Research Areas, with Examples

Three directions of Pragmatic AI: methods that keep working in the messy real world. Each example links to the paper (journal or conference version where one exists, otherwise the preprint).

Robust AI for Language and Vision

AI, language and vision, that works under high noise and realistic human text.

• LLM that understands noisy patient textpreprint
• LLM that asks good follow-up questionsScientific Reports
• LLM that understands critical voice communicationspreprint
• Neural networks that read license plates from impossible anglespreprint
• Neural networks that detect anomalies in cyber-physical (vehicle) modelsSAE WCX
• ML that detects tunnels from weak seismic signalsUS Army report

Spec-Driven Generative AI for Language, Vision and Materials

AI that generates data according to strict specifications.

• Diffusion model that designs new molecules by their propertiespreprint
• LLM that generates text with a precise stylepreprint
• LLM that generates text with multi-aspect sentimentspreprint
• LLM that simulates students with specified knowledge gapspreprint

Multi-Agent AI without Explicit Coordination

Multiple AI agents that solve a common task without prior knowledge or coordination.

• Drones working together on a pool of targetspreprint
• LLMs evaluating and ranking each otherpreprint