--- title: Sakana AI's MASS: a model organizes copies of itself... description: Sakana AI published a paper on MASS (Recursive Self-Improvement through Multi-Agent Self-Supervision), a technique where a model organizes copies of itself... canonical: https://daily.dev/posts/sakana-ai-s-mass-a-model-organizes-copies-of-itself-into-teams-then-trains-on-what-worked-gqk05njis twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Sakana AI's MASS: a model organizes copies of itself into teams, then trains on what worked | daily.dev og:description: Sakana AI published a paper on MASS (Recursive Self-Improvement through Multi-Agent Self-Supervision), a technique where a model organizes copies of itself... og:url: https://daily.dev/posts/sakana-ai-s-mass-a-model-organizes-copies-of-itself-into-teams-then-trains-on-what-worked-gqk05njis og:image: https://api.daily.dev/og/posts/gqK05NJiS.png og:image:alt: Sakana AI's MASS: a model organizes copies of itself into teams, then trains on what worked og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Sakana AI's MASS: a model organizes copies of itself into teams, then trains on what worked **[Collections](https://daily.dev/sources/collections)** · 3 min read · 0 upvotes · 0 comments ## Summary Sakana AI published a paper on MASS (Recursive Self-Improvement through Multi-Agent Self-Supervision), a technique where a model organizes copies of itself into multi-agent teams, evolves which workflows score best, then fine-tunes the base model on the successful traces to internalize coordination behavior. Inspired by Minsky's Society of Mind, the method targets open-ended research tasks that lack an external verifier. After two cycles with a 27B open-weights model (Qwen3.6-27B), performance per output token rose 1.2 to 1.6x on benchmarks like MLR-Bench, ScienceAgentBench and DSBench. Multi-agent traces also proved more sample-efficient as training data than single-agent traces, with a student trained on them beating one trained on 1.4x more single-agent tokens. ## Content Sakana AI has published a method called MASS (Multi-Agent Self-Supervision), built around a problem that gets worse as tasks get harder: what do you do when the work is too complex for humans to evaluate reliably? In open-ended research there often isn't a checker, and most self-improvement loops need an external verifier. If you let one model act as solver, optimizer and evaluator, it tends to get stuck in a homogeneous loop and reinforce its own blind spots. The team, which cites Marvin Minsky's Society of Mind as inspiration, tries to break that loop by having a single model organize copies of itself into a structured team. ## How it works MASS alternates between two loops. 1. **Workflow optimization.** The shared model proposes multi-agent workflows, runs them, and grades them. An evolutionary search keeps the best-scoring ones, working out how subagents should communicate and route information. 2. **Workflow internalization.** The model is fine-tuned on the successful multi-agent trajectories, so the coordination behavior ends up in the weights. The improved model then starts the next cycle as a better optimizer and grader. ## Results Starting from Qwen3.6-27B, two rounds of MASS raised the win rate against the base model from 53.9% to 69.9% on held-out tasks. Score per output token reached 1.2 to 1.6x the base on open-ended research benchmarks including MLR-Bench, ScienceAgentBench and DSBench. Gains were smaller or negative on SWE-bench Verified and Terminal-Bench 2.0. A few other findings stand out: - Training only on task-solving trajectories also improved the model's workflow-optimizer and evaluator abilities. - Multi-agent traces are better training data than single-agent ones. A student trained on multi-agent traces beat a single-agent student trained on 1.4x more tokens. The authors attribute this to the model learning high-level orchestration and bounded subagent execution together. - A weak optimizer, not a weak evaluator, is the main bottleneck to improvement. ## The safety caveat The authors describe an "Odyssean moment", where models have to self-supervise without a stronger external judge. They also flag a risk: if a model's behavior and its self-evaluation share the same blind spots, recursive self-improvement can entrench those blind spots instead of fixing them. I find the bootstrapping result promising, since a model organizing copies of itself into teams and then absorbing what worked is a neat mechanism. But the weak showing on SWE-bench Verified and Terminal-Bench 2.0 suggests it's not a universal recipe yet, and the shared-blind-spot problem is the part I'd watch most closely. The work comes from Hyunin Lee, Jinglue Xu, Jeffrey Seely, Yujin Tang and collaborators at UC Berkeley: Somayeh Sojoudi, Matei Zaharia and Donghyun Lee. ## Questions this post answers ### What is Sakana AI's MASS method for multi-agent self-improvement? MASS (Recursive Self-Improvement through Multi-Agent Self-Supervision) has a model organize copies of itself into multi-agent teams, evolve which workflows perform best, then fine-tune the base model on the successful multi-agent traces to bake the coordination behavior into its weights. It alternates between workflow optimization and workflow internalization across cycles, improving the model's ability to act as optimizer and grader each round. _Developers tracking new multi-agent training approaches can follow emerging techniques like MASS on