Post-AGI Physical World Foundation Models: Event-Conditional Dynamics, Topology Transfer, and Risk-Limiting Guarantees

Linda Hong Cheng

Track: Track 1: Technical Foundations for a Post-AGI World

Keywords: foundation models, world models, physical decision-making, dynamical systems, time-series foundation models, event-conditioned modeling, graph neural networks, topology transfer, adapters, conformal prediction, uncertainty quantification, scalable oversight, shadow-mode evaluation, risk-limiting deployment

TL;DR: A post-AGI evaluation contract for physical-world foundation models: event-conditioned dynamics world modeling with topology transfer and shadow-mode, risk-limiting guarantees for deploy vs abstain decisions.

Abstract:

Physical decision-making is a limit case for post-AGI foundations: compound-
ing errors, distribution shift, rare events, and institutional demand for auditabil-
ity (Amodei et al., 2016). Time-series foundation models establish “pretrain then
transfer” feasibility for forecasting (Das et al., 2023; Ansari et al., 2024), but high-
stakes dynamical control adds two requirements that do not appear in benchmark-
only regimes: (i) event-conditional, topology-aware world models that represent
cross-stage propagation and delayed effects and (ii) evaluation contracts that con-
nect uncertainty to deploy or abstain decisions. We propose Industrial Dynam-
ics Foundation Models as a testable blueprint with three concrete components:
unified tokenization of multi-rate telemetry and sparse operational events; graph-
conditioned topology adapters that transfer a frozen backbone across plants; and
shadow-mode guarantees built from regime-aware conformal calibration (Shafer
& Vovk, 2008; Angelopoulos & Bates, 2021; Stankeviciute et al., 2021). Guaran-
tees become operational via a decision gate: recommend interventions only when
safety constraints hold for all trajectories in a calibrated uncertainty set, and oth-
erwise abstain. The full proposal is falsifiable with modest infrastructure: logged
telemetry, event streams, and a process graph compatible with common industrial
abstractions (OPC Foundation, n.d.; International Society of Automation, n.d.).

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Presenter: ~Linda_Hong_Cheng2

Format: Yes, the presenting author will definitely attend in person because they attending ICLR for other complementary reasons.

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Submission Number: 36