Brown AI _🤖 Graduate Student Luncheon Series

From AI companions and mental health chatbots to automated research assistants and content moderation systems, large language models (LLMs) are increasingly deployed in high-stakes contexts that require interpretation of human communication — inferring implicit intentions, identifying emotions, and navigating plurality in possible interpretations. Yet we have little understanding of where they fail at this kind of deep interpretation or how to make their behavior in these contexts transparent to users. This talk examines both questions through four studies focused on narrative as a rich testbed for interpretation. Working with professional writers, I first show that state-of-the-art LLMs make faithfulness errors in over half of short story summaries, struggling most with subtext and implicit emotion. Building on this study, I introduce StorySumm, a benchmark revealing that no existing automatic metric reliably catches the hardest of such errors. Then, I present the Ambiguity Rewrite Metric (ARM), which addresses the inherent subjectivity of narrative evaluation by rewriting ambiguous claims with explanations, yielding substantially higher annotator agreement and more transparent model outputs. Finally, in collaboration with psychologists, I measure race and gender biases in framing and emotional content in LLM summaries of real personal narratives — raising important questions about using LLMs as tools for qualitative research. Together, these projects argue for a human-centric approach to NLP evaluation and model design that takes seriously both the complexity and diversity of the human experience.

Video generative models trained on expert demonstrations have been utilized as performant text-conditioned visual planners for solving robotic tasks. However, generalization to unseen tasks remains a challenge. Whereas improved generalization may be facilitated by leveraging learned prior knowledge from additional pre-collected offline data sources, such as web-scale video datasets, in the era of experience we aim to design agents that can continuously improve in an online manner from self-collected behaviors. In this work we thus propose the Self-Improving Loops for Visual Robotic Planning (SILVR), where an in-domain video model iteratively updates itself on self-produced trajectories, and steadily improves its performance for a specified task of interest. We apply SILVR to a diverse suite of MetaWorld tasks, as well as two manipulation tasks on a real robot arm, and find that performance improvements continuously emerge over multiple iterations for novel tasks unseen during initial in-domain video model training. We demonstrate that SILVR is robust in the absence of human-provided ground-truth reward functions or expert-quality demonstrations, and is preferable to alternate approaches that utilize online experience in terms of performance and sample efficiency.

The attention mechanism is a foundational building block of modern large language models (LLMs). However, a fundamental question in mechanistic interpretability is: why does an attention head attend to a specific token pair? This talk presents a unified, first-principles framework to address this question. We first establish a sound theoretical and empirical justification for identifying attention features, demonstrating that the singular vectors of attention matrices robustly align with feature representations. Building on this observation, we introduce ACC++, a method for extracting low-dimensional signals that are causal to attention, using only a single forward pass. We then use ACC++ to trace per-prompt circuits, showing that circuits are intrinsically prompt-specific, and that averaging over prompts obscures distinct underlying mechanisms. We demonstrate how prompt-specific circuits cluster into families, and we propose the notion of a representative circuit to analyze the differences between these families. Finally, we introduce a fully automated pipeline that leverages ACC++ signals to discover highly interpretable circuits across multiple LLMs, shedding light on algorithmic differences between prompt-specific circuits.

Jailbreak transferability is the surprising phenomenon where an adversarial prompt compromising one model also elicits harmful responses from other models. Despite widespread demonstrations, there is little consensus on why transfer is possible: is it a quirk of safety training, an artifact of model families, or a more fundamental property of representation learning? In this talk, we present evidence that transferability emerges from shared representations rather than incidental flaws. We find two factors that systematically shape transfer: (1) model representational similarity and (2) the strength of the jailbreak on the source model. Furthermore, our qualitative analyses reveal systematic transferability patterns across different types of jailbreaks. Together, these results reframe jailbreak transfer as a consequence of representation alignment rather than a fragile by-product of safety training.

Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce "emergent misalignment" that generalizes broadly. It remains unclear whether this brittleness reflects fundamental limits of alignment or merely our failure to leverage model capabilities. Here we use targeted weight pruning as a causal intervention to probe the internal organization of harmfulness in LLMs. We demonstrate that harmful content generation depends on a compact set of weights that are general across harm types and distinct from benign capabilities. Aligned models show greater compression of harm generation weights than unaligned counterparts, indicating that alignment reshapes harmful representations internally---even as safety guardrails remain brittle at the surface. This explains emergent misalignment: if weights of harmful capabilities are compressed, fine-tuning that engages these weights in one domain can trigger broad misalignment. Consistent with this, pruning harm-generation weights in a narrow domain substantially reduces emergent misalignment. Notably, harm generation and reasoning are dissociable---models can lose the ability to produce harmful content while retaining the ability to recognize and explain it. Together, these results reveal a coherent internal structure for harmfulness in LLMs that may serve as a foundation for more principled approaches to safety.