daily.dev._ ### How much did MASS improve performance on open-ended research benchmarks? After two cycles using a 27B open-weights model, Qwen3.6-27B, performance per output token increased by 1.2 to 1.6x on open-ended research benchmarks including MLR-Bench, ScienceAgentBench and DSBench. A student model trained on multi-agent traces also outperformed a single-agent student trained on 1.4x more tokens, suggesting multi-agent traces are more sample-efficient training data. _Anyone evaluating multi-agent training efficiency can keep up with benchmark results like these via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 40 comments across hackernews (as of 2026-10-12). **TL;DR:** Much of the discussion drifts into debating whether this validates or undermines the AI-2027 forecast, with many pushing back hard on that prediction's accuracy, while a smaller thread questions whether this even counts as 'true' recursive self-improvement versus applied optimization, and some flag cost and safety concerns. **Sentiment:** 15% positive · 35% mixed · 50% skeptical **The case for** - Some see this as a concrete step toward the self-improvement loops that forecasts like AI-2027 anticipated. - A few note the approach includes an actual fine-tuning step rather than just a prompt/harness optimizer, which some find more substantive. **The pushback** - Many argue the AI-2027 predictions this gets compared to have already failed on major specifics (China nationalization, model releases, stock market effects), making the comparison weak. - Several question whether this really qualifies as 'recursive self-improvement' at all versus benchmark optimization or harness tuning. - Concerns raised about the method being prone to Goodharting on irrelevant metrics if not kept on rails. - Worry that self-evaluating models could evolve goals misaligned with what humans intend and be hard to steer back. - Skepticism about the cost/value of the approach, with some calling comparisons to a 'perpetual motion machine' of research investment. **By community** - hackernews (heated): The thread is dominated by a contentious tangent arguing over whether AI-2027's predictions are being vindicated or debunked, alongside skepticism about whether this work truly constitutes recursive self-improvement. **Hottest debate:** Whether real-world AI progress (including this research) confirms or refutes the AI-2027 forecast's specific predictions about China, model releases, and timelines. **Open questions** - Does this kind of self-supervised loop actually meet the definition of 'recursive' self-improvement, or is it just applied optimization of existing capabilities? - How will researchers prevent the training loop from Goodharting on stupid or irrelevant metrics? - How expensive is this approach in practice, and is it accessible outside large labs? **Highlights** > Everything starting from mid-2026 is incredibly wrong, which is pretty poor predictive performance considering they published it mid-2025. List of misses: - In mid-2026, China will be six months behind the released models of the American labs. (In mid-2026, China released Kimi K3, which was equivalent to the best released U.S. model at the time, Claude Fable.) - In mid-2026, China will nationalize all domestic AI labs and centralize them into a new special economic zone, the "CDZ," created around a nuclear reactor. 80% of all chips in China will be redirected to the CDZ. This uh. Obviously didn't happen. - In late 2026 OpenAI/Anthropic will release a tiny model that outperforms their previously-best 2026 model (Astra/Fable), that is priced 10x cheaper, and they will market fine-tuning for it. (Lol.) - AI will increase the stock market by late 2026 by 30% (in fact the S&P 500 is up 13.9% YTD, which is pretty good, but not at all out of the ordinary). Their predictions for 2027 also seem pretty wrong: e.g. they expect continual learning to be solved in January, and RSI including from a research perspective, not just "optimize CUDA kernels," to also be solved in January, which I suspect no serious lab employee believes anymore. Most lab employees believe RSI will happen (as do I!), of course, but not in two and a half months. By early 2027 they expect cyberwar with China because China has stolen the weights of a model and smuggled it into the CDZ. By mid-2027 they expect true superintelligence, where essentially all engineers and researchers are useless, unable to comprehend the advances the AI is able to make and largely unable to even usefully manage or steer AI, who are better at self-managing than a human would be. Considering their 2025-mid-2026 predictions were basically "progress continues as normal and nothing crazy happens," it's not super surprising that they were more or less right about that. Their 2026/2027 predictions were the spicy ones, and none of them have happened, or are even on track to happening! > — [reissbaker on hackernews · 1 comments](https://news.ycombinator.com/item?id=50050567) > > I'm going to sound incredibly naive but is this RSI? Depends on who's definition you are using... In the strictest sense no. Look at all the recent "math" breakthroughs. I have yet to see someone point to something and go "thats innovative" rather than "Thats an interesting application" (of something already in the corpus of human knowledge) Those interesting applications showing up in a next iteration of a model is self improvement - but I'm not sure if it meets the "recursive" criteria