Polybee AI Research · Publication 01
F3-AGI v4.0
Sample-efficient abstract reasoning with neuro-symbolic architecture and active inference.
Investigating how to learn from few examples, synthesize programs and explore environments under computational constraints.
FISHER
FRUGALITY
This preprint has not undergone peer review. The code is public, with a software-artifact DOI on Zenodo. Publication of the manuscript on arXiv is pending endorsement. These are local research evaluations, not official ARC Prize scores.
01 / Scientific summary
Fewer examples.
More structure for reasoning.
F3-AGI v4.0 investigates the induction of algorithmic transformations from a few demonstrations, combining program synthesis, continual learning and active inference in a CPU-first architecture.
In abstract reasoning tasks, different programs can explain the same examples yet fail on new inputs. This work addresses that multiplicity of hypotheses through structural and complexity constraints, local error repair and controlled exploration.
Three mechanisms are proposed: F3-Metabolic-Optim, combining mistake-gated plasticity, Fisher-EMA preconditioning and virtual ATP gating; AACR with residual patch synthesis, searching for compositions in a 30-operator library; and a normalized 512-dimensional random projection, used to guide exploration and prune actions with no effect.
The manuscript presents three separate empirical studies: a synthetic continual-learning stream, the public ARC-AGI-1 evaluation set and a local interactive suite inspired by ARC-AGI-3. The contribution is an architecture and its evaluation in these settings; the work does not demonstrate artificial general intelligence.
02 / Empirical evidence
Three studies.
Explicitly distinct scopes.
Results are reported in the preprint and repository logs. This institutional page does not constitute an independent reproduction of the experiments.
ARC-AGI-1 · local public-set evaluation
57 exact solutions across 400 tasks.
| Subset | pass@1 | pass@2 | Interpretation |
|---|---|---|---|
| Calibration · 50 tasks | 19/50 · 38.00% | 19/50 · 38.00% | Demonstrations informed library development. |
| Uninspected · 350 tasks | 37/350 · 10.57% | 38/350 · 10.86% | Generalization subset reported by the author. |
| Total · 400 tasks | 56/400 · 14.00% | 57/400 · 14.25% | Combines the two subsets above. |
How to read these results: pass@1 considers the first attempt; pass@2 accepts an exact solution in either attempt. The 50-task subset informed operator design, so its 38% result does not measure generalization without prior inspection.
- Observed total runtime
- 647.5 s
- Mean across 400 tasks
- 1.62 s/task
- Autonomous patches in the uninspected subset
- 7 tasks
Sources: preprint, Table 4 · 400-task evaluation logs ↗
Continual learning · synthetic stream
Selective plasticity with task retention.
A stream of five sequential tasks, two classes per task, 64-dimensional vectors and 6,000 online steps, without replay. The experiment isolates memory stability and update cost; it does not evaluate learning on ARC grids.
The M₁ plasticity cost was 2,736.81, compared with 7,986.85 for EWC: a 65.7% reduction in this proxy. M₁ represents the L₁ sum of weight changes, not measured electricity consumption.
Sources: preprint, Table 2 · ablation results and configuration ↗
Interactive exploration · local suite
15 levels completed across five games.
Local 64 × 64 grid environments, with three levels per game and no natural-language instructions. The suite is inspired by ARC-AGI-3; it is neither the official benchmark nor a competition-scored submission.
The efficiency metric weights levels by 1, 2 and 3. The unweighted average was 90.72%. In the ablation, active inference without dead-signature pruning reached 46.24% on the weighted metric, with 32 dead actions.
Sources: preprint, Tables 6–7 · results by agent and game ↗
03 / Methodology
Constraints as part
of system design.
Five levels. One architecture.
- Level 4Consolidation and global workspace
NREM memory, operator library and rule transfer.
- Level 3Synthesis, residual repair and exploration
AACR, ResidualPatchSynthesizer and normalized 512D random projection.
- Level 2Metaplastic memory
Fisher-EMA preconditioning and anchor regularization.
- Level 1.5Bioenergetic control and complexity
Mistake gating, virtual ATP budget and MDL criterion.
- Level 1Multiscale perception
Topological invariant extraction and scene representation.
Institutional overview of the architecture described in the manuscript. Components are evaluated in distinct studies.
Program synthesis
Invariants extracted from demonstrations guide a search through a parametric library. Checking the smallest pair enables early rejection of incompatible candidates; evaluation uses the available demonstration pairs.
Residual repair
Cycle-3 searches for local patches when a global transformation is close to the solution. In the 350-task uninspected subset, seven solutions were attributed to this mechanism.
Gated plasticity
F3-Metabolic-Optim combines selective mistake-driven updates, Fisher-EMA preconditioning and a virtual bioenergetic budget to investigate the trade-off among adaptation, retention and plasticity cost.
Exploration with memory
A fixed, normalized 512D random projection supports the state-coverage measure. DeadSignatureTracker records actions with no effect to avoid repeating them.
Reported experimental conditions
Local execution on a laptop with 16 GB of RAM, a CPU-first NumPy engine and a reported footprint below 400 MB. Seeds and configurations are documented in scripts and logs. Runtimes depend on hardware and execution conditions; this summary provides no confidence intervals.
04 / Scope and limitations
What these results
support.
- Calibration and generalization are different subsets.
Ten operators were developed after inspecting demonstrations in the first 50 tasks. The 38/350 result on the uninspected subset is the study’s reported generalization measure.
- Public-set evaluation is not an official competition score.
The 400 tasks belong to the public ARC-AGI-1 corpus and were evaluated locally. The interactive suite is also local. These are neither private test sets nor official ARC Prize results.
- The experiments do not cover every domain of intelligence.
The continual-learning experiment uses synthetic data. The 512D module is an analytical random projection, not an end-to-end trained JEPA encoder. AGI is not demonstrated.
- Virtual ATP and M₁ are cost models.
These mechanisms are not measurements of energy in joules or device electricity consumption. Bioinspired efficiency and physical energy savings must be distinguished.
- Open evaluation questions remain.
The second attempt added only one exact solution across the full set. Independent evaluations, repeated runs and tests on new datasets are needed to assess robustness and transfer.
Based on Section 6 of the manuscript. The roadmap includes ARC-AGI-2 and hybrid Wake-Sleep consolidation; these steps are not presented as completed results.
05 / Open science
Read. Examine.
Reproduce.
Download the PDF (697 KB) · ORCID 0009-0004-2730-292X ↗
Reproducing the studies
The repository README describes installation, dependencies and execution commands. The software is distributed under the MIT License. Check the terms of each dataset and other materials before reusing them.
| Study | Entry point | Public evidence |
|---|---|---|
| Continual learning | f3_metabolic_optim.py | Configuration and results ↗ |
| Public ARC-AGI-1 | kaggle_submission/f3_agi_v4_kaggle_submission.py | 400-task report ↗ |
| Local interactive suite | f3_agi_v4_arc3_runner.py --ablation | Four-agent ablation ↗ |
Kaggle notebooks
Explore the three studies as notebooks, with direct access on Kaggle.
Original manuscript title
F3-AGI v4.0: Bioenergetically-Gated Active Inference, Metaplastic Preconditioning, and Residual Program Synthesis for Sample-Efficient Abstract Reasoning
06 / Citation
How to cite this work.
Use the manuscript reference when discussing the paper and the software reference to identify the version executed. The DOI below identifies the software artifact; it is not an arXiv identifier for the manuscript.
Oliveira, H. (2026). F3-AGI v4.0: Bioenergetically-Gated Active Inference, Metaplastic Preconditioning, and Residual Program Synthesis for Sample-Efficient Abstract Reasoning. Polybee AI Research Laboratory. Preprint, October 29, 2026.
@misc{oliveira2026f3agi_preprint,
author = {Henrique Oliveira},
title = {F3-AGI v4.0: Bioenergetically-Gated Active Inference,
Metaplastic Preconditioning, and Residual Program
Synthesis for Sample-Efficient Abstract Reasoning},
year = {2026},
month = oct,
howpublished = {Polybee AI Research Laboratory, preprint},
url = {https://polybee.ai/research/f3-agi-v4/},
note = {Manuscript dated 2026-10-29. Not peer reviewed.}
}Oliveira, H. (2026). F3-AGI v4.0: CPU-First Neuro-Symbolic Reasoning (v4.0.0) [Software]. Zenodo. 10.5281/zenodo.23816665.
@software{oliveira2026f3agi_software,
author = {Henrique Oliveira},
title = {F3-AGI v4.0: CPU-First Neuro-Symbolic Reasoning},
year = {2026},
version = {4.0.0},
publisher = {Zenodo},
doi = {10.5281/zenodo.23816665},
url = {https://github.com/henrique-polybee/f3-agi-v4}
}Research in dialogue
Reproduce, question
and build together.
We welcome technical feedback, reproduction reports and collaboration proposals. Use the repository to discuss code and experiments. For research partnerships, contact Polybee directly.