← back

Price

—

Travel

—

Access

—

Social

—

ABSTRACT: In this talk, I will discuss 3 major pitfalls and challenges of current state-of-the-art AI agents, and present potential solutions for: (1) Teaching agents to be trustworthy and reliable collaborators based on: social/pragmatic multi-agent interactions via speaker-listener confidence calibration, learning to balance positive and negative persuasion, and multi-agent AI safety through the lens of compositional attacks, theory-of-mind, and belief-steering; (2) Discovering and improving skills/world models needed for efficient, robust action and collaboration based on: learning programmatic skills, weakness-driven adaptive data/environment generation for skill improvement, and structured, selective world model discovery and inference; (3) Planning of long-horizon memory for multi-step reasoning and generation over continuously evolving, conflicting, and scattered information. We will cover diverse domains (math, commonsense, coding, tool-use, computer use, etc.), modalities (text, images, videos, audio, layouts, etc.), and real-world applications (early medical diagnosis and classroom education engagement).

• Other