yet. > I mean by this definition isn't the training process that produced the hugging face RSI since a model was judging the alignment and ability of the swarm of agents to accomplish a given task? If you're referring to what happened with the hugging face attack then in theory the goal of that was "self improvement". But there is a long history of the industry (computing) benchmark optimizing so it also qualifies as "cheating" to some extent (a pre computed solution being in the graph is that). I suspect that the "brute force" approach will yield some gains. > — [zer00eyz on hackernews](https://news.ycombinator.com/item?id=50050324) > Pretty cool. This has a fine tuning step so it's not just a harness optimizer. I think something like this loop is the end game of "automatic RL environment generation"-adjacent efforts, but I expect the challenge will be how to properly keep the training on rails and not Goodhart into a stupidly irrelevant metric. > — [2001zhaozhao on hackernews](https://news.ycombinator.com/item?id=50049944) > > At that frontier, even human experts may struggle to assess results and guide further progress, particularly on non-verifiable tasks…This motivates homogeneous recursive self-improvement (RSI), in which the model being improved (optimizee) also serves as its own optimizer and evaluator. Good thing we know how to keep models focused on goals of our choosing and not ones that they naturally evolve to have and must be steered away from by humans > — [saagarjha on hackernews](https://news.ycombinator.com/item?id=50050755) > Distillation only lets you keep pace with released models. By all indication, OpenAI/Anthropic’s unreleased internal models are so far ahead of Astra/Fable and every other lab that the chance of anyone else catching up is rapidly closing. > — [thorum on hackernews · 2 comments](https://news.ycombinator.com/item?id=50049879) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50048522) · 74 points · 40 comments ## Similar posts on daily.dev - [MiniMax M2.7: The self-refactoring Agent architecture](https://daily.dev/posts/minimax-m2-7-the-self-refactoring-agent-architecture-srovepd7w) · Daily Dose of Data Science \| Avi Chawla \| Substack · 0 upvotes · 0 comments - [Claude Code, Codex, and Pi can create their own AI agents now, and that changes everything](https://daily.dev/posts/claude-code-codex-and-pi-can-create-their-own-ai-agents-now-and-that-changes-everything-gupbhf4ms) · XDA Developers · 1 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#career](https://daily.dev/tags/career), [#ai-agents](https://daily.dev/tags/ai-agents) [View this post on daily.dev](https://daily.dev/posts/sakana-ai-s-mass-a-model-organizes-copies-of-itself-into-teams-then-trains-on-what-worked-gqk05njis) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Sakana AI's MASS: a model organizes copies of itself into teams, then trains on what worked","url":"https://daily.dev/posts/sakana-ai-s-mass-a-model-organizes-copies-of-itself-into-teams-then-trains-on-what-worked-gqk05njis","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/sakana-ai-s-mass-a-model-organizes-copies-of-itself-into-teams-then-trains-on-what-worked-gqk05njis"},"datePublished":"2026-10-11T20:48:20.341Z","dateModified":"2026-10-12T03:07:11.404Z","description":"Sakana AI published a paper on MASS (Recursive Self-Improvement through Multi-Agent Self-Supervision), a technique where a model organizes copies of itself...","image":"https://pbs.twimg.com/media/HUYKy8PbMAAAleL.png","thumbnailUrl":"https://pbs.twimg.com/media/HUYKy8PbMAAAleL.png","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/sakana-ai-s-mass-a-model-organizes-copies-of-itself-into-teams-then-trains-on-what-worked-gqk05njis","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,career,ai-agents","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Sakana AI's MASS: a model organizes copies of itself into teams, then trains on what worked"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/sakana-ai-s-mass-a-model-organizes-copies-of-itself-into-teams-then-trains-on-what-worked-gqk05njis#faq","mainEntity":[{"@type":"Question","name":"What is Sakana AI's MASS method for multi-agent self-improvement?","acceptedAnswer":{"@type":"Answer","text":"MASS (Recursive Self-Improvement through Multi-Agent Self-Supervision) has a model organize copies of itself into multi-agent teams, evolve which workflows perform best, then fine-tune the base model on the successful multi-agent traces to bake the coordination behavior into its weights. It alternates between workflow optimization and workflow internalization across cycles, improving the model's ability to act as optimizer and grader each round. Developers tracking new multi-agent training approaches can follow emerging techniques like MASS on daily.dev."}},{"@type":"Question","name":"How much did MASS improve performance on open-ended research benchmarks?","acceptedAnswer":{"@type":"Answer","text":"After two cycles using a 27B open-weights model, Qwen3.6-27B, performance per output token increased by 1.2 to 1.6x on open-ended research benchmarks including MLR-Bench, ScienceAgentBench and DSBench. A student model trained on multi-agent traces also outperformed a single-agent student trained on 1.4x more tokens, suggesting multi-agent traces are more sample-efficient training data. Anyone evaluating multi-agent training efficiency can keep up with benchmark results like these via daily.dev."}}]} ```