NSF Institute for Trustworthy AI in Law & Society (TRAILS)

Strong machine-translation benchmark scores do not by themselves show whether translation tools help people communicate. Marine Carpuat presents human-centered research on how nonexpert users understand and rely on machine translation, and draws implications for evaluation and system design. The talk then considers open research questions about making successful communication a central objective of large-language-model research and development.

Autonomous robots and other large dynamical systems often have complex, uncertain or partly unknown models that limit traditional model-based control. Machine learning and reinforcement learning offer adaptable decision-making methods, but their interpretability and guarantees of safe, correct behavior remain limited. This talk brings formal methods together with control theory and learning: expressive specification languages capture objectives and safety requirements, while verification, synthesis and proofs establish desired behavior. Examples from autonomous driving and robotic manipulation illustrate controllers designed to scale and adapt while remaining safe and interpretable.

Multimodal large language models perform well on many visual and textual reasoning tasks, but assistive applications expose concerns about privacy and reliable real-world evaluation. This talk examines document understanding and spatial reasoning in three-dimensional environments. An agentic privacy-protection method masks personal information selectively while preserving visual context needed for question answering. SceneCritic, a symbolic evaluator, applies explicit spatial heuristics to test the semantic and geometric consistency of generated indoor layouts. Together, these projects investigate how multimodal models can be useful despite their limitations as reasoners and evaluators.