150M-Param Model Sets New Cost-Accuracy Frontier on ARC-AGI-1
pathwaycom · hf · 2026-08-11
A new model named BDH-CQ has been released on Hugging Face. Using only 150M parameters, it achieves a new cost-accuracy frontier on the ARC-AGI-1 benchmark by combining recurrent latent reasoning with in-context learning, demonstrating the reasoning potential of small parameter models under specific architectures.
More from Research
- New Chunked KL Loss Slashes Knowledge Distillation VRAM to <6GB for 32K Context — ikergarcia1996 · 2026-08-11
- RAG Me Up: Open-Source RAG Tutorial and Codebase for Engineers — FutureClubNL · 2026-08-11
- Google Introduces Science One Framework to Eliminate AI Research Hallucinations — dl_weekly · 2026-08-11
- Why Do Agent Memory Systems Always Fail After Two Months? Devs Discuss Forgetting — False-Excitement-886 · 2026-08-11
- LLM Watermarking Can Be Repurposed for Imperceptible Text Steganography — dyn___ · 2026-08-11
- Dynamic Data Strategies for Fine-Tuning When Compute is Tight — Puzzleheaded_Box2842 · 2026-08-11