Avatar

Aldo Pacchiano

Assistant Professor / Visiting Scientist

Boston University / Broad Institute of MIT and Harvard

I am an Assistant Professor at the Boston University Center for Computing and Data Sciences and a Visiting Scientist at the Schmidt Center of the Broad Institute of MIT and Harvard. I obtained my PhD at UC Berkeley where I was advised by Peter Bartlett and Michael Jordan. My research leverages the principles of sequential decision making to design intelligent and adaptive systems. I aim to advance our statistical understanding of learning phenomena in adaptive environments and translate these insights into the design of AI systems for autonomous discovery in large scale scientific and societal domains.

The Pacchiano’s Lab for Adaptive and Intelligent Algorithms (PLAIA) site can be found here.

A sample of my literary writings, including short stories and notes in both English and Spanish, can be found here.

Interests

  • LLMs and Discovery
  • Reinforcement Learning
  • AI for Science
  • Theory of Sequential Decision-Making

Education

  • PhD in Computer Science, 2021

    University of California Berkeley

  • MEng in Computer Science, 2014

    Massachusetts Institute of Technology

  • Masters of Advance Study in Pure Mathematics, 2013

    Cambridge University

  • Bachelors of Science in Computer Science and Theoretical Mathematics, 2012

    Massachusetts Institute of Technology

Recent Posts

Autonomous Discovery from Data

Sequential decision making algorithms are history dependent policies. Modern sequence prediction models such as transformer …

The Dissimilarity Dimension

For a time the Eluder dimension has been used to provide bounds for optimistic algorithms in function approximation regimes. We …

Neural Optimism for Genetic Perturbation Experiments

This work provides a theoretically sound framework for iteratively exploring the space of perturbations in pooled batches in order to …

Model Selection for Contextual Bandits and Reinforcement Learning

In the problem of model selection the objective is to design ways to select in online fashion the best suitable algorithm to solve a …

Beyond the Standard Assumptions in Reinforcement Learning

In Reinforcement Learning it is standard to assume the reward to be an additive function of per state feedback. In this work we …

Publications

Language Model Personalization via Reward Factorization

COLM 2025. Also presented as an oral at the 2nd Workshop on Test-Time Adaptation Putting Updates to the Test (PUT) 2025 at ICML and at …

Multiple-policy Evaluation via Density Estimation

ICML 2025. Also presented at the Foundations of Reinforcement Learning and Control – Connections and Perspectives Workshop, ICML …

A Theoretical Framework for Partially Observed Reward-States in RLHF

ICLR 2025. Also presented at the Aligning Reinforcement Learning Experimentalists and Theorists Workshop and the Workshop on Models of …

Learning Rate-Free Reinforcement Learning: A Case for Model Selection with Non-Stationary Objectives

Presented at the Failure Modes of Sequential Decision-Making in Practice Workshop, RLC 2024.

Provably Sample Efficient RLHF via Active Preference Optimization

Presented at the Theoretical Foundations of Foundation Models Workshop, ICML 2024.

Experiment Planning with Function Approximation

NeurIPS 2023, also presented at the PAC-Bayes Meets Interactive Learning Workshop, ICML 2023.

Anytime Model Selection in Linear Bandits

NeurIPS 2023, also presented at the PAC-Bayes Meets Interactive Learning Workshop, ICML 2023.

Supervised Pretraining Can Learn In-Context Reinforcement Learning

NeurIPS 2023, also presented at the New Frontiers in Learning, Control, and Dynamical Systems Workshop, ICML 2023.

Transfer RL via the Undo Maps Formalism

Presented at the New Frontiers in Learning, Control, and Dynamical Systems Workshop, ICML 2023.

Meta Learning MDPs with Linear Transition Models

AISTATS 2022; also presented in the Workshop on Reinforcement Learning Theory, ICML 2021.

On the Theory of Reinforcement Learning with Once-per-Episode Feedback

NeurIPS 2021; also presented as an oral talk in the Workshop on Reinforcement Learning Theory, ICML 2021.