150M-parameter BDH-CQ breaks ARC-AGI cost-accuracy frontier
moschles · reddit · 2026-08-15
New research introduces BDH-CQ, a reasoning system performing in-context learning via recurrent memory and iterative computation in a latent space, without verbalizing intermediate reasoning.
- Mechanism: Updates recurrent memory with demos and solves queries via iterative latent computation, bypassing explicit Chain-of-Thought.
- Performance: A 150M-parameter config achieves 29.5% pass@2 on ARC-AGI-1 at $0.0007 per task, breaking the previous cost-accuracy Pareto frontier.
Related event: 150M Model Achieves 29.5% on ARC-AGI(3 posts)→
More from Research
- SPLIT-RAG at EMNLP 2026: multi-agent graph partitioning cuts agentic RAG overhead — flosalim · 2026-09-21
- Neural network trainability boundary is fractal, shows Jascha Sohl-Dickstein — alec_helbling · 2026-09-21
- Qwen-Image 2.1 LoRA training plagued by anatomy failures, community hunts for fixes — ResidentFrame4195 · 2026-09-21
- EventCV: Open-Source Rust Library Lowers the Barrier to Event Cameras in Robotics — ssh4net · 2026-09-21
- UCL+FAIR unveil neural reasoning architecture that learns to generate and select rules — PMinervini · 2026-09-21
- NLP's surprising origins: IEEE traces natural language processing back to Kabbalist mystics — ArtificialOther · 2026-09-21