🤖 machine learning

Sakana Fugu Technical Report

The paper introduces Sakana Fugu, a family of orchestrator language models that dynamically construct agentic scaffolds to combine and amplify the specialized capabilities of multiple LLMs, achieving state-of-the-art performance across diverse challenging benchmarks.

Original authors: Yujin Tang, Edoardo Cetin, Jinglue Xu, Qi Sun, Stefan Nielsen, Vincent Richard, Haruto Goda, Iaroslav Tymchenko, Nhan Nguyen, Hyunin Lee, Mari Ashiga, Shashank Kotyan, So Kuroki, Tarin Clanuwat

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yujin Tang, Edoardo Cetin, Jinglue Xu, Qi Sun, Stefan Nielsen, Vincent Richard, Haruto Goda, Iaroslav Tymchenko, Nhan Nguyen, Hyunin Lee, Mari Ashiga, Shashank Kotyan, So Kuroki, Tarin Clanuwat

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a team of world-class experts, each a master of a specific field. One is a genius at math, another is a coding wizard, a third is a cybersecurity detective, and a fourth is a scientific researcher. Individually, they are brilliant. But what if you could hire a super-smart manager who knows exactly which expert to call for every single problem, and how to get them to work together?

That is exactly what Sakana Fugu is.

According to the technical report from Sakana AI, Fugu isn't just one giant brain trying to do everything. Instead, it's a "conductor" or a "manager" that orchestrates a team of other powerful AI models (like GPT-5.5, Claude Opus, and Gemini) to solve problems better than any single one of them could alone.

Here is a simple breakdown of how it works and what it can do:

1. The Two Versions: The "Speedy Waiter" and the "Master Chef"

The team released two versions of this manager, designed for different needs:

Fugu (The Speedy Waiter): This version is built for everyday use. When you ask a question, it acts like a fast waiter who instantly knows which expert in the kitchen is best for your order. It picks one expert to do the job immediately. It's fast, low-latency, and great for daily tasks, but it still picks the best person for the job.
Fugu-Ultra (The Master Chef): This version is for the hardest, most complex problems where speed doesn't matter as much as getting the answer perfect. Instead of just picking one person, it builds a whole kitchen crew. It might ask one expert to draft a plan, a second to check the math, a third to write the code, and a fourth to debug it. It takes longer, but it combines the strengths of the whole team to tackle problems that are too hard for a single model.

2. How It Learns: The "Talent Scout"

You might wonder, "How does Fugu know which expert is best for which job?"

The paper explains that Fugu was trained like a talent scout.

The Training: The researchers gave the Fugu manager thousands of problems. They watched how the different experts (the "workers") performed on each one.
The Learning: Fugu learned a pattern. For example, it learned that "Gemini is amazing at science questions," while "GPT is a wizard at complex math," and "Claude Opus is the best at finding security bugs."
The Result: Instead of memorizing answers, Fugu learned to route questions. It looks at your question, thinks, "Ah, this needs a math expert," and instantly sends it to the right model.

3. What It Can Do: The "Swiss Army Knife" Effect

The report shows that by using this "teamwork" approach, Fugu beats the individual experts in many areas. It's like having a Swiss Army knife where every tool is the best version of itself, and the handle knows exactly which tool to pull out.

Here are some specific examples from the paper:

Coding & Software: In tests where AI had to fix bugs in real software, Fugu-Ultra was the top performer. It would have a "builder" model write the code, then switch to a "debugger" model to find the mistakes, and then switch back to fix them. This back-and-forth was better than any single model trying to do it alone.
Science & Math: On difficult science exams (like the "Humanity's Last Exam"), Fugu-Ultra got higher scores than the individual models. It knew to use the math-heavy model for physics problems and the science-heavy model for biology questions.
Creative Problem Solving: The paper describes a test where the AI had to figure out the reading order of ancient, messy Japanese letters scattered across a page. Fugu-Ultra didn't just guess; it wrote a custom program to solve the puzzle, improved it, and solved it better than any single model could.
Design: It even designed a working mechanical camera iris (the part that opens and closes in a lens) that actually moved correctly, while other models created designs that were broken or didn't work.

4. Why This Matters: The "Orchestra" vs. The "Soloist"

The main point of the paper is that we don't necessarily need to build one single, massive, super-expensive brain to get better AI. Instead, we can build a system that knows how to use the brains we already have.

The Old Way: Try to make one model so big and smart that it can do everything perfectly.
The Fugu Way: Keep the specialized experts (the math guy, the coding guy, the science guy) and hire a smart manager (Fugu) to coordinate them.

The paper claims this approach allows them to achieve "state-of-the-art" results—meaning the very best scores on difficult tests—without needing to train a new, massive model from scratch. It turns a collection of specialized tools into a single, incredibly capable system.

In short: Sakana Fugu is a smart manager that knows exactly which AI expert to call for the job, whether it's a quick answer or a complex, multi-step puzzle, resulting in a team that is smarter than the sum of its parts.

Technical Summary: Sakana Fugu

Problem Statement

Frontier Large Language Models (LLMs) have achieved expert-level performance across diverse domains, yet they exhibit increasingly differentiated specializations. For instance, while some models excel at mathematical reasoning or theorem proving, others dominate in software engineering, cybersecurity, or algorithmic implementation. Furthermore, within single domains, models often possess fine-grained strengths (e.g., direct algorithm implementation vs. complex planning).

Current approaches to scaling AI capabilities rely primarily on training larger, more expensive models. However, the paper argues that the next frontier may not be achieved by a single monolithic model, but by systems that can identify, combine, and amplify the complementary strengths of multiple specialized models. Additionally, while "agentic scaffolds" (structured prompting, tool use, environment feedback) have enhanced individual model performance, these are often fixed, task-specific systems. There is a need for a system that can dynamically orchestrate a team of agents, adapting the workflow, communication topology, and tool usage to the specific requirements of each query, effectively treating collective intelligence as a scalable axis beyond raw model size.

Methodology

Sakana Fugu is a family of learned orchestrator models designed to expose a multi-agent system through a single model interface. Instead of requiring users to design workflows, Fugu models dynamically construct "agentic scaffolds" over a pool of frontier LLM workers. The system releases two variants optimized for different operating points: Fugu (latency-aware) and Fugu-Ultra (performance-prioritized).

1. Fugu: Balancing Performance and Latency

Fugu is designed for interactive use, selecting a single worker per input to minimize latency while maximizing the chance of selecting the most capable agent.

Parametrization: Fugu uses a pre-trained language model backbone with a lightweight selection head attached to the final hidden layer. Unlike previous coordinators that assign roles, Fugu's head outputs logits for $L$ worker models, selecting one to act as a worker. To adapt the backbone efficiently without full fine-tuning, the system employs singular-value fine-tuning, training only the singular-value scales of selected parameter matrices.
Training Paradigm:
- Supervised Fine-Tuning (SFT): The model is trained on single-step tasks (coding, math, reasoning). Labels are generated by running all worker models on a task and ranking them based on performance against ground truth. The orchestrator is trained to predict a soft probability distribution over workers (via softmax) rather than a hard single label, providing a richer signal.
- Evolutionary Optimization: Following SFT, Fugu is optimized using sep-CMA-ES (separable Covariance Matrix Adaptation Evolution Strategy) on end-to-end multi-turn tasks (e.g., coding assistants with tool use and environment feedback). This stage optimizes the expected terminal reward of the interaction trajectory, allowing the model to learn practical capabilities like tool reliability and iterative debugging that static benchmarks miss.

2. Fugu-Ultra: Prioritizing Performance

Fugu-Ultra is designed for complex, multi-step workloads where answer quality supersedes latency. It builds upon the Conductor framework, which generates natural language agentic workflows.

Workflow Generation: The orchestrator outputs a sequence of workflow steps, defining subtasks, assigning specific worker agents, and specifying access lists (which previous outputs to include in the context). This allows for arbitrary topologies, from sequential chains to parallel trees.
Training: Trained using GRPO (Grouped Relative Policy Optimization) on a reward function that penalizes formatting errors and rewards correct task completion.
Advanced Memory & Isolation:
- Intra-workflow Isolation: To prevent "orchestration collapse" (where the first agent's trajectory biases all subsequent agents), agents within a single workflow are isolated. They only see context explicitly permitted by the access list, preserving their independent solution paths.
- Persistent Shared Memory: Agents retain memory of interactions across different workflows (multi-turn conversations) to avoid redundant tool calls and maintain context, while remaining isolated within the current workflow.
Agent Pool: The system orchestrates a diverse pool of frontier models (e.g., Gemini-3.1-Pro, Claude-Opus-4.8, GPT-5.5).

Key Contributions

Learned Orchestration as a Scaling Axis: The paper proposes model orchestration as a complementary scaling axis to increasing model size. It demonstrates that combining existing specialized models can yield performance exceeding the best individual models without requiring new training runs or parameter access.
Dynamic, Query-Adaptive Scaffolds: Unlike fixed agentic frameworks, Fugu models learn to dynamically devise scaffolds (routing, topology, communication strategies) specific to each query. This includes selecting the optimal aggregator (e.g., using a math-specialist model for aggregation in math-heavy tasks and a fact-retrieval specialist for trivia).
Production-Ready System: The report details the infrastructure and design principles (e.g., singular-value fine-tuning for efficiency, evolutionary strategies for robustness) that turn multi-agent research concepts into a production system capable of handling real-world interactive tasks.
Two-Tier Release: The introduction of Fugu (low latency, single-agent selection) and Fugu-Ultra (high performance, multi-agent workflows) provides practical options for different use cases.

Results

The Fugu models were evaluated on a suite of challenging benchmarks, often outperforming state-of-the-art frontier models and proprietary "preview" models.

Agentic Coding: Fugu-Ultra achieved state-of-the-art results on SWE-Bench Pro (73.7%) and Terminal Bench 2.1 (82.1%), surpassing the next best public models by 5-6 percentage points. It demonstrated fine-grained adaptivity, alternating between GPT-5.5 for building and Claude-Opus-4.8 for debugging.
Scientific Reasoning: Both Fugu and Fugu-Ultra achieved 95.5% on GPQA-Diamond, surpassing Gemini-3.1-Pro (94.3%) and GPT-5.5 (93.6). The system successfully routed math-heavy subtasks to GPT and science-heavy tasks to Gemini.
Other Benchmarks:
- LiveCodeBench Pro: 90.8% (Fugu-Ultra) vs 88.4% (GPT-5.5).
- CharXiv Reasoning: 86.6% (Fugu-Ultra) vs 84.1% (GPT-5.5).
- Humanity's Last Exam: 50.0% (Fugu-Ultra), outperforming all baselines.
Qualitative Tasks:
- AutoResearch: In autonomous ML training optimization, Fugu-Ultra achieved lower validation bits-per-byte (0.9774) than single-model baselines, demonstrating superior consistency and peak performance in long-horizon search.
- Classical Japanese Reading: Fugu-Ultra achieved a mean Normalized Edit Distance (NED) of 0.776 on recovering reading orders of scattered classical Japanese letters, significantly outperforming the best frontier baseline (0.642).
- CAD Generation: Fugu-Ultra consistently produced mechanically valid and complete CAD designs (e.g., camera iris mechanisms) where other models produced incomplete or structurally weak results.

Significance and Claims

The paper claims that Sakana Fugu represents a shift in how frontier AI capabilities can be accessed and scaled.

Democratization of Capability: By composing models at the behavioral level rather than the parameter level, Fugu can incorporate closed-source frontier models and heterogeneous providers without requiring access to their weights. This makes frontier-level capabilities more modular and accessible.
Economic and Geopolitical Implications: The authors suggest that treating orchestration as a scaling axis could distribute the benefits of frontier AI more broadly, allowing organizations to achieve high performance without the massive compute resources required to train the largest models.
Collective Intelligence: The work validates the hypothesis that "collective intelligence" is an emergent property of coordinated model collectives. It demonstrates that a system can engineer software like the strongest coder, reason mathematically like the strongest mathematician, and dynamically decide when to deploy each, effectively creating a "super-agent" from existing components.

The authors position Fugu as a first step toward a future where AI progress relies not just on training larger models, but on learning how specialized models can collaborate through dynamic, query-adaptive scaffolds.