For the strategist, analyst, or policy maker drowning in data, the era of the autonomous research agent has arrived. Tokyo-based Sakana AI has launched ‘Sakana Marlin’, positioning it not as a conversational assistant, but as a virtual Chief Strategy Officer. This B2B tool is designed to ingest a single topic and autonomously execute the research, synthesis, and verification required for high-level decision-making, effectively compressing weeks of manual labour into hours.
Unlike standard chatbots that respond in seconds, Marlin operates on a different timescale. A single session can run for up to eight hours, issuing hundreds to thousands of LLM queries to construct a comprehensive output. The result is a structured dossier comprising a long-form report and a presentation slide deck, generated via image-generation AI.
What is Sakana Marlin
Marlin functions as an enterprise-grade research engine rather than a chat interface. The user provides a prompt, and the system independently formulates hypotheses, browses sources, and validates findings. The output is tailored for executives, featuring a main body, references, and appendices.
The scale of the work is significant. While the Japanese announcement mentioned reports of dozens of pages, the English release specifies outputs of up to 100 pages, citing between 60 and 80 sources in recent demonstrations. The system underwent a closed beta in April 2026, where approximately 300 professionals tested the tool on real-world tasks including strategy formulation, market research, risk analysis, and competitive intelligence. Sakana has since secured strategic investment from Citigroup and partnered with MUFG.
Inside AB-MCTS: Wider or Deeper
The engine driving Marlin is AB-MCTS, or Adaptive Branching Monte Carlo Tree Search, derived from Sakana’s earlier research titled “Wider or Deeper? Scaling LLM Inference-Time Compute with Adaptive Branching Tree Search”. This algorithm treats complex reasoning as a tree-search problem, making a decision at every step to either expand the search space or refine a specific path.
Standard sampling methods typically generate multiple parallel answers and hope for the best. AB-MCTS is more surgical. It can go ‘wider’ by generating a new candidate answer or ‘deeper’ by refining an existing promising one. A multi-LLM variant further enhances this by routing steps to different models entirely.
In reported ARC-AGI-2 experiments, this collaborative approach yielded superior results. A combination of o4-mini, Gemini 2.5 Pro, and DeepSeek-R1 solved approximately 27.5% of tasks, compared to 23% when using the o4-mini model in isolation. Marlin applies this same adaptive logic to long-horizon research projects.
The second pillar of Marlin is workflow automation technology from Sakana’s “AI Scientist” project, which demonstrated autonomous scientific discovery and was published in Nature.
Interactive demo: The embedded widget (marlin-abmcts-demo.html) visualises the “wider or deeper” decision process. Pressing Run allows you to watch the search tree expand. Nodes with higher scores appear greener, and the optimal path is highlighted. Toggling “Multi-LLM” reveals how steps are routed across different models.
AB-MCTS: “Wider or Deeper?”, interactive search
Search state
Decision log
high score
best path
Key takeaways
- Sakana Marlin shifts the paradigm from instant chat responses to deep, autonomous research sessions lasting up to eight hours, delivering 60–100 page reports with 60–80 citations.
- The core AB-MCTS algorithm dynamically decides whether to expand the search breadth or refine existing hypotheses, outperforming standard parallel sampling methods.
- Multi-LLM collaboration within the search tree proved effective in testing, solving roughly 27.5% of tasks compared to 23% for a single model.
- Backed by Citigroup investment and partnerships with MUFG, the tool has been stress-tested by 300 professionals on complex strategy and risk analysis tasks.
