Coverage of Darwin Gödel Machine and The AI Scientist in MIT Technology Review article. @technologyreview.com www.technologyreview.com/2025/08/06/1...

フロンティアAIモデルを「混ぜて使う」── 「試行錯誤」と「集合知」で新たな推論時スケーリングへ ブログ： sakana.ai/ab-mcts-jp/ 論文： arxiv.org/abs/2503.04412 このたびSakana AIは新アルゴリズム「AB-MCTS」を開発し、ARC-AGI-2ベンチマークで有望な結果を得ました。

AIにもっと“試行錯誤”と“集合知”を─Sakana AIが開発する新アルゴリズム wired.jp/article/saka... フロンティアモデルと呼ばれるAIを単体ではなく“混ぜて”使えば、個々のモデル─ChatGPT、Gemini、DeepSeek─を使うよりも大幅に上回る成績を出すことが可能だと、日本発AIスタートアップのSakana AIが発表した。

AIも「3人寄れば文殊の知恵」、Sakana AIが新しい推論手法を開発 3人集まれば1人よりも優れた知恵が出るということわざ「3人寄れば文殊の知恵」が、AIにも当てはまった格好だ。 🐡🐟🐠 xtech.nikkei.com/atcl/nxt/new...

Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search arxiv.org/abs/2503.04412

Sakana AIではApplied Teamの立ち上げを急速に進めており、優秀なApplied Research Engineerを引き続き募集しています🚀 sakana.ai/careers/#app... 正社員だけでなく学生インターンシップも歓迎です✨ 金融・保険などのエンタープライズ分野から政府・防衛などの公共分野での業務に興味のある方 最先端のAI技術を実社会に導入してインパクトを出したい方 雇用期間や勤務スタイルの相談もできますのでぜひご応募ください！

Just published a blog post on our new LLM answer search method: "Multi-LLM AB-MCTS”🚀 It's designed to flexibly explore how to search and which LLM to use for any given problem. We've also open-sourced the implementation and experiments. Check it out! 🙌

We’re excited to introduce AB-MCTS! Our new inference-time scaling algorithm enables collective intelligence for AI by allowing multiple frontier models (like Gemini 2.5 Pro, o4-mini, DeepSeek-R1-0528) to cooperate. Blog: sakana.ai/ab-mcts Paper: arxiv.org/abs/2503.04412

Inspired by the power of human collective intelligence, where great achievements arise from the collaboration of diverse minds, we believe the same principle applies to AI. Individual models possess unique strengths and biases, which we view as valuable resources for collective problem-solving.

AB-MCTS (Adaptive Branching Monte Carlo Tree Search) harnesses these individualities, allowing multiple models to cooperate and engage in effective trial-and-error, solving challenging problems for any single AI.

Our initial results on the ARC-AGI-2 benchmark are promising, with AB-MCTS combining o4-mini + Gemini-2.5-Pro + R1-0528, current frontier AI models, significantly outperforming individual models by a substantial margin.

This research builds on our 2024 work on evolutionary model merge, shifting focus from “mixing to create” to “mixing to use” existing, powerful AIs. At Sakana AI, we remain committed to pioneering novel AI systems by applying nature-inspired principles such as evolution and collective intelligence.

We believe this work represents a step toward a future where AI systems collaboratively tackle complex challenges, much like a team of human experts, unlocking new problem-solving capabilities and moving beyond single-model limitations. Algorithm (TreeQuest): github.com/SakanaAI/tre...

Inference-Time Scaling and Collective Intelligence for Frontier AI sakana.ai/ab-mcts/ We developed AB-MCTS, a new inference-time scaling algorithm that enables multiple frontier AI models to cooperate, achieving promising initial results on the ARC-AGI-2 benchmark.

「時間を使って考える」AIの新パラダイム ：Continuous Thought Machine（CTM）を提案 🧠 日本語ブログ ： sakana.ai/ctm-jp インタラクティブレポート ： pub.sakana.ai/ctm Sakana AIは、時間情報を明示的に扱う新しいAIモデル「Continuous Thought Machine（CTM）」を発表しました 。

“Continuous Thought Machines” Blog → sakana.ai/ctm Modern AI is powerful, but it's still distinct from human-like flexible intelligence. We believe neural timing is key. Our Continuous Thought Machine is built from the ground up to use neural dynamics as a powerful representation for intelligence.

The Continuous Thought Machine (CTM) incorporates neuron-level temporal processing and neural synchronization, moving beyond current AI limitations. Interactive Paper (with web-demo): pub.sakana.ai/ctm/ Full Paper: arxiv.org/abs/2505.05522 GitHub Project: github.com/SakanaAI/con...

Japan’s Sakana AI sees opportunity with US uncertainty “While there are many AI companies in the US and China, Japanese firms have had little global presence. We believe there’s a demand—particularly among government agencies—for domestically developed AI solutions.” asia.nikkei.com/Business/Tec...

1/ Transformer-Squared: Self-adaptive LLMs Paper: openreview.net/forum?id=dh4... Transformer-Squared adapts its weights on the fly for each query, achieving strong performance across tasks and enabling parameter-efficient life-long learning.

2/ Agent Skill Acquisition for Large Language Models via CycleQD Paper: openreview.net/forum?id=Kvd... CycleQD is an ecological niche-inspired model-merging approach that achieves great performance on computer science tasks while retaining language capabilities.

3/ An Evolved Universal Transformer Memory Paper: openreview.net/forum?id=s1k... Neural Attention Memory Models (NAMMs) is an evolved memory system trained to improve performance and efficiency on language transformers, and zero-shot transferring to vision and RL foundation models.

4/ TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models Paper: openreview.net/forum?id=cqs... TAID is a novel knowledge distillation method that uses a time-dependent intermediate distribution addressing common challenges in distilling LLMs.

5/ Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization Paper: openreview.net/forum?id=gx1... Drop‑Upcycling: We show that partial re‑initialization can turn dense LLM into a sparse MoE!

6/ Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search Paper: openreview.net/forum?id=3HF... Workshop: fm-wild-community.github.io AB-MCTS is a novel inference-time framework that unifies repeated sampling with principled multi-turn exploration and exploitation.

7/ Large Language Models to Diffusion Finetuning Paper: openreview.net/forum?id=Wu5... Workshop: workshop-llm-reasoning-planning.github.io New finetuning method empowering pre-trained LLMs with some of the key properties of diffusion models and the ability to scale test-time compute.

8/ We will give a talk at #ICLR2025 Workshop, ICBINB (Apr 28), who let us perform an experiment of submitting papers fully generated using The AI Scientist-v2 into peer review. We’ll discuss our experiment and learnings. Blog: sakana.ai/ai-scientist... Workshop: sites.google.com/view/icbinb-...

9/ Our researcher, Yutaro Yamada, will further discuss more details, and progress made with The AI Scientist-v2 at the #ICLR2025 social event (“AI Co-scientist Discussion”), on Saturday, April 26th, 5pm SGT: Social@ICLR2025 Website: ai-researcher.net/social-iclr-...

10/ Sakana AI Co-Founder and CEO, David Ha, will be giving a talk at the #ICLR2025 World Models Workshop, at a panel to discuss the Current Development and Future Challenges of World Models. Workshop Website: sites.google.com/view/worldmo...

Sakana AI will be presenting 7 super interesting papers and giving 3 talks at #ICLR2025❗ 🧵 Check out this thread 👇

Will you be at #ICLR2025 in Singapore this week? 🇸🇬 Sakana AI is presenting several works at the conference. Come chat with our researchers! 🐙🐟🐠🐡 🧵 Scroll down this thread to see our list of presentations 👇

「世界トップのエンジニアがSakana AIに集まるワケ」（伊藤錬インタビュー） 文春オンラインにて、Sakana AI共同創業者・COOの伊藤錬のインタビューが掲載されました。 bunshun.jp/articles/-/7... 世界の第一線のエンジニア・リサーチャーがSakana AIに集まる背景について語りました。

Their team successfully applied SFT and RL to DeepSeek-R1-Distill-Qwen-14B. You can find more details about their Kaggle solution here: www.kaggle.com/competitions...

Congratulations to Sakana AI researcher Yuichi Inoue (@inoichan.bsky.social) and his collaborators on winning a gold medal (placing 9th out of 2,211 teams) at the AI Mathematical Olympiad! 🏅 This Kaggle competition focuses on LLMs’ math problem-solving capabilities.

