{"slug": "bdh-cq-hits-29-5-on-arc-agi-1-with-only-150m-parameters", "title": "BDH-CQ hits 29.5% on ARC-AGI-1 with only 150M parameters", "summary": "BDH-CQ, a 150M-parameter model, achieved a 29.5% pass@2 score on the ARC-AGI-1 benchmark at a cost of $0.00070 per task, demonstrating that recurrent latent state reasoning can outperform larger models on abstract reasoning tasks. The model, detailed in a paper on arXiv (2608.09888), performs in-context learning by updating a recurrent memory and iterating internally without decoding intermediate reasoning steps into language.", "body_md": "# BDH-CQ hits 29.5% on ARC-AGI-1 with only 150M parameters\n\nBasically, it handles in-context learning by updating a recurrent memory when it sees a new task. When it comes time to solve the actual query, it iterates internally. The critical takeaway here is that intermediate reasoning states are never decoded into language. It's essentially \"silent\" reasoning.\n\n## How the latent workspace actually functions\n\nUnlike typical LLM agents that rely on prompt engineering to force a step-by-step verbal breakdown, BDH-CQ integrates memory and inference into the same computational fabric. The process looks like this:\n\n1. **Memory Update:** The model takes in demonstrations of a task it has never seen before. Instead of just storing these in a KV cache, it updates its recurrent memory.\n\n2. **Latent Iteration:** The query is processed through iterative computation. It cycles through the latent space to refine the answer.\n\n3. **Direct Output:** It jumps straight to the solution without verbalizing the \"scratchpad\" steps.\n\nThis architecture avoids the need for task identifiers or specific demonstration pairs during the training phase. Even more impressive is that no parameters are updated during inference; it's all happening through the recurrent state.\n\n## Performance and Efficiency\n\nThe numbers on this are wild when you consider the model size. A 150M-parameter setup—which is tiny compared to the behemoths we usually see—hit a 29.5% pass@2 on ARC-AGI-1.\n\n**Model Size:** 150M parameters**ARC-AGI-1 Pass@2:** 29.5%**Cost per task:**$0.00070\n\nComparing this to larger models that try to brute-force AGI benchmarks with massive prompt windows, BDH-CQ proves that recurrent latent states can be way more efficient than token-heavy reasoning. It's a real-world example of how moving away from \"thinking out loud\" might actually lead to better generalization on abstract reasoning tasks.\n\nIf you're looking for a deep dive into the math, the full paper is available here:\n\n```\nhttps://arxiv.org/abs/2608.09888\n```\n\n[A 150M parameter model hitting 29.5% on ARC-AGI-1 is insane 8h ago](/en/news/6379/)\n\n[BDH-CQ is hitting tasks at $0.007 and that crushes OpenAI Luna 1d ago](/en/news/6211/)\n\n[Pathway's 150M model just hit 29. 3d ago](/en/news/5979/)\n\n[Next Steam catapults might be making a comeback on US aircraft →](/en/news/6430/)\n\n[a practical ChatGPT prompt guide](https://tanyan888.com/), with plenty of directly applicable cases.", "url": "https://wpnews.pro/news/bdh-cq-hits-29-5-on-arc-agi-1-with-only-150m-parameters", "canonical_source": "https://promptcube3.com/en/news/6433/", "published_at": "2026-08-15 12:31:17+00:00", "updated_at": "2026-08-15 12:41:32.913711+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-research"], "entities": ["BDH-CQ", "ARC-AGI-1", "arXiv"], "alternates": {"html": "https://wpnews.pro/news/bdh-cq-hits-29-5-on-arc-agi-1-with-only-150m-parameters", "markdown": "https://wpnews.pro/news/bdh-cq-hits-29-5-on-arc-agi-1-with-only-150m-parameters.md", "text": "https://wpnews.pro/news/bdh-cq-hits-29-5-on-arc-agi-1-with-only-150m-parameters.txt", "jsonld": "https://wpnews.pro/news/bdh-cq-hits-29-5-on-arc-agi-1-with-only-150m-parameters.jsonld"}}