Pathway's BDH-CQ hits 29.5% on ARC-AGI-1 at $0.0007 per task with 150M params
bendee983 · x · 2026-10-07
A notable AI trend: recurrent reasoning, letting models reason in their own continuous latent space instead of token-by-token discrete CoT. The looped transformer gained attention after GPT-6 Astra's announcement.
Pathway's BDH-CQ combines latent reasoning with in-context learning, previously seen only in autoregressive transformers:
- 29.5% on ARC-AGI-1, setting a new Pareto frontier
- $0.0007 per task, only 150M parameters
- Reasons on novel tasks from examples while being highly compute-efficient
Researchers say scaling prospects are promising.
More from Models
- Western labs less likely to distill from Claude but more likely from GLM 5.3 — andrew_n_carr · 2026-10-07
- With OpenAI's Luna, cloud beats open weights on price — at least GPUs heat the house — BLUECOW009 · 2026-10-07
- V4.1 Flash review: top of Flash tier but wild hallucination swings, tester wary of V4.1 Pro — teortaxesTex · 2026-10-07
- OpenAI Launches Decisions API in Public Beta, Up to 10x Faster Than GPT-6 Luna — stevenheidel · 2026-10-07
- OpenAI to watermark ChatGPT outputs by default in the EU under AI Act — Ars Technica AI · 2026-10-07
- Mistral trained ML4 on 3,800 Grace Blackwell GPUs in its European datacenter, more clusters coming — rakyll · 2026-10-07