ARC-IT: Rule-Conditioned Transformer for ARC-AGI

A novel architecture that solves abstract reasoning tasks (ARC-AGI) by explicitly extracting transformation rules from demonstration pairs and applying them to new inputs:

• GridTokenizer -- Embeds discrete ARC grids (0-11) into continuous patch tokens
• RuleEncoder -- Extracts transformation rules from demo input/output pairs via cross-attention
• RuleApplier -- Applies the learned rules to a test input via cross-attention
• SpatialDecoder -- Converts output tokens to 64x64 grid logits

Architecture

Demo Pairs -> GridTokenizer -> RuleEncoder (cross-attention + aggregation) -> Rule Tokens
Test Input  -> GridTokenizer -> RuleApplier (cross-attention to rules) -> SpatialDecoder -> Predicted Grid

Training

• 2-stage training: Full Training -> Hard Focus (AGI-2 oversampling)
• Test-Time Training (TTT): Per-task fine-tuning on demonstration examples

Model Details

• Training step: 18000
• Best validation accuracy: 0.733029360572497
• Hidden size: 384
• Rule Encoder: 2 pair layers, 2 agg layers, 64 rule tokens
• Rule Applier: 4 layers, 8 heads
• Canvas size: 64

Usage

import torch
from arc_it.models.arc_it_model import ARCITModel

model = ARCITModel.from_config(config)
ckpt = torch.load("model.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["model_state_dict"])

Links

• Repository: github.com/REDDITARUN/arc_it
• ARC-AGI: arcprize.org