Research
Research
My work includes an ICLR main-conference paper and four papers accepted at NeurIPS workshops, including a spotlight. It spans language-model negotiation, probability calibration, adversarial examples, uncertainty, and collective behavior. See also Google Scholar.
2026 NeurIPS · SLM-Agents workshop · Poster
Do Small Language Models Learn to Negotiate? A Controlled Scaling Study of RL-Trained Sellers
Pedro Tabacof and Sagar Joglekar.
A controlled study of model size, learning rate, and buyer-model generalization across 1,152 negotiations. Accepted as a poster at the NeurIPS 2026 Workshop on SLMs for Agentic Systems.
2020 ICLR · Main conference
Probability Calibration for Knowledge Graph Embedding Models
Pedro Tabacof and Luca Costabello.
Calibrating knowledge-graph embeddings when ground-truth negatives are unavailable.
2018 NeurIPS · LatinX in AI workshop
Adversarial Attacks on Variational Autoencoders
George Gondim-Ribeiro, Pedro Tabacof, and Eduardo Valle.
A framework for attacking and quantitatively evaluating variational autoencoders across MNIST, SVHN, and CelebA.
2017 PLOS ONE
The role of the interaction network in the emergence of diversity of behavior
Alan Godoy, Pedro Tabacof, and Fernando J. Von Zuben.
How small-world interaction networks create specialised behaviour among otherwise identical agents.
2016 NeurIPS · Bayesian Deep Learning workshop
Known Unknowns: Uncertainty Quality in Bayesian Neural Networks
Ramon Oliveira, Pedro Tabacof, and Eduardo Valle. Code.
An anomaly-detection test for uncertainty quality, plus a faster variational sampling method for Bayesian neural networks.
2016 NeurIPS · Adversarial Training workshop · Spotlight
Adversarial Images for Variational Autoencoders
Pedro Tabacof, Julia Tavares, and Eduardo Valle. Code.
A targeted attack on latent representations that makes an autoencoder reconstruct a different image.
2016 IJCNN
Exploring the Space of Adversarial Images
Pedro Tabacof and Eduardo Valle. Code.
A study and visualisation of where adversarial images appear in pixel space, from MNIST to ImageNet.
Open source
- adversarial — experiments for Exploring the Space of Adversarial Images.
- bayesian-nn-uncertainty — uncertainty experiments with Bayesian neural networks.
- rust-trees — decision trees, random forests, and causal forests in Rust.
- AmpliGraph — representation learning on knowledge graphs; I contributed to the project and calibration work.