Non-LLM Reasoning System Scores 100% on ARC-AGI-3 ft09 with Zero Model Calls
Living_Substance1274 · reddit · 2026-08-10
Developer Orivael built an experimental reasoning system that uses absolutely no LLMs in the loop for perception, planning, or action, achieving notable results on ARC-AGI-3.
- Key Results: Scored 100% on ft09 (6/6 levels in 80 actions, beating the 208-action human baseline) with a total model inference cost of $0.00. Performance on other tasks like tr87 and cd82 remained low.
- Failure Analysis: The author found that major failures usually stemmed from drawing perfectly reasonable conclusions based on an incorrect representation of the environment. Examples include mistaking a sprite for a wall due to shared color values, or classifying buttons as inert because they were tested in the wrong state.
- Core Challenge: While the agent becomes highly efficient once it identifies a game's mechanics, the core problem remains: how to recognize a new world without carrying over assumptions from previous ones.
More from Research
- MLCommons launches MLPerf Endpoints, a benchmark that measures inference endpoints the way buyers actually purchase them — TheKanter · 2026-10-16
- DIY LEGO bench tests AI brick-building, GPT-6 Astra beats Opus — TheLegoless · 2026-10-16
- tszzl's theory: Tornado → Apache license → apaches → Southwest → Colorado — tszzl · 2026-10-16
- DIRECT: retrieval-based next-question suggestions for conversational agents — burkov · 2026-10-16
- A Qwen3 neuron tracks both Colorado and Tornado HTTP code; tszzl offers a theory — tszzl · 2026-10-16
- AI-Driven Timing-Aware Power Optimization Uses Learned Ranking With Baseline-Preserving Repair — Amit Sharma · 2026-10-16