Dive deeper into AI Scientist-v2 methods, the experiment, our analysis and ethical considerations in our technical report: pub.sakana.ai/ai-scientist... The project is open-sourced on GitHub: github.com/SakanaAI/AI-... What's the future of AI in science? We would love to hear your thoughts❗

江戸時代の古文風テキストで会話できるチャットボット「からまる」を公開 ブログ → sakana.ai/karamaru/ デモ → huggingface.co/spaces/Sakan... Sakana AIが江戸時代のテキストで学習した「からまる」は現代日本語で質問すると、江戸時代の世界観と当時の古文風テキストで回答してくれます。

Sakana AIで働く研究者インタビュー ブログ → sakana.ai/cv-interview... 「コンピュータービジョン最前線」にて、Sakana AIの秋葉拓哉、井ノ上雄一、黒木颯の3名の研究者が、入社の理由から、日々の働き方、AI技術の未来まで、幅広く語りました。 「仕事は100%ガッツリ研究」で、「それはどう面白いか」を日々議論しています。 書籍ページ → www.kyoritsu-pub.co.jp/book/b101234... この度、共立出版様のご快諾を得て、インタビュー記事の冒頭部分を転載します。

Sakana AI super-powers AI reasoning using Japan’s own Sudoku Puzzles! At NVIDIA GTC 2025, Llion Jones announced the release of our new reasoning benchmark based on the modern variant Sudoku to challenge the AI community. Read more here → sakana.ai/sudoku-bench

Sakana AI、防衛イノベーションの日米コンペティションで受賞 ブログ → sakana.ai/defense-chal... 日本の防衛装備庁と米国国防総省傘下のDefense Innovation Unit（DIU）が主催するチャレンジで、バイオディフェンスと偽情報対策の両部門で新たなAIソリューションを提案し、世界６０社の応募の内、日本企業として唯一の受賞を果たしました。

Sakana AI Wins Award at US-Japan Competition for Defense Innovation sakana.ai/defense-chal... We proposed novel AI solutions at a challenge jointly hosted by US Department of Defense’s DIU and Japan Ministry of Defense’s ATLA. Out of 60 companies from around the world, Sakana AI won 1 of 3 awards.

This was a fun experiment we conducted while developing The AI Scientist-v2. With the permission of ICLR, we submitted an AI-generated paper to an ICLR workshop that passed the peer-review process. We documented the entire process and what we have learned in a blog post: sakana.ai/ai-scientist...

The AI Scientist Generates its First Peer-Reviewed Scientific Publication We’re proud to announce that a paper produced by The AI Scientist-v2 passed the peer-review process at a workshop in ICLR, a top AI conference. Read more about this experiment → sakana.ai/ai-scientist...

Can AI Improve Itself? In an interview with Machine Learning Street Talk, co-first authors of “The AI Scientist” paper, Chris Lu, Cong Lu, and Robert Lange share their views on the recent and future development of Self-Improving AI. Watch the full podcast on YouTube → youtu.be/1kwbp8hRRfs

A key problem with the standard distillation is the mismatch between student-teacher distributions. With TAID, we dynamically interpolate student and teacher distributions through an adaptive intermediate distribution, gradually shifting from the student’s initial distribution towards the teacher’s.

“TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models” has been accepted at #ICLR2025 as a Spotlight Paper ✨ Blog post (English): sakana.ai/taid/ Paper (OpenReview): openreview.net/forum?id=cqs...

We are releasing a dataset of the lifeforms found by ASAL from our large scale searches! View the datasets: pub.sakana.ai/asal/data/il... pub.sakana.ai/asal/data/il... To load them up and play around with them visit the ASAL GitHub or this Colab notebook: colab.research.google.com/github/Sakan...

Introducing ASAL: Automating the Search for Artificial Life with Foundation Models Blog: sakana.ai/asal/ We propose a new method called Automated Search for Artificial Life (ASAL) which uses foundation models to automate the discovery of the most interesting and open-ended artificial lifeforms!

Sakana AIが進化の先に見る“自然としてのAI” (Wired JAPAN) 「わたしたちがこの会社を始めたとき、周りと必ずしも同じ方向ではない、自分たちが興味深いと信じるアイデアに取り組もうとしました。そうした研究開発のアプローチは理想主義的かもしれませんが、ビジネスの観点からも、日本では米国のAI開発者たちよりもう少し賢く物事を行なう必要がありました。単に米国のビジネスモデルをコピーすることはできません。」 wired.jp/article/sz-s...

“AI creating AI” 2024 was a blast. Sakana AI has been in operation for just over a year, and we have built not only a world class research team with some of the most creative and innovative minds, but also a world class applied engineering and business team in Tokyo. sakana.ai/careers/ A Thread 👇

Neural Attention Memory Models are evolved to optimize the performance of Transformers by actively pruning the KV cache memory. Surprisingly, we find that NAMMs are able to zero-shot transfer its performance gains across architectures, input modalities and even task domains! arxiv.org/abs/2410.13166