F3-AGI tests a frugal route to abstract reasoning—with a calibration caveat

An abstract grid puzzle can have several plausible explanations. A rotation, a color substitution or a rule about enclosed objects might each fit the demonstrations, yet make different predictions on the next input. More search does not automatically resolve that ambiguity: it can simply produce more convincing wrong answers.

In a preprint, Henrique Oliveira of Polybee AI Research presents F3-AGI v4.0, a CPU-based architecture that combines symbolic program search, a separate continual-learning optimizer and an interactive exploration controller. The paper reports promising results within these settings, but does not establish general intelligence or a measured reduction in electricity consumption.

Search for structure, then repair the remainder

For static ARC-AGI puzzles, the system searches combinations of grid operations, including rotations, object manipulation and cavity filling. Its registered library contains 30 operators, alongside core primitives. Candidate programs are judged using output dimensions, foreground-balanced pixel agreement, color palettes and connected objects, rather than ordinary pixel accuracy alone.

The distinction matters when most of a grid is background. A program that outputs almost nothing can achieve high pixel agreement while missing the transformation entirely. Giving foreground structure more weight makes that shortcut less attractive. This is an inductive bias, not a guarantee that the intended rule has been identified.

The search also checks the demonstration with the smallest input first. If a candidate produces the wrong output dimensions there, it cannot exactly solve every demonstration and can be rejected before further evaluation.

A third search cycle addresses near-misses. Instead of discarding a largely correct transformation, the system freezes it and examines where its predictions disagree with the demonstrated outputs. It then searches for local repair rules. Those repairs are induced from demonstrations; unseen test outputs are not available to guide them.

The important split inside the headline score

On the 400-task public ARC-AGI-1 evaluation set, the author reports 56 exact solutions with the first prediction and 57 when either of two predictions can count: 14.00% and 14.25%, respectively. Mean reported runtime was 1.62 seconds per task, with no large-language-model calls during evaluation.

However, 50 tasks had served as a development slice. Their training demonstrations informed operator engineering. The system solved 19 of those 50 tasks, or 38%, but that figure should not be read as performance on wholly uninspected problems.

With the library frozen, it solved 38 of the remaining 350 tasks using up to two predictions: 10.86%. The author describes these tasks as strictly uninspected during development. Seven were solved through the residual-repair stage. This subset provides the cleaner evidence of generalization, subject to the paper’s account of its development procedure.

The reported 42.5-fold speedup also has a specific denominator: an earlier version of the same system on the 50-task development slice. It is not a demonstrated speed advantage over contemporary competing systems. Moreover, eight of 61 exact demonstration fits did not transfer exactly to test outputs. Structural scoring reduces neither ambiguity nor overfitting to zero.

Three experiments, not one demonstrated capability

A separate experiment tests the learning optimizer on synthetic vectors: five successive two-class tasks, ten classes overall, without replaying previous examples. The optimizer skips updates on sufficiently well-predicted samples, adjusts updates using a running estimate of parameter importance and constrains them through simulated resource reserves.

It reports 91.67% final average accuracy and 9.17% forgetting, compared with 90.67% accuracy and 11.25% forgetting for Elastic Weight Consolidation in that experiment. It skipped 93.47% of updates and reduced cumulative absolute weight changes by 65.7% relative to that baseline.

These are results on a controlled vector-learning problem, not evidence that the optimizer improves ARC puzzle solving. The paper explicitly identifies its “virtual ATP” and weight-change costs as algorithmic proxies, not hardware energy measurements.

The interactive experiment is separate again. A fixed random projection maps observations into 512-dimensional representations used to encourage exploration of unfamiliar states. A pruning mechanism blocks actions recorded as having no effect and transfers rules across levels, including prior knowledge about inert distractors.

The full controller completed all 15 levels in five local games, scoring 89.44% on a level-weighted action-efficiency measure. These were ARC-AGI-3-inspired environments, not the official server-side benchmark; the result cannot be treated as an official ARC-AGI-3 score.

What remains open

The contribution is a concrete approach to bounded search: reject impossible candidates early, favor meaningful structure and repair localized errors. Its limits are equally concrete: a partly development-informed public benchmark, synthetic continual-learning data and a small local interactive suite. This preprint reports the author’s experiments, not independently established performance. Broader evaluation remains necessary before its efficiency claims can travel beyond those settings.