The ARC-AGI Without Pretraining author claims ARC-AGI tasks can be solved competitively without pretrained foundation-model knowledge, challenging pretraining as a prerequisite for abstract reasoning.
state: expiredheat: lowuncertainty: highconvergesscott: mediumarc-agi reasoning-research
What is this?
CompressARC is a 76K-parameter neural model that starts from random initialization and trains separately on each target ARC-AGI puzzle at inference time, using minimum description length rather than pretrained knowledge or an external dataset. The supplied research post and arXiv snippet report scores of 34.75% on the training set and 20% on the evaluation set, with roughly 20 minutes of processing per puzzle on an RTX 4070. The author, identified in the supplied sources only by the handle “iliao2345,” presents this as evidence that abstract generalization on ARC-AGI can emerge from compression-driven test-time learning without foundation-model pretraining; the snippets do not independently validate the claimed competitiveness or broader implications.
Why it matters to Scott
CompressARC extends Scott’s inference-time-scaling position from structured search and verification into per-task weight adaptation: useful reasoning may be assembled during inference rather than residing entirely in pretrained priors. If independently validated, its tiny randomly initialized model and compression objective could inform Scott’s reasoning-system and local-inference experiments, but the supplied evidence is still a single author’s unverified benchmark claim; the radar tracks adjacent ARC-AGI and test-time-learning work, not this same development.
ip:concept.inference-time-scalingip:source.the-hidden-architecture-of-better-ai-reasoningdev:concept.serial-intelligence-loopdev:concept.hardware-aware-local-inferenceradar:concept.arc-agiradar:ttt-discover-test-time-learningradar:littlelearner-pretraining-capability-ceilingradar:pathway-recurrent-arc-efficiency
queries asked of Scott's wikis
- test-time training versus pretrained reasoning
- compression and minimum description length as intelligence
- small task-specific models versus foundation models
- ARC-AGI and abstract generalization benchmarks
- inference-time learning and model adaptation
- reasoning without retrieval or prior knowledge
Measured heat
no measured readings yet — the hourly heat pass fills this in
How the heat travelled
no chain yet — the hourly chain pass fills this in
Evidence (2) — ⭐ canonical anchor
Interpretation history
2026-09-01T20:54:33Z
Re-evaluation found no new validation, implementation, or discussion beyond the historical authors’ claim. With the ARC-AGI neighborhood cool and no confirming event expected, this no longer merits active monitoring.
2026-09-01T20:45:00Z
grounded: converges/medium — CompressARC extends Scott’s inference-time-scaling position from structured search and verification into per-task weight adaptation: useful reasoning may be ass
2026-09-01T20:41:34Z
origin walked (codex/luna, conf 0.99): anchor hn.story.49527531 -> echo.blog.d91c77df19 by Isaac Liao and Albert Gu
2026-09-01T20:40:19Z
case created — The linked original research write-up advances a bounded claim about solving ARC-AGI without pretrained knowledge.
Decision trace
- 09-02 06:54expireRe-evaluation found no new validation, implementation, or discussion beyond the historical authors’ claim. With the ARC-AGI neighborhood cool and no confirming event expected, this no longer merits ac
- 09-02 06:54alert_silentThe only delta is an administrative re-evaluation with unchanged evidence and engagement; there is nothing consequential to deliver before a normal briefing.
- 09-02 06:54alert_routeThe only delta is an administrative re-evaluation with unchanged evidence and engagement; there is nothing consequential to deliver before a normal briefing.
- 09-02 06:50alert_silentThe original author post establishes that CompressARC and its reported no-pretraining ARC-AGI results exist, but this is a resurfacing of 2025 work rather than a new release or validation event. Its 2
- 09-02 06:50surface_candidateThe original author post establishes that CompressARC and its reported no-pretraining ARC-AGI results exist, but this is a resurfacing of 2025 work rather than a new release or validation event. Its 2
- 09-02 06:50alert_routeThe original author post establishes that CompressARC and its reported no-pretraining ARC-AGI results exist, but this is a resurfacing of 2025 work rather than a new release or validation event. Its 2
- 09-02 06:45groundCompressARC extends Scott’s inference-time-scaling position from structured search and verification into per-task weight adaptation: useful reasoning may be assembled during inference rather than resi
- 09-02 06:41promote_anchororigin walk conf 0.99
- 09-02 06:40createThe linked original research write-up advances a bounded claim about solving ARC-AGI without pretrained knowledge.