Research and engineering
The foundations of intelligence.
Five connected areas for investigating how AI can be understood, built and controlled.
This agenda presents the questions and themes that guide Polybee’s work.
Publications and artifacts
From a research agenda to results that can be examined.
F3-AGI v4.0
Abstract reasoning with neuro-symbolic architecture, active inference and CPU execution.
tasks solved in local evaluation
on public ARC-AGI-1 · pass@2
Area 01
Agentic systems
How can we coordinate agents, memory and tools in systems that can be evaluated and controlled?
Agentic systems combine models, memory and tools to carry out sequences of actions. We investigate how to organize these parts, observe their behavior, and maintain evaluation and control criteria throughout execution.
Research themes
- Harnesses and orchestration
- Execution and coordination environments that connect models, tools and work stages.
- Memory and tools
- Ways to preserve context and make the use of information and external resources explicit.
- Multi-agent architectures
- Distributing tasks among agents with clear responsibilities.
- Specification-driven development
- Building from specifications that make goals and acceptance criteria verifiable.
Connections across the agenda
Agents depend on models, retrievable knowledge and evaluation. Their execution also raises questions about infrastructure and autonomy.
Area 02
Models & intelligence
How can we adapt models to specific contexts and treat Portuguese as a first-class language?
Studying models means understanding their possibilities and limits in different contexts. Our agenda includes adaptation, specialization and efficiency, with attention to Portuguese, voice and the combination of information modalities.
Research themes
- Fine-tuning
- Adapting models to specific data, tasks and contexts.
- Distillation and SLMs
- Knowledge distillation and smaller models for different operating constraints.
- Specialized models and multimodality
- Connections among language, voice, images and other kinds of information.
- Portuguese from the outset
- Data, methods and evaluations that place the language and its contexts at the center of research.
Connections across the agenda
Model research connects to data quality, behavioral evaluation and real computing conditions.
Area 03
Data & knowledge
How can we organize data and knowledge so AI systems can retrieve and use them with traceability?
Data must be organized to support the creation, retrieval and use of knowledge. Our agenda brings together methods and structures that connect information, preserve context and track its origin in systems.
Research themes
- Datasets and pipelines
- Preparing, organizing and transforming information.
- Retrieval
- Finding the knowledge relevant to a task.
- Knowledge graphs and ontologies
- Representing concepts and their relationships.
- Memory and knowledge infrastructure
- Maintaining, updating and making knowledge available to models and agents.
Connections across the agenda
Retrievable knowledge brings models and agents closer to their contexts. Traceability helps us observe and evaluate how it is used.
Area 04
Evaluation & governance
How can we measure behavior, identify limits and establish reliability criteria for AI systems?
Evaluation must follow the whole system, beyond a single answer. We investigate ways to observe behavior, recognize limits and make development and operating criteria explicit.
Research themes
- Evals
- Examining behavior and limits under defined tasks and conditions.
- Observability
- Understanding what happens during execution.
- Traceability and safety
- Investigating paths, dependencies and risks behind outcomes.
- Policies and reliability criteria
- Explicit rules to guide decisions about the use of AI systems.
Connections across the agenda
Evaluation spans the whole agenda: models, data, agents and infrastructure must be understood under real operating conditions.
Area 05
Compute, edge & robotics
How can we bring intelligence to different environments while accounting for latency, connectivity, resources and control?
AI must work wherever it is needed. We investigate local inference, distributed infrastructure and intelligent systems beyond centralized cloud services.
Research themes
- Local inference
- Running models close to the people, devices or processes that need them.
- Hybrid infrastructure
- Combining cloud, local clusters and private environments.
- Embedded systems
- Operating within hardware and connectivity constraints.
- Robotics
- Models and systems for machines that perceive and act.
Connections across the agenda
The choice of computing environment shapes autonomy, cost, latency, privacy and the kinds of applications we can build.
Openness and democratization run through the entire agenda.
Shared knowledge, tools and methods can widen participation in building intelligence. Our open-core philosophy connects that openness to sustainable research.
Our open-core philosophy