Research

Five research topics shaping AI infrastructure, power systems, and policy strategy

PAI’s research spans 14 projects in efficient LLM scheduling, energy-aware computation, data center dynamics, grid foundation models, and electricity strategy.

Open-Source Community Maintained by PAI: PowerAgent

PowerAgent is PAI’s open-source community for agentic intelligence in power systems, spanning foundation models, tool interfaces, and workflow design.

poweragent.seas.harvard.edu

Project 01

Don’t Stop Me Now: Embedding Based Scheduling for LLMs

TRAIL recycles intermediate transformer embeddings to predict remaining sequence length and preempt requests in a way that cuts latency and time-to-first-token.

Read paper

Diagram illustrating embedding-based scheduling for LLM requests.

Project 02

Fast Inference for Augmented Large Language Models

This work schedules API-augmented requests by forecasting memory demand over time and choosing request-specific KV cache handling during external calls.

Read paper

Figure showing fast inference strategies for API-augmented large language models.

Project 03

Intra-Request Branch Orchestration for Efficient LLM Reasoning

Lightweight probes on layer activations identify promising reasoning branches early so the system can stop dead ends and spend compute on higher-value paths.

View preprint

Visualization of branch orchestration for efficient LLM reasoning.

Project 04

Ongoing Work: Load Balancing in Mixture-of-Experts Systems

A plug-and-play routing algorithm adapts to gate score distributions in MoE inference to reduce load imbalance and improve throughput without retraining.

View preprint

Figure showing load balancing behavior in mixture-of-experts systems.

Project 05

TrainMover: An Interruption-Resilient and Reliable ML Training Runtime

TrainMover keeps training productive through interruptions with standby machines, delta-based communication, and shadow iterations that enable second-level recovery.

View preprint

Diagram of the TrainMover runtime for interruption-resilient machine learning training.

Project 06

Power-Aware MoE Inference Framework

This ongoing project aligns Mixture-of-Experts inference settings with grid supply by co-tuning software and hardware configurations for power-aware operation.

Ongoing research

System diagram for a power-aware mixture-of-experts inference framework.

Project 07

Data Center Control Against Sub-Synchronous Resonance: A Data-Driven Approach

This project studies whether grid-connected data centers can trigger sub-synchronous resonance and develops data-driven control around power-factor-correction converters.

View preprint

Figure illustrating data-driven control against sub-synchronous resonance in data centers.

Project 08

LILAD: Learning In-context Lyapunov-stable Adaptive Dynamics Models

LILAD learns adaptive dynamics models and Lyapunov certificates together so new systems can be identified quickly while preserving stability guarantees.

View preprint

Visualization of LILAD adaptive dynamics modeling with Lyapunov stability guarantees.

Project 09

Crucial Role of Foundation Models in Enhancing the Interaction of AI and Power Systems: Achieving Integrated Frameworks

This work frames foundation models as both a new source of grid stress and a tool for smarter energy management across data centers, operations, and markets.

Read article

Integrated framework linking foundation models, data centers, and electric power systems.

Project 10

PowerAgent: A Road Map Toward Agentic Intelligence in Power Systems: Foundation Model, Model Context Protocol, and Workflow

PowerAgent outlines a roadmap for context-aware AI assistants in power systems built on foundation models, standardized tool interfaces, and structured workflows.

Read article

PowerAgent roadmap for agentic intelligence in power systems.

Project 11

PowerMamba: A Deep State Space Model for Time Series Prediction in Power Systems

PowerMamba combines state-space modeling and deep learning to forecast multivariate power-system time series while incorporating high-resolution external forecasts.

Read article

PowerMamba figure showing deep state space modeling for power system time series prediction.

Project 12

Unlocking Multi-Task Electric Energy System Intelligence: Data Scaling Laws and Performance with Limited Fine-Tuning

This study examines whether data scaling laws can produce multi-task, cross-timescale foundation models for power systems that generalize to unseen operations.

View preprint

Figure on scaling laws and limited fine-tuning for multi-task electric energy system intelligence.

Project 13

AI Data Centers, and the U.S. Electric Grid: A Watershed Moment

This policy brief analyzes the surge in U.S. data-center electricity demand and identifies engineering and regulatory strategies for more flexible, equitable grid expansion.

Read policy brief

Policy brief figure on AI data centers and the U.S. electric grid.

Project 14

Flexibility-Aware Framework for Efficient Planner-Initiated Siting of Data Center

This preprint introduces a planner-initiated siting workflow that screens reliable locations, evaluates market impacts under standardized flexibility envelopes, and ranks pre-certified interconnection-ready sites to accelerate large data center deployment while preserving grid reliability and market stability.

View preprint

Figure showing a flexibility-aware framework for planner-initiated siting of data centers.