CMU Researcher on Long-Horizon Agents: Information Flow, User Modeling and Social World Models
niloofar_mire · x · 2026-10-21
Niloofar Mireshghallah (CMU / humans&) presented "What Goes Where?" at the Copenhagen NLP Symposium on information flow in long-horizon AI agents. Slides cover two themes: bringing humans into the loop via user simulators and the Persimmon project modeling how people respond and disclose information; and social world models for dynamic environments — asking whose beliefs changed, which goals still hold, and whether yesterday's permission is still valid today, informed by the Social World Models paper (arXiv 2509.00559). She also discussed long-horizon agentic tasks in privacy and chemistry, and is recruiting for both CMU LTI and humans&.
More from coding & agent
- Testing Jev, an RLCD decision model, to build topical maps of Anthropic's 547-page blog — cneuralnetwork · 2026-10-21
- NVIDIA's FOX blueprint brings factory-manager agents built on Cosmos and Nemotron to Europe — NVIDIAAI · 2026-10-21
- Solo dev reproduces agent failures and shrinks them to minimal cases, but can't find customers — ibackstrom · 2026-10-21
- Codex community hackathon comes to Berlin with dev, founder and GTM tracks — paw_lean · 2026-10-21
- Airbnb staff engineers show how they actually run AI agents in production — msharmas · 2026-10-21
- dsh-context: a visual dashboard for diagnosing agent context bloat, cache hits and token spend — aigclink · 2026-10-21