Principal Researcher · PhD Student

Building agents that make decisions at scale.

Hi! I'm Bryan, an AI researcher and PhD Student in Computer Science at the Federal University of Goiás (UFG), based in Goiânia, Brazil. I work at CEIA and AKCIT, where I lead interdisciplinary research teams at the intersection of Deep Reinforcement Learning and decision-making at scale.

Bryan de Oliveira
Bryan de OliveiraGoiânia · Brazil

Current work

Leading research and development of autonomous agents.

My work focuses on building agents that solve real problems — through both task-specific models trained with Deep RL and more general LLM-powered systems. The agents I'm interested in building have to deal with the uncertainty of the world and perform long-horizon tasks. At AKCIT, I lead research on two main fronts: how agents can reason and act across different abstraction levels, from language models seeking information through dialogue to embodied agents planning in physical environments; and the empirical and theoretical foundations of modern RL algorithms. I also lead a team working on autonomous humanoid robots, covering teleoperation and sim-to-real transfer. Through CEIA, I additionally lead applied R&D with industry partners across energy, finance, and software development.

01

Autonomous Humanoids

Teleoperation, policy training, digital twins, Isaac Sim, and sim-to-real transfer for autonomous humanoid robots, with industry applications (WEG).

02

Policy Gradient Methods

Theoretical and empirical study of GRPO and PPO in classical reinforcement learning environments.

arXiv
03

Embodied Agents

LLM-based embodied planning under uncertainty, with Bayesian belief-state tracking.

04

Information-Seeking Agents

Dialogue agents for optimal information-seeking under uncertainty, grounded in information theory.

arXiv
05

Small Language Models

Continued pre-training and post-training RL for offline Brazilian Portuguese LLMs in education (Positivo).

06

LLM Agents for Business

Multi-turn LLM evaluation (Conv-to-Bench), deep research agents for business documents, and RLM-based agents that build and evolve other agents (HUG Labs).

arXiv
07

LLM Post-Training

RLHF, RLAIF, and RLVR pipelines for alignment, reasoning, and tool-use in a large language model for the Brazilian energy sector (CEMIG).

08

Debt Pricing & Collection

Applied RL to orchestrate debt pricing and collection strategies at a large financial institution (Recovery / Itaú).

OpenReview
09

Marketplace Personalization

RL for recommendation systems for travel offers (Bancorbrás).

OpenReview
10

Advertising Optimization

RL and generative models for multi-channel ad generation and campaign optimization (bettrAds).

11

Offline RL for Digital Marketing

Conservative Q-Learning for personalized marketing (Acordo Certo).

Work

Research and projects.

Research

My research spans Deep Reinforcement Learning, representation learning, world models, and autonomous decision-making.

View Google Scholar profile
LLM agents · NeurIPS 2026Mapping LLM-Agent Engineering Decisions from the Literature with an Autonomous Research System Agent safety · NeurIPS 2026The Safety Decisions Behind LLM Agents: Options, Conventions, and Open Questions Embodied AI · NeurIPS 2026EvoEmbody: Self-Evolving Embodied AI Agents in Interactive Text Environments Information theory · NeurIPS 2026What Makes LLMs Effective Information Seekers? Diagnosing Question Asking and Belief Tracking Text-to-SQL · NeurIPS 2026SQL-Zero: Self-Evolving Text-to-SQL Brazilian LLMs · BRACIS 2026What Sub-Billion Models Don't Know About Brazil Language model evaluation · ICLR 2026CONV-TO-BENCH: Evaluating Language Models via User–Assistant Dialogues in Code Tasks Language model reasoning · AAAI 2026Partial Reasoning in Language Models: Search and Refinement Guided by Uncertainty Reinforcement learning · NeurIPS 2025Learning Without Critics? Revisiting GRPO in Classical Reinforcement Learning Environments Visual reinforcement learning · ICML 2025Sliding Puzzles Gym: A Scalable Benchmark for State Representation in Visual Reinforcement Learning Information-seeking · RLC 2025InfoQuest: Evaluating Multi-Turn Dialogue Agents for Open-Ended Conversations with Hidden Context Applied RL · RLC 2025Reinforcement Learning for Debt Pricing: A Case Study in Financial Services Fairness in RL · RLC 2025Personalizing Fairness: Adaptive RL with User Diversity Preference for Recommender Systems Offline RL · NeurIPS 2021PulseRL: Enabling Offline Reinforcement Learning for Digital Marketing Systems via Conservative Q-Learning

Highlighted projects

Research with code and benchmarks, and public projects on GitHub.

Harness Engineering

2026 · Harness engineering

Harness Engineering

From model to system: building reliable agent harnesses in layers — context, tools, permissions, verification, and evaluation.

Open homepage
InfoGainme

2026 · Information seeking

InfoGainme

A multi-turn framework that measures the information gain of each question, evaluating 14 models across domains and observability modes.

View on GitHub
Sliding Puzzles Gym

2025 · Benchmark

Sliding Puzzles Gym

A scalable benchmark for state representation in visual reinforcement learning, with 200,000+ sliding puzzle games.

View on GitHub
InfoQuest

2025 · Benchmark

InfoQuest

A benchmark for evaluating multi-turn dialogue agents in open-ended conversations with hidden context.

View dataset
Data-Driven Debt Pricing SLR

2023 · Systematic review

Debt Pricing SLR

A systematic literature review of machine learning, especially reinforcement learning, for data-driven debt pricing.

OpenReview
PulseRL

2021 · Offline RL

PulseRL

Offline reinforcement learning for digital marketing systems via Conservative Q-Learning.

View on GitHub
Multiagent Soccer Environment

2021 · Reinforcement learning

Multiagent Soccer Environment

A pre-compiled Soccer-Twos environment with multi-agent Gym-compatible wrappers, built on Unity ML-Agents.

View on GitHub
Cellular Automata Framework

2021 · Parallel computing

Cellular Automata Framework

A cellular automata program built with C++, OpenGL, CUDA, and OpenMP.

View on GitHub
IEEE VSSS Team

2018 · Robotics

IEEE VSSS Team

Computer vision, navigation, control, and coordination software for the Latin-American Robotics Competition.

View on GitHub

Interests & mission

What drives the work.

I'm deeply interested in the fundamental questions of artificial intelligence and neuroscience — particularly how we can develop AI systems that learn better representations and world models to effectively plan and adapt. I believe in a principled, interdisciplinary approach to AI research that combines theory with rigorous empirical validation and real-world impact.

Research interests

  • Deep Reinforcement Learning
  • Reasoning
  • Information-Seeking
  • World Models
  • Representation Learning
  • LLMs
  • Autonomous Agents
  • Bayesian Methods
  • Decision-making under uncertainty
  • Embodied AI
Just cause

To help humans and machines make better decisions, expanding human agency and accelerating the production of knowledge.

Contact

Have a problem worth solving?

I'm always open to conversations about research, collaboration, and interesting challenges.

Whether it's exchanging ideas, exploring a collaboration, or discussing a complex technical problem, feel free to get in touch.

Open contact card