Over the past few years I have had the chance to work across a fairly wide range of collaborative settings - competitive hackathons, academic research groups, industry internships, a data study group at the Alan Turing Institute. The contexts were different, the problems were different, the people were very different. But one thing kept showing up in roughly the same form.
The teams that did the most interesting work were almost never the ones with the deepest shared expertise. They were the ones where people came in with genuinely different mental models, different instincts about what mattered, and different vocabularies for describing the same thing. That combination - uncomfortable as it can sometimes feel - consistently produced better questions, broader solution spaces, and results that none of the individuals involved could have reached working within their own discipline alone.
I want to say something more specific than “diversity is good,” because the vague version of that argument is easy to agree with and easy to ignore.
Three settings, the same pattern
The pattern becomes more convincing when it holds across settings that are structurally quite different.
A competitive hackathon with a tight deadline and a hard problem - predicting which biological pathways novel compounds would affect, using high-content imaging data. The team included people from computational chemistry, software engineering, and biology. The solution we built - using transformer encoders on molecular representations alongside image embeddings - came from a conversation between someone who thought about molecules and someone who thought about language models, mediated by someone who understood what the biology actually needed to tell us. No single person had all three pieces.
A similar format, a different problem. Predicting how mutations affect protein function using deep mutational scanning data. The team spanned protein biochemistry, deep learning, and statistics. The biochemist knew which parts of the sequence space were biologically meaningful. The statistician kept asking whether what the model was learning was actually what we thought it was learning. The ML work sat in the middle, depending on both. The result was a model that generalised well precisely because it was built with those constraints in mind from the start.
This one was different in character - no competition, no winning team, just a group of researchers brought together to work on a hard problem over a week. Without a competitive frame, the collaboration became more openly exploratory. People were more willing to say they did not understand something, more willing to follow a line of thinking into territory they were not expert in. The output was a published report, but what I remember most is the quality of the conversations - particularly the ones that happened when someone from outside the domain asked a question that the domain experts had stopped asking because they assumed they knew the answer.
Three different settings, three different problems, three different team compositions. The same underlying dynamic: the most useful contributions often came from people who were not the deepest experts in the relevant field.
What diversity of background actually does
It breaks down jargon
Every field has terminology that carries compressed meaning - meaning that can quietly distort how you think about a problem. When someone asks you to explain something without the jargon, you often find the concept is less solid than you thought, or that the jargon was obscuring an assumption worth questioning. This is mildly uncomfortable and consistently useful.
It widens the solution space
A biologist looking at a classification problem will reach for one set of tools. A software engineer will reach for another. A statistician will question whether it is actually a classification problem at all. Three different framings of the same problem is more valuable than three people with identical training each arriving at the same framing independently - even if reconciling those framings takes more time upfront.
It produces better questions
The questions that come from genuine confusion - from someone who does not understand why you are approaching something a particular way - are frequently the ones that matter most. Not because naivety is a virtue, but because accumulated expertise builds accumulated assumptions, and those assumptions are often invisible from inside the field.
“The most useful question I heard in three days of a hackathon came from a software engineer who had never worked in biology. He asked why we were treating the problem as single-stage rather than decomposing it. It was not a biological insight. But none of the biologists had thought to ask it.”
The bridge-builder is the most underrated role in science
In every one of these settings, a significant part of what made things work was not just the diversity of expertise, but the presence of people who could move between domains - who had enough of a foundation in more than one area to translate, to hold context across multiple conversations, to know when two people were actually agreeing without realising it and when they were genuinely at odds.
These people are often hard to place in a traditional academic hierarchy. They do not have the deepest publication record in any single area. They are not the most cited person in the room. But they are frequently the reason the room functions. They carry the thread of a conversation across domain lines. They notice when an assumption from one field is about to cause a problem in another.
It is a kind of intellectual generosity: a genuine interest in what other people know, and a willingness to do the work of understanding it rather than waiting for it to be translated into your own terms. Combined with enough breadth to actually follow the conversation, that quality is consistently valuable in ways that are hard to quantify and easy to overlook in hiring.
Why we keep building teams that look the same anyway
Shared expertise feels safer. A team of specialists who all speak the same language is easier to manage, easier to evaluate, and easier to explain to a funder. The risk is legible. The counterfactual - what might have been discovered by a broader team - is not.
The incentive structures in academia push against it. Grants, publications, and career progression all reward depth. A researcher who spends time building bridges between disciplines is doing something genuinely valuable that the system has limited ways of recognising. That does not make it wrong - but it does make it harder to sustain.
Hiring is also hard across domain lines. It is difficult to evaluate someone whose expertise is partly outside your own, which means interdisciplinary candidates often get filtered out early in processes designed to find domain specialists. The people who are hardest to evaluate are sometimes exactly the people a team needs most.
That is what I think interdisciplinary teams are actually for. Not just to produce better results, though they often do. But to generate the kind of thinking that only happens when people who see the world differently are given a problem and asked to work on it together. That thinking is harder to measure than a citation count or a winning model. It is also, I suspect, where most of the important ideas come from.