Hugging Face Daily Papers · August 11, 2026 · 3 min read

BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

<a href=\"https://cdn-uploads.huggingface.co/production/uploads/651310e3b7994cff61d71320/A-H7LZX0210YoZssD-tKP.png\" rel=\"nofollow\"><img src=\"https://cdn-uploads.huggingface.co/production/uploads/651310e3b7994cff61d71320/A-H7LZX0210YoZssD-tKP.png\" alt=\"arc-agi-1-bdh-cost-efficiency-paper-Aug8\"></a></p>\n","updatedAt":"2026-08-11T10:07:00.620Z","author":{"_id":"651310e3b7994cff61d71320","avatarUrl":"/avatars/161cb412c7cb284fa7483e790807e4d3.svg","fullname":"Jan Chorowski","name":"janchorowski","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.47814205288887024},"editors":["janchorowski"],"editorAvatarUrls":["/avatars/161cb412c7cb284fa7483e790807e4d3.svg"],"reactions":[{"reaction":"🔥","users":["muditjps","sakubhai06","themanas021","Brain-synapse","architaamahajan","Medha1308","weran1te","ivanp25","clrnt","chrisociepa","Starlitakash","Prakhar7017","aksh1845","Akurzrok","mabhusubhani001"],"count":15},{"reaction":"🚀","users":["themanas021","weran1te","sakubhai06","ivanp25","Starlitakash","Prakhar7017"],"count":6},{"reaction":"❤️","users":["themanas021","weran1te","ivanp25","Starlitakash","Prakhar7017"],"count":5},{"reaction":"🧠","users":["sakubhai06","ivanp25","Starlitakash","Prakhar7017"],"count":4}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.09888","authors":[{"_id":"6a7af16fb7183653340c11cd","user":{"_id":"66d17d0d2fa8a088dc6992b3","avatarUrl":"/avatars/0aeb72e86b49d85db5577b507751147a.svg","isPro":false,"fullname":"Björn Engdahl","user":"bjornengdahl","type":"user","name":"bjornengdahl"},"name":"Björn Engdahl","status":"claimed_verified","statusLastChangedAt":"2026-08-11T16:45:04.704Z","hidden":false},{"_id":"6a7af16fb7183653340c11ce","name":"Adrian Kosowski","hidden":false},{"_id":"6a7af16fb7183653340c11cf","user":{"_id":"651310e3b7994cff61d71320","avatarUrl":"/avatars/161cb412c7cb284fa7483e790807e4d3.svg","isPro":false,"fullname":"Jan Chorowski","user":"janchorowski","type":"user","name":"janchorowski"},"name":"Jan Chorowski","status":"claimed_verified","statusLastChangedAt":"2026-08-11T16:45:04.695Z","hidden":false},{"_id":"6a7af16fb7183653340c11d0","user":{"_id":"68dd50bb9d2cc46fef645799","avatarUrl":"/avatars/699f84aa771851854a4201cb1df3543b.svg","isPro":false,"fullname":"Zuzanna Stamirowska","user":"Brain-synapse","type":"user","name":"Brain-synapse"},"name":"Zuzanna Stamirowska","status":"claimed_verified","statusLastChangedAt":"2026-08-11T16:45:04.700Z","hidden":false},{"_id":"6a7af16fb7183653340c11d1","name":"Przemysław Uznański","hidden":false},{"_id":"6a7af16fb7183653340c11d2","name":"Junlin Jiang","hidden":false},{"_id":"6a7af16fb7183653340c11d3","name":"Rohan Phadke","hidden":false},{"_id":"6a7af16fb7183653340c11d4","name":"Remigiusz Kinas","hidden":false},{"_id":"6a7af16fb7183653340c11d5","name":"Richard Zhong","hidden":false}],"publishedAt":"2026-08-10T00:00:00.000Z","submittedOnDailyAt":"2026-08-11T00:00:00.000Z","title":"BDH-CQ: In-Context Learning with Recurrent Latent Reasoning","submittedOnDailyBy":{"_id":"651310e3b7994cff61d71320","avatarUrl":"/avatars/161cb412c7cb284fa7483e790807e4d3.svg","isPro":false,"fullname":"Jan Chorowski","user":"janchorowski","type":"user","name":"janchorowski"},"summary":"We introduce BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. Inputs presented at inference time continuously update the model's recurrent memory; the model then solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning. We evaluate the model on the public ARC-AGI-1 evaluation set and use controlled ARC-like interventions to study what it learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult. A 150M-parameter configuration reaches 29.5% pass@2 at a computed inference cost of \\$0.0007 per task. This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.","upvotes":231,"discussionId":"6a7af170b7183653340c11d6","projectPage":"https://pathway.com/blog/pathway-150m-model-breaks-arc-agi-1-cost-efficiency-frontier","githubRepo":"https://github.com/pathwaycom/arc-task-gen","githubRepoAddedBy":"user","ai_summary":"A 150M-parameter reasoning model using recurrent latent reasoning and in-context learning achieves a new cost-accuracy frontier on ARC-AGI-1.","ai_keywords":["in-context learning","recurrent latent reasoning","high-dimensional latent space","ARC-AGI-1","cost-accuracy Pareto frontier"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":2,"organization":{"_id":"6697b1a8b981e21b48bc575f","name":"pathwaycom","fullname":"Pathway","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6697aead07b36ccd016f0888/nOEvf0HpKBQJcVDfljgAc.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"68dd5359ce22e16f17bcf21b","avatarUrl":"/avatars/7b476a02105cdcced1546eac7024b832.svg","isPro":false,"fullname":"Saksham Garg","user":"sakubhai06","type":"user"},{"_id":"65b67ff2138466e48b756d08","avatarUrl":"/avatars/bb1ad1f01fbb97401e16da0d3dfeb7ee.svg","isPro":false,"fullname":"Akshay","user":"Cappybara12","type":"user"},{"_id":"64952bfcb70715a4bc2b2c83","avatarUrl":"/avatars/4af11732aa4e0095ec4519b17353bfa1.svg","isPro":false,"fullname":"Lakshay Dagar","user":"lakshay315","type":"user"},{"_id":"6a7a202f908d777551108ad4","avatarUrl":"/avatars/2c38975af0bcde8b1a21353313d26353.svg","isPro":false,"fullname":"Archita Mahajan","user":"architaamahajan","type":"user"},{"_id":"64aa3e925a69e2ca88ac6295","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64aa3e925a69e2ca88ac6295/m8-Bksow3gE44CgvqW-lV.jpeg","isPro":false,"fullname":"Manas Chopra","user":"themanas021","type":"user"},{"_id":"6a7af7ade8c0dd917e9ee135","avatarUrl":"/avatars/65a378091f1e8d6537f235206f29d985.svg","isPro":false,"fullname":"Himansu Maurya","user":"HimansuMaurya","type":"user"},{"_id":"68f7f3780a277a9bf7383596","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/xLeyezSaI7cBguRdSywQn.jpeg","isPro":false,"fullname":"AmanVadev","user":"vasdevaman6","type":"user"},{"_id":"6a76106da3f602f644d2f409","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/zQv7KcYf4bhcgQdDiZgNf.png","isPro":false,"fullname":"Prajjwal singh","user":"prajjwal1112","type":"user"},{"_id":"68dcf4154552213537963e0e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/HRoJIsk7Q32zT51jvQY41.png","isPro":false,"fullname":"Shaurya Raj","user":"shauryaraj694","type":"user"},{"_id":"63a48fe7658851481f766a24","avatarUrl":"/avatars/4d1738e80fd7677a81197fb158ba8c44.svg","isPro":false,"fullname":"Adrian Kosowski","user":"dxtrous","type":"user"},{"_id":"673c86d8c1f31aade295d922","avatarUrl":"/avatars/4ff28923d0bc9ad98d25e0900b588405.svg","isPro":false,"fullname":"Ronit Mathur","user":"Ronit0001","type":"user"},{"_id":"661164d8db728ea9676dbb8e","avatarUrl":"/avatars/35710c69ef0070a2730a21928250dec6.svg","isPro":false,"fullname":"Vinayak Goyal","user":"Vinayak-goyal","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":1,"organization":{"_id":"6697b1a8b981e21b48bc575f","name":"pathwaycom","fullname":"Pathway","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6697aead07b36ccd016f0888/nOEvf0HpKBQJcVDfljgAc.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.09888.md","query":{}}">

Papers

arxiv:2608.09888

BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

Published on Aug 10

· Submitted by

Jan Chorowski on Aug 11

Authors:

Björn Engdahl ,

Jan Chorowski ,

Zuzanna Stamirowska ,

Abstract

A 150M-parameter reasoning model using recurrent latent reasoning and in-context learning achieves a new cost-accuracy frontier on ARC-AGI-1.

Generated by thinkingmachines/Inkling-Small

We introduce BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. Inputs presented at inference time continuously update the model's recurrent memory; the model then solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning. We evaluate the model on the public ARC-AGI-1 evaluation set and use controlled ARC-like interventions to study what it learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult. A 150M-parameter configuration reaches 29.5% pass@2 at a computed inference cost of \$0.0007 per task. This operating point breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier, establishing a new state of the art in benchmark cost efficiency.