Researches | Pablo Piantanida

Collaborative & Resource-Aware Agentic Intelligence

Mon, 01 Jan 0001 00:00:00 +0000

My research investigates how agents with different information can communicate and cooperate. It connects information-theoretic foundations of distributed inference with pragmatic models of retrieval and dialogue.

A central theme is that communication should be understood through its contribution to a shared task: what an agent communicates, what another agent can infer, and how the exchange changes subsequent decisions.

Key questions

What information should agents exchange to solve a shared task?
How can communication account for another agent’s knowledge and beliefs?
What are the benefits and limits of interaction under communication constraints?

Selected contributions

EMNLP · 2025

Collaborative Rational Speech Act: Pragmatic Reasoning for Multi-Turn Dialog

Studying collaborative dialogue between agents with private information.

Paper ↗ arXiv ↗ Code ↗

SIGIR · 2025

Rational Retrieval Acts: Leveraging Pragmatic Reasoning to Improve Sparse Retrieval

Applying pragmatic reasoning to the information conveyed by retrieval terms.

Paper ↗ arXiv ↗ Code ↗ Poster ↗

IEEE Transactions on Information Theory · 2019

Collaborative Information Bottleneck

Studying trade-offs between communication complexity and relevant information in a cooperative setting.

Paper ↗ arXiv ↗

IEEE Transactions on Information Theory · 2017

The Three-Terminal Interactive Lossy Source Coding Problem

Investigating fundamental limits of interactive communication for reconstructing distributed sources.

Paper ↗ arXiv ↗

Ongoing research

Building on these contributions, I am interested in how collaborative systems should allocate information, computation, and communication to support reliable decisions under limited resources.

Mathematical Foundations of Learning, Representation & Reasoning

Mon, 01 Jan 0001 00:00:00 +0000

My research investigates the mathematical principles underlying learning and representation. Drawing on information theory and statistical decision theory, I study how representations retain task-relevant information, how their usefulness can be compared, and how knowledge can be transferred across models.

The objective is to connect the design of learning systems with precise questions about information, sufficiency, and downstream decisions.

Key questions

What makes one representation more useful than another?
How can information-theoretic principles guide learning objectives?
How can multiple models contribute to reusable representations?

Selected contributions

NeurIPS · 2024

When Is an Embedding Model More Promising Than Another?

Comparing representations through their potential usefulness across downstream tasks.

Paper ↗ arXiv ↗ Code ↗ Poster & presentation ↗

NeurIPS · 2025

Learning Task-Agnostic Representations through Multi-Teacher Distillation

Investigating how knowledge from multiple teachers can support representations that remain useful across tasks.

Paper ↗ arXiv ↗ Code: NLP ↗ Code: molecules ↗ Code: vision ↗ Poster & presentation ↗

Machine Learning · 2023

The Role of Mutual Information in Variational Classifiers

Examining the role of mutual information in classification objectives and learned predictive representations.

Paper ↗ arXiv ↗

ICML · 2022

A Differential Entropy Estimator for Training Neural Networks

Connecting continuous information measures with objectives used to train neural networks.

Paper ↗ arXiv ↗ Code ↗ Poster ↗

Uncertainty, Evaluation & Reliable Decision-Making

Mon, 01 Jan 0001 00:00:00 +0000

My research studies how intelligent systems can recognize uncertainty, identify potential failures, and support decisions about when to act, reject a prediction, or seek additional evidence.

This work connects statistical measures of uncertainty with practical evaluation methods. It spans classification and regression, generated-content assessment, and the privacy risks associated with trained models.

Our work on prediction errors leads to a practical question for auditors: what evidence should they collect next under a limited testing budget? Our position paper on language-model auditing examines this question in changing deployment contexts.