@sakanaai.bsky.social
Sakana AI is an AI R&D company based in Tokyo. 🗼🧠
Try Sakana Chat, Translate, Marlin, Namazu, Fugu → https://chat.sakana.ai/
@hardmaru.bsky.socialReposted by @sakanaai.bsky.socialNew Sakana AI paper: Recursive Self-Improvement through Multi-Agent Self-Supervision (MASS).
Blog: pub.sakana.ai/mass/
Paper: arxiv.org/abs/2610.12176
How do we keep improving AI when tasks become too complex for humans to reliably evaluate?
The model eventually hits a supervision bottleneck.
@sakanaai.bsky.socialThis work was a fantastic collaboration between Sakana AI and researchers at UC Berkeley.
Check out the full paper and codebase above to dive into the details!
Blog: pub.sakana.ai/mass/
Paper: arxiv.org/abs/2610.12176
Code: github.com/SakanaAI/mass
@sakanaai.bsky.socialFinally, we believe the community needs to investigate homogeneous RSI deeply to sustain AI safety.
When a model is its own evaluator, shared blind spots allow mistakes to be accepted as progress. To safely improve beyond human expertise, we must understand these failures.
@sakanaai.bsky.socialWe made 3 key observations: Role generalization sustains RSI: training only on task-solving data improves both the optimizer and evaluator. Multi-agent trajectories are more efficient training data than single-agent ones. Optimizer capacity, not evaluator capacity, is the bottleneck for RSI speed
@sakanaai.bsky.socialTraining on the best team’s executions provides self-supervision by distilling the team’s collective experience back into the shared model. After running two cycles on a 27B open-weights model, the score per output token reached 1.2–1.6× the base model’s level across four research benchmarks.
@sakanaai.bsky.socialWe asked: Can a model learn from the collective intelligence of its own team?
With MASS, one shared model solves tasks through a team of agents, searches for a better workflow, and uses its own judgments to select the best one.
pub.sakana.ai/mass/
@sakanaai.bsky.socialTo advance beyond humanity’s accumulated knowledge, AI itself may need to become its own optimizer and evaluator, since human experts will struggle to assess results.
The challenge? Obtaining useful self-supervision when a model is trapped in a homogeneous loop.
@sakanaai.bsky.socialIntroducing Multi-Agent Self-Supervision (MASS): Recursive self-improvement through collective intelligence 🧵
Blog: pub.sakana.ai/mass/
Paper: arxiv.org/abs/2610.12176
Code: github.com/SakanaAI/mass
@sakanaai.bsky.socialOur goal is to give reviewers useful support in checking research, with human expertise and judgment at the center.
Open Review: openreview.net/forum?id=7iX...
@sakanaai.bsky.socialIts feedback emphasized different aspects of the work from human reviews, offering a complementary perspective, while its assessments of paper quality remained broadly consistent with human judgments.
@sakanaai.bsky.socialIn our evaluations, the system detected more errors than the other review systems we tested, including on papers withdrawn because of real mistakes.
@sakanaai.bsky.socialWe also introduce Multi-Layered Review, an AI review system inspired by the 3-Pass Approach to reading papers. It outlines the main ideas, then examines the details and potential weaknesses, finally brings observations together into a review. The simple idea: understand the paper before judging it.
@sakanaai.bsky.socialNot all errors carry the same weight. Some undermine a paper’s central findings; others affect smaller details. We map the connections between each paper’s claims, methods, evidence in a knowledge graph to estimate severity of each contradiction and better assess what different systems can catch.
@sakanaai.bsky.socialWe build an automated pipeline that deliberately introduces contradictions into papers by adding statements that conflict with information elsewhere in the manuscript. These planted errors give us clear targets for testing whether AI reviewers can spot and explain what is wrong.
@sakanaai.bsky.socialAI review systems are emerging as a potential solution, but much of their development and evaluation focuses on how closely they imitate human reviews. We explore how AI can support the review process without losing human touch. Focus on one demanding but essential task: catching errors in papers.
@sakanaai.bsky.socialBeyond Imitation: A Framework and Benchmark for LLM-Assisted Peer Review
Peer review needs support, not substitutes. Accepted at TMLR: our new paper on using AI to help reviewers catch errors in papers.
As submissions grow, so does workload for experts who evaluate them.
@sakanaai.bsky.socialニュースレター「Sakana AI Insider」、登録受付中です🐟
・プロダクトのリリース情報
・研究開発のアップデート
・Swagプレゼントやイベントの優先案内
先日公開し、多数ご応募いただいたイベントでは購読者の皆様への先行受付を実施しました！
登録はこちら 👇
forms.gle/E4siqGLT8MQF... 🐟
@hardmaru.bsky.socialReposted by @sakanaai.bsky.socialRegistration now open: From World Models to Real-World AI
Luma: luma.com/kjf0z516
Join us on October 26 in Tokyo for a symposium with Jürgen Schmidhuber, who recently joined Sakana AI as Chief Scientific Advisor.
@sakanaai.bsky.social日時：2026年10月26日（月）15:00〜18:00
会場：一橋講堂（東京・千代田区）
言語：英語（通訳なし）
参加費：無料（事前登録制）
※ 席に限りがありますので、お早めにお申し込みください。
お申し込みはこちら：
luma.com/kjf0z516 🐟
@sakanaai.bsky.social【受付開始】From World Models to Real-World AI — 日本から拓くAIの新時代
Chief Scientific Advisorとして当社に参画したJürgen Schmidhuber博士を迎えるシンポジウムを、10月26日（月）に東京・一橋講堂で開催します。
本日より一般申し込みを開始しました。Schmidhuber博士による基調講演とQ&A、 CEO David Haとの対談、Sakana AI研究者によるショート講演を予定しています。
お申し込みはこちら：
luma.com/kjf0z516 🐟
