ARC-AGI

Two years, 18 solver architectures across 9 families.

The finding

The binding constraint is object-level relational structure that is not recoverable from pixel context — not compute.

18
architectures
9
distinct families
1,545
commits unique to those lines

Commit counts are those unique to each branch, measured from the common ancestor. Total branch length would count shared history many times over and flatter every approach equally.

End-to-end neural

Latent-process CNN

2024-08 – 2024-10

OpinionAI@main

Pixel-level embeddings with no downsampling, primitive selection over 64 operations, cross-attention between demonstrations and test input, and coupled shape/content heads.

What it fought

Pixel-level embeddings with no downsampling, and selection over 64 primitive operations. The base both later OpinionAI branches fork from.

CNN-only baseline

2024-08 – 2025-09 · 54 commits

OpinionAI@cnn-only

Deliberately stripped convolutional baseline, to establish what pixel-context alone can reach.

What it fought

A full codebase overhaul: content- and shape-aware encoders, a joint decoder, and shape prediction for the N-latent variant.

Vision encoder

2025-09 – 2025-10 · 152 commits

Opinion@vision

Treat the grid as an image and lean on vision-style feature extraction.

What it fought

Padding, and variable grid size. PAD removed from the vocabulary, then reweighted, then supervised; grids cropped properly; then device management and multiprocessing.

Meta-learning

Meta-learning

2024-08 – 2024-10

OpinionAI@meta-learning

Treat each task as an episode and learn an initialisation that adapts from the demonstration pairs alone.

What it fought

Episodic adaptation from the demonstration pairs alone. Shares the base history; no commits unique to the line.

Reptile meta-learning

2025-09 – 2025-10 · 7 commits

Opinion@meta-learning

Reptile-style first-order meta-learning across tasks.

What it fought

Five reformulations in seven commits: VQ-VAE, Gumbel-Softmax VQ-VAE, fewer codes, a 64-element basis, then Reptile.

RL

Reinforcement learning

2024-08 – 2024-10 · 53 commits

OpinionAI@rl

Grid construction as a sequential decision problem, with reward shaping tuned to give usable gradients rather than sparse terminal signal.

What it fought

Reward shaping. Granular reward functions, rewards retuned specifically to give usable gradients, and repeated reverts to previously working prediction and evaluation code.

Bayesian

Active inference

2024-08 – 2024-11 · 2 commits

OpinionAI@active-inference

Free-energy formulation: pick the action that minimises expected surprise over the predicted grid.

What it fought

Two commits on the line: a deeper network, and meta-learning combined with active inference. Dropped almost immediately.

Variational solver

2025-09 – 2026-03 · 214 commits

Opinion@main

A variational encoder distils the transformation rule from the demonstration pairs into a latent z; a FiLM-conditioned delta decoder applies z. Trained on an ELBO with a coupled constraint forcing demo reconstruction and test prediction to improve together.

What it fought

Converged on the variational solver: a coupled ELBO forcing demo reconstruction and test prediction down together, under cosine annealing.

Statistical inversion

2025-09 – 2026-05

Opinion@claude/arc-agi-bayesian-approach

Experiment harness treating the task as inverting a statistical generating process.

What it fought

An experiment harness for treating the task as inverting a statistical generating process. Branched from the variational line.

Relational

Hypergraph GNN + meta-learning

2024-08 – 2025-06 · 43 commits

OpinionAI@gnn-meta

Grids as hypergraphs over cells and regions, message passing with HypergraphConv, wrapped in a meta-learning outer loop.

What it fought

Getting MAML right, a distance-based loss, and predicting discrete grid heights and widths rather than regressing them.

Relational model

2025-09 – 2026-04 · 215 commits

Opinion@relational

Explicit relational structure over grid objects, with the coupled loss and D4 symmetry augmentation. Migrated back from OpinionAI.

What it fought

The loss, again. DenseOrdinalReplicaLoss with adaptive class weighting and focal loss, mask-signal weighting, a balance penalty, bounding-box predictions, and debug logging added to see what training was actually doing.

Program synthesis

SAT / hypergraph

2024-08 – 2025-06 · 35 commits

OpinionAI@sat

Constraint framing over a hypergraph encoding of the grid, searching for an assignment consistent with every demonstration.

What it fought

Hypergraph convolutions under a meta-learning loop, latent projector capacity, and penalties built around exact matches — inverse scaling and a multiplicative perfect-match penalty.

Combinator grammar (SKI)

2025-09 – 2025-10 · 4 commits

Opinion@grammar

Grammar-constrained combinator learning: select and compose formal operations from the demonstration pairs, replacing the prototype-based architecture.

What it fought

Probabilistic SKI, then robustness and device fixes. Four commits on this line before it was left.

TyleDSL

2025-06

TyleDSL@main

A typed reformulation of Michael Hodel's ARC-DSL, using the type system to prune the space of candidate programs.

What it fought

A typed reformulation of the ARC-DSL. The type system prunes candidate programs before they are ever evaluated.

Transformer

Traditional transformer

2025-09 – 2025-10 · 137 commits

Opinion@traditional-transformer

Standard encoder-decoder over serialised grids with a dense per-cell loss.

What it fought

The loss formulation, circularly. L2 replaced L1, then L1 was restored, then a dense per-cell loss replaced both.

Elastic transformer

2025-09 – 2026-01 · 270 commits

Opinion@transformer

Transformer over an ElasticConv1d front end sized to variable grid dimensions.

What it fought

Attention cost first, then stability. O(n^2) attention was replaced by an O(n) adaptive CNN, a PVR refinement loop with a verifier and refiner was added, a cosine schedule with linear warmup after that — and the line ends on gradient explosion in ElasticConv1d.

Cellular automata

Neural cellular automaton

2025-09 – 2025-11 · 159 commits

Opinion@nca

Output shape learned implicitly through CA dynamics rather than predicted: seed the test input at the centre of a 30×30 canvas, evolve, and let the automaton learn where to stop.

What it fought

The loss and the automaton's dynamics. L1 against L2, coverage switched to L1, automatic interpolation, demo-conditioned features, and perturbations to the evolving state.

LLM

LLM + LoRA

2025-09 – 2025-10 · 200 commits

Opinion@llm

Low-rank adaptation of a pretrained language model over serialised grids.

What it fought

Termination and length. EOS fallback handling, repeated token-limit and max-length increases, evaluation fixes, and eventually a dedicated diagnosis command to see why generations were failing.

Each entry is a real branch in the Opinion, OpinionAI or TyleDSL repositories. Most of these repositories are private, so branch names are given as provenance rather than links. 'What it fought' is summarised from the commits unique to that branch — it is what the history shows each approach wrestling with, not a stated conclusion. Where it finally broke, and the benchmark numbers, are still to be written up: no branch records a solve rate.