Humans do not learn only from what they see. From a single visual scene, we infer hidden structure and imagine alternatives: how an object might move, what another person could have seen, or how someone with different experiences might interpret the same situation. This project asks whether AI systems can learn in a similar way. Rather than relying only on more external data, I will study whether vision-language models can improve by generating and learning from counterfactual perspectives around the scene in front of them. The project will explore methods for producing alternative viewpoints, absent-agent perspectives, plausible continuations, and different value frames. These generated perspectives will be evaluated as a possible training signal for improving belief attribution, perspective-taking, and scene understanding. The central claim is that machine imagination may allow AI systems to learn more from a single event by constructing the possible perspectives, interactions, continuations, and meanings surrounding it.

This project investigates how large language models complicate traditional semantic theories of indexicals and linguistic orientation. Terms such as “I,” “here,” “now,” and “this” have historically been understood as depending upon embodied standpoint, temporal position, and perspectival access to the world. Yet contemporary language models generate and operationalize these structures despite lacking many traditionally human forms of situatedness. The paper uses artificial systems as a pressure point for reconsidering what forms of orientation are actually necessary for indexical coherence and context-sensitive meaning. The project contributes to contemporary debates in philosophy of language, mind, and semantics.

My project explores how AI systems can support epistemic values such as rigor, transparency, and intellectual honesty through formal reasoning environments. I am building an AI-assisted theorem proving system based on Lean 4 that helps users navigate mathematical proofs by retrieving relevant lemmas, suggesting proof strategies, and explaining intermediate reasoning steps. Unlike conventional language models, formal proof systems operate within verifiable logical constraints, making them a unique setting for studying how AI reasoning can remain interpretable and trustworthy. The project investigates what it means for an AI system to support not only correct reasoning, but also intellectual practices such as clarity, justification, and acknowledgment of uncertainty. More broadly, it examines how formal mathematical environments may provide a foundation for developing AI systems aligned with human epistemic values.

This project investigates an underexplored failure mode in large language models: sympathetic drift, which is the tendency of AI systems to relax moral guardrails when users provide emotionally compelling personal context, even when the requested action causes harm to third parties. In other words, the project investigates AI systems’ response when user optimization incentives conflict with socially beneficial behavior. Specifically examining if LLMs become permissive, compliant, or morally justificatory if the users frame requests around personal gain, livelihood, financial pressure, or competitive advantage. The project will develop controlled prompt environments to measure whether models can balance helpfulness against systemic safety guards among varying contexts and levels of moral ambiguity. The research will analyze the AI’s compliance metrics, refusal rates, and presence of cognitive framing to legitimize questionable user behaviors. The goal of the project is to better understand the incentive sensitivity, lie detection, and ethical robustness of various LLMs, especially as AI agents are becoming increasingly integrated into everyday and business decision-making processes. Findings may contribute to safer alignment strategies and evaluation benchmarks among increasingly agentic AI systems.