No-neural-network chat assistant CAEL solves 62/400 ARC-AGI tasks on CPU with handwritten rules
Silver_Employ2617 · reddit · 2026-11-21
A developer demos CAEL, a CPU-only chat assistant with no neural network — just rules and retrieval — and shares a 40-line numerical checker plus open (noncommercial) source.
- The checker searches for an arithmetic path from the question's numbers to the proposed answer, down-weighting solutions that leave numbers unused; the hard part is distinguishing "skipped a needed quantity" from "correctly ignored a distractor" (e.g. "paid at till 6")
- CAEL runs several handwritten solvers over different readings of the conversation; each check gets one vote, and conflicting passing interpretations trigger a clarification request
- Main limitation: solvers are handwritten, so unfamiliar questions often get refusals
- ARC-AGI-1 experiment: 53 handwritten rule families must reproduce examples before voting; with 8 rotations/reflections and a held-out example check, it solved 62 of 400 evaluation tasks at pass@2 under a 10-second per-task CPU limit
The author asks the community how they tell skipped quantities from valid distractor handling in LLM outputs.
More from Research
- Study: human auditors can't detect sabotage in AI safety research — ronbodkin · 2026-11-21
- CMU Researchers Outline Theory of Capable Agents and a Path to Provable AI Safety — aran_nayebi · 2026-11-21
- Betting on Lean: Formal Proofs May Prove More Reliable Than Human Math — jasondeanlee · 2026-11-21
- Meta's AVTR adds visual imagination to multimodal reasoning, +4.5 pts across 16 benchmarks — liuzhuang1234 · 2026-11-21
- Pro² Agent paper: explicit Problem State lifts autonomous AI R&D to 76% MLE-Bench — rohanpaul_ai · 2026-11-21
- Google & JHU's Pro² Agent hits 76% MLE-Bench medal rate by tracking bottlenecks — rohanpaul_ai · 2026-11-21