Primer

The Diversity Problem: Institutions, Science, and What Makes Research Possible

David Hilmer Rex · June 2026

Summary

Research ecosystems tend to converge on a single dominant institutional form — the university — regardless of whether it fits the problem at hand. This is not a failure of policy but the predictable outcome of how funding, career paths, and reputational systems have co-evolved over decades. The result is systemic fragility: an ecosystem that cannot adequately address problems whose structural requirements don't match the university model. This primer examines what external conditions historically produce institutional diversity, what makes different forms viable at a given moment, and what the current wave of new institutions — Focused Research Organisations, AI labs, ARPA-style agencies — requires to persist rather than converge back toward the dominant model. The larger stake is adaptive capacity: whether a research and innovation system can reconfigure quickly when its external environment delivers a shock, a discontinuity, or a problem no existing form is shaped to meet — the forces mapped in the companion ecosystem forces primer.

On institution–problem fit ↑

The fit between a research problem and its institutional home is a perennial question, not a discovery of the last decade. Churchill's observation that "we shape our buildings and afterwards our buildings shape us" applies to institutions as much as to architecture: the structures we create for science determine which problems get attacked, by whom, on what timelines, and with what tolerance for failure. What is less often acknowledged is that the relationship runs in both directions. External conditions — new technology paradigms, geopolitical pressures, economic structures, the problems a society urgently needs to solve — determine which institutional forms become viable, and when.

The history of great research organisations is less a story of deliberate design than of precipitation: structures called into being by specific external circumstances, sustained by those circumstances, and declining or transforming when they changed. Bell Labs required AT&T's monopoly revenues and Cold War urgency. DARPA required Sputnik shock and a government willing to route research funding outside the university system. The Danish enterprise foundations required a particular confluence of commercial law, social trust, and industrial structure that accumulated over more than a century. None of these were engineered from first principles. They were the forms that specific historical conditions made possible — and when those conditions changed, the institutