ARC-JEPA v1

Object-centric transformation JEPA with latent-guided program search for ARC-AGI-2 (code: https://github.com/acco-cyber/ARC-AGI-2/tree/main/arc-jepa). The package is the input of the solver: a hierarchical transformation-JEPA (cell/object/relation encoders, EMA target, predictor, rule latent, program encoder, scorer) plus a transformation memory, used as a prior for a typed-DSL program search with exact demonstration verification.

Measured results (every number is measured, none is projected)

version	date	public eval (120 tasks, competition metric)	val 150 (training-distribution split)	params	notes
v1	2026-09-27	0.0083 (1/172 outputs; 0 tasks with an exact-fit program; 48 near misses >= 90% cells)	0.12 (18/150, symbolic search at 20 s/task, model=None; same code family)	0.56M (debug config)	Smoke-trained debug model (0.56M params, 20 steps per stage) as the prior of the 72-op typed DSL search. Pipeline validation release: measured 0.0083 on the 120

"public eval" is the official 120-task ARC-AGI-2 evaluation set, never used for training or tuning. "val 150" is the held-out 150-task split of the 1,000 training tasks (easier distribution). The Kaggle leaderboard score, when available, is reported in the GitHub release notes.

This version

Smoke-trained debug model (0.56M params, 20 steps per stage) as the prior of the 72-op typed DSL search. Pipeline validation release: measured 0.0083 on the 120-task public eval. Not a competitive model.

Files

model.safetensors (weights), config.json (model + search config), vocab.json (program tokenizer), program_memory.npz + programs.json (transformation memory: synthetic + training tasks only), metrics.json.

Load

from arcjepa.model.arcjepa import ARCJEPA
model = ARCJEPA.load_package("path/to/this/repo", device="cuda")