Real-world grounding in agentic AI

Summary

Foundation models deployed as physical AI agents face critical risks when natural language hallucinations translate into physical law violations and operational hazards. To ensure physical consistency and reliability, researchers from the University of California, San Diego and Amazon Fulfillment Technology introduced four grounding pillars. Physics-guided deep learning embeds symmetries and differential equations into pretraining to obey conservation laws while reducing dataset requirements. The UQ4CT framework applies mixture-of-experts architectures to produce calibrated uncertainty estimates, cutting expected calibration error by over 25% and halting operations when safety thresholds are breached. Furthermore, the Adapting-While-Learning framework and formal verifiers like Hilbert integrate numerical simulators and interactive proof systems to enforce mathematical and scientific correctness.

Context

As foundation models transition from digital tasks to physical AI agents in settings like Amazon fulfillment centers, ungrounded model hallucinations risk violating physical laws, damaging equipment, or causing safety hazards. Standard language models lack inherent understanding of mathematical precision, physical conservation laws, and confidence calibration.

Approach / What changed

The author outlines four grounding approaches developed through research at UC San Diego and Amazon Fulfillment Technology: physics-guided deep learning incorporating first principles during pretraining, uncertainty-aware reasoning via the UQ4CT mixture-of-experts framework, the Adapting-While-Learning framework for dynamic simulator tool adaptation, and verifier-augmented grounding systems like Zephyrus and Hilbert using external formal verification.

Takeaways

• Physics-guided deep learning integrates first-principle physics like symmetries and differential equations during model pretraining, ensuring outputs obey physical laws while reducing required training data.
• The UQ4CT framework uses a mixture-of-experts approach to provide calibrated uncertainty estimates, reducing expected calibration error by over 25% and allowing agents to request human intervention when safety thresholds are crossed.
• Post-training with the Adapting-While-Learning framework enables dynamic tool calling of numerical simulators, yielding a 29 percent increase in answer accuracy and 12 percent better simulator tool usage over standard models.

Related reading

• Two Sigma · Jan 21, 2026

  AI in Investment Management: 2026 Outlook (Part II)

  Two Sigma leadership outlines the evolving role of artificial intelligence across quantitative investing workflows heading into 2026. The firm is embedding frontier large language models into internal systems, incident management, and feature generation pipelines to accelerate research tasks that previously took months into days. Technical focus across the broader field is shifting from raw parameter scaling toward efficiency optimizations, multimodal unified representations, and mechanistic interpretability circuits. In forecasting pipelines, rapid automated hypothesis generation introduces severe risks of overfitting and compromised backtesting, particularly when pre-trained models already contain historical regime knowledge prior to their cutoff dates. Consequently, engineering success requires strong institutional research discipline, production monitoring, and skepticism alongside the adoption of automated agentic tooling.

  Emily Majewski

• Two Sigma · Jan 12, 2026