living with cognitive impairment, this talk will examine how AI systems confront human vulnerability, fluctuating cognition, and relational decision-making. Particular attention will be given to anosognosia, impaired self-awareness frequently observed in Alzheimer’s disease, as a critical but often overlooked challenge for trustworthy AI design.

Rather than focusing solely on system performance, this presentation will discuss how clinically grounded perspectives can help develop AI systems that are safe, adaptive, ethically responsible, and aligned with the lived realities of people living with dementia, care partners, healthcare teams, and support communities.

Foundation Models (FMs) are rapidly democratizing technology across many domains, and software engineering is no exception. Today, even individuals without a programming background can build functional software through natural language interactions with AI, a phenomenon called Vibe Coding. While this lowers the barrier to software creation, developing production-quality, reliable software demands something far more rigorous. This need motivates Agentic Software Engineering (Agentic SE): a discipline in which FM-powered agents plan, reason, act, and verify across the entire software engineering lifecycle, with humans maintaining meaningful oversight.

In this talk, I present our recent work along two fronts of Agentic SE. First, I introduce a dynamic benchmarking framework grounded in the Bloom’s Taxonomy that transforms static SE benchmarks into structured, multi-layered evaluations. Applied to automated program repair and bug-report-driven test case generation, the framework examines FM-powered agents across progressively demanding cognitive layers, from recalling known fixes to generalizing into entirely new project contexts.

Our findings reveal that current SE agents rely heavily on memorized patterns and suffer sharp performance drops under modest perturbations, exposing a significant gap between benchmark scores and genuine capability. Second, I share lessons learned from building agentic systems for cross-platform mobile app translation from Android to both iOS and HarmonyOS.

As AI becomes a fundamental partner in the software development lifecycle, a critical gap has emerged: while modern models excel at predicting code tokens, they remain cognitively blind to the high-level reasoning, intentionality, and memory constraints of human programmers. This cognitive misalignment results in AI models that often fail to match human judgment, leading to diminishing returns as task complexity increases. The first half of the talk focuses on decoding the developer.

By establishing a systematic methodology to model developers’ cognitive processes, Dr. Huang uncovers the complex relationship between programming and natural language processing while providing a framework to quantify human expertise for software engineering. The second half focuses on encoding this expertise to drive measurable gains in AI for SE, moving from data-centric augmentation techniques to a cognitive theory-grounded approach.

The results show that integrating human cognitive signals, such as expert attention patterns, significantly surpasses the performance of traditional cognitively-unaware AI models, with substantial improvements in tasks like code summarization and comprehension.

Software engineering demands artifacts that satisfy rigorous syntactic and semantic constraints, yet large language models are inherently probabilistic and non-deterministic. As LLMs are increasingly used to automate software engineering tasks, ensuring the reliability of the artifacts they produce becomes a critical challenge. This talk presents recent efforts toward reliable AI-based generation of software artifacts, particularly domain-specific languages and other structured representations.

It first demonstrates how current LLMs frequently violate both syntactic and semantic constraints, then presents two complementary approaches: AbsCon, a post-processing framework that leverages self-consistency to ensure provable specification compliance, and Projectional Decoding, a semantics-guided decoding approach that proactively steers generation toward valid outputs. Together, these approaches illustrate emerging directions toward LLM-based software automation that developers can trust.

Artificial Intelligence (AI) and Machine Learning (ML) systems are increasingly explored to support law-enforcement investigations, including crime linkage analysis and online child sexual abuse investigations. However, many AI-enabled systems fail in practice because user requirements, operational constraints, and human responsibilities are not sufficiently understood during system design.

This talk presents a series of collaborative studies conducted with UK law-enforcement agencies on the engineering and evaluation of AI-assisted investigative systems. First, we report findings from qualitative studies examining how analysts make decisions in practice and identify requirements for scalability, explainability, trustworthiness, adaptability, and human oversight.

Second, we describe our experience applying a Goal-Oriented Requirements Engineering framework to the design of an ML-enabled suspect identification system, highlighting the strong interdependencies between goals, data, and ML performance. Finally, we present results from an industrial evaluation of an AI-enabled crime-linkage tool, showing that analysts selectively considered AI predictions while validating outputs using traditional behavioural evidence.

Reinforcement learning has enabled the development of effective agents for critical tasks across domains such as robotic control, autonomous driving, AI coding assistance, and computer use. In this talk, I will discuss the trustworthiness challenges of RL-based agents. I will first uncover robustness issues in RL agents and show how curiosity-driven mechanisms can be leveraged to expose flaws in their decision-making. Next, I will discuss how to provide runtime verification and protection for these agents. We will also examine the hidden risks in the prevailing practice of fine-tuning agents on trajectories, including backdoor attacks and watermarking. Finally, I will share a broader outlook on the intersection of reinforcement learning and software engineering.

Cities promise smart infrastructure, real-time data, and seamless digital services. Behind the scenes, however, many municipal systems still rely on fragmented processes, siloed governance, and aging legacy technology. This keynote cuts through the hype to show why applying software engineering to society is both essential and incredibly complex.

Drawing on the Ville_IA experience, we will explore decentralized data ownership, multi-actor coordination across utilities and provincial agencies, and the messy reality that undermines assumptions about digital twins and urban AI. From incompatible datasets to projects managed by multiple independent entities, municipal digitalization is ultimately a socio-technical maze rather than a simple technical challenge.

This talk delivers an honest look at why cities can struggle to fully leverage AI, offering concrete insights on what it will actually take to move from fax-era workflows to sovereign, data-driven urban systems.

We are shifting toward an era of AI-based software engineering, where large language models and AI agents are being continuously integrated into software systems. While their integration improves software performance, these non-deterministic models also turn software systems into unexplainable black boxes. This talk reveals transparency challenges in major AI-based software systems and proposes low-cost, high-performance strategies to promote software transparency and trustworthiness.

Autonomous driving systems increasingly rely on AI-enabled perception systems based on multi-sensor fusion, yet their robustness and reliability under real-world conditions remain insufficiently understood. As these systems are deployed in increasingly complex and safety-critical environments, ensuring their dependable operation has become a fundamental challenge for both academia and industry. This talk first discusses the key robustness challenges faced by modern autonomous driving perception systems, then presents recent research on automated testing techniques for improving effectiveness and efficiency through realistic multi-modal test generation and dynamic driving scenario synthesis. The talk closes with a broader perspective on advancing robust autonomous driving systems in the era of foundation models and generative AI.

Developing scientific software capable of addressing real-world fluid dynamics problems is inherently complex, particularly for researchers who were never formally trained as software engineers. In this talk, I will share my personal journey in computational fluid dynamics software development, beginning with a blank page as a PhD student in 2009, passing through ambitious attempts to build “one-size-fits-all” simulation frameworks, and eventually arriving at a more pragmatic philosophy centered on specialized tools for specific classes of flows and applications.

Beyond the technical challenges of numerical methods and classical software maintenance, onboarding new students into rapidly evolving software ecosystems is becoming an ever more challenging task.

Foundation models are among humanity’s most complex software systems. They are currently engineered through tribal knowledge and ad-hoc processes, despite quarterly release cadences and rapidly expanding model size, context length, and modality.

This 90-minute technical briefing presents SE4FM, a disciplined software engineering framing for foundation models, organized around four pillars: DataOps, ExperimentOps, EvalOps, and FieldOps. These pillars cover data decision-making, curation, optimization, and governance; hypothesis-driven and scaling-law-aware experimentation; systematic testing, prioritization, and debugging; and release, monitoring, and compliance practices.

The session gives software engineering researchers an insider view of how foundation models are built and operated at scale, while surfacing concrete software engineering challenges and opportunities where new methods, tools, and theory are urgently needed.

Software Engineering is shifting from AI-assisted development, often described as SE 2.0, to Agentic Software Engineering, or SE 3.0, where autonomous AI agents act as collaborators. This technical briefing establishes the foundations of a disciplined SE 3.0 practice.

The tutorial synthesizes a large-scale characterization of agent activity in the wild based on the AIDev dataset, an empirical study of 567 Claude Code pull requests across 157 open-source software projects, and the Structured Agentic Software Engineering framework that reframes actors, processes, artifacts, and tools.

Attendees will leave with evidence-backed mental models, evaluation patterns for agent-authored pull requests, and a practical vocabulary, including agent command environment and agent execution environment, for designing trustworthy human-agent workflows.

AI coding agents powered by large language models are rapidly entering real-world software development, yet their reliability is still evaluated mostly through task-correctness benchmarks. This talk presents a behavior-driven fuzzing framework for systematically testing coding agents under realistic multi-step workflows. By mining real-world agent failure reports, we derive reusable interaction patterns and action types, compose them into repository-grounded fuzzing cases, and evaluate agent behavior through execution traces, file diffs, generated artifacts, and manual validation.