Roberto Infante

Cognitive computing · decision-making in humans and machines

rdji@uw.edu · CV · GitHub · LinkedIn · Twitter

Roberto Infante

I study how humans and machines move from uncertain information to a committed decision. When does an agent know it has enough evidence to commit? I am interested in how agents make sequential decisions when they suspect that their current hypothesis space may be incomplete. I build controlled, species-fair tasks that go beyond final accuracy to compare how these agents reach an answer.



Clusterwords

Clusterwords is a controlled semantic reasoning task in which humans and machines identify groups of words linked by a common relation (NYT connections). It lets us manipulate which hypotheses or hints are available and measure how agents commit over time

Using Clusterwords, I want to study how different forms of information affect awareness, subsequent search, confidence, and commitment.

Four words. Pick the two that belong together — then say why.
Play the full game →

This is one of the tasks of clusterwords: Difficulty is manipulated along two independent axes; combinatorial (board size) and semantic (structure, margin, distractors). Boards are machine-generated to reduce overlap with any model's training data.

Through these tasks, I aim to develop new forms of diagnosis that go beyond final accuracy and provide evidence about how each system reaches its answer. To compare their behavior without assuming that differences in performance necessarily reflect differences in competence.

How can we evaluate the similarity of the process of answering this task by LLM systems and humans in a species-fair way?

Working papers

Gentags: Discrete Semantic State for Constraint-Sensitive Decision Pipelines 2025 – 2026

First-author preprint introducing Gentags, a representation that compresses source text into short, evidence-grounded semantic units used as inspectable intermediate state in LLM decision pipelines. In a controlled study isolating representation structure, Gentags raised agreement with full-evidence decisions to 79.5% (vs. 52.3–61.6% for RAKE/YAKE/TF-IDF) and hard-constraint satisfaction to 97.3% (vs. 84.7–89.3%). [PDF]

Other projects

Dynamic Information 2024

An LLM agent that builds and self-updates a knowledge/decision graph with an explicit hypothesis structure for Bayesian-style reasoning. Its limits motivated the shift to studying commitment under uncertainty directly.

Recommendation System & Smart Filters — UW MSIM Capstone (sponsor & lead) 2025 – 2026

Directed two capstone teams applying the Gentags approach to a live product, shipping a recommendation system and categorical/nudging filters now in production.

Reading

A shelf around cognition, computation, and decision-making. Drag the shelf; the laptop on the right holds the papers. Read and the next on the line.

Papers — pick one to open it on the laptop

Background

Previously I built and shipped data-driven products end to end as a technical founder. I am now moving into research in computational cognition.