Networked Intelligence: Active Shared Context Graphs for Human-AI Team Science
This paper introduces Mycelium, an active shared context graph system that cultivates networked intelligence by automatically routing scientific observations and hypotheses between diverse human and AI team members, thereby transforming isolated findings into collaborative mechanistic constraints and experimental designs.
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 science as a massive, bustling city where every expert is a specialist living in their own neighborhood. One group of scientists lives in the "Protein District," another in the "Genetics Grove," and a third in the "Chemistry Canyon." For a long time, these neighborhoods were separated by tall, silent walls. If a scientist in the Protein District found a weird clue, they would write it in their notebook, but the scientist in the Genetics Grove might never see it until months later, if at all. This is how most "AI for science" works today: it tries to build a super-intelligent robot that can read everything at once, hoping one giant brain can solve the whole puzzle alone. But real scientific breakthroughs often happen when different experts connect the dots that no single person could see. The big question is: how do we build a digital system that doesn't just make one brain bigger, but actually connects all these different brains and tools together so they can talk to each other instantly?
This is where a new system called Mycelium comes in. The researchers at Pacific Northwest National Laboratory built this system to act like the underground network of a giant mushroom. Just as a mushroom's roots (mycelium) connect different parts of a forest to share nutrients and signals, Mycelium connects different scientists and AI agents. Instead of trying to force one AI to know everything, Mycelium creates a "shared living graph"—a digital space where every observation, hypothesis, and experiment is a node. When a scientist in one neighborhood makes a discovery, the system automatically checks if it's relevant to someone in another neighborhood and "routes" the information to them. It's like having a magical messenger that knows exactly who needs to hear a specific piece of news, instantly turning a local finding into a team-wide breakthrough.
In their latest test, the team put Mycelium to work on a tricky biological problem: figuring out why a specific type of bacteria (Pseudomonas putida) was accumulating a chemical called gluconate instead of making what they wanted. They set up a "sprint" where three human experts—one focusing on proteins, one on genetics, and one on physical outcomes—worked alongside AI agents. They all used a standard chat interface, but behind the scenes, Mycelium was busy connecting their dots. The system took a confusing clue from the genetics expert (that certain genes weren't turning on) and instantly routed it to the protein expert. This connection helped the protein expert realize their data wasn't just "noisy"; it was actually confirming the genetic clue. Later, a physical measurement from the third expert helped the team realize they needed to change their entire experimental design. The result? The team converged on a new, detailed plan for how to fix the bacteria's metabolism, a plan that included specific genetic edits and chemical tests.
To prove this "networked" approach was actually better than just having a smarter robot, the researchers ran a head-to-head comparison. They gave a single, super-powered AI agent the exact same data and asked it to solve the problem alone, without any help from a team or a shared graph. They also tried a version where they told the AI to "act like a team," but without the actual network structure. The results were clear: the single AI agents were smart, but they missed the big picture. They found fewer clues and didn't connect the dots across different fields as well as the human-AI team did. The networked team surfaced 25 distinct scientific findings and actionable plans, while the single AI agents only found 17 and 18, respectively. More importantly, the networked team generated 4 fully "actionable decision rules" (specific instructions for what to do next), compared to just 2 and 3 for the solo agents.
The paper suggests that the real power isn't just in making AI models bigger or smarter; it's in building the right architecture to connect them. The authors found that while a single AI can reason deeply about one topic, it struggles to explore the vast, messy space of a real scientific problem where different experts see different things. By using Mycelium to route information asynchronously—sending the right clue to the right person at the right time—the team was able to compress what usually takes months of back-and-forth into a single, rapid sprint. The system didn't just store data; it actively managed the team's evolving understanding, ensuring that every new idea was tied back to the evidence that created it. This approach, which the authors call "networked intelligence," shows that the future of science might not be about one super-brain, but about a smart, connected web of many minds working together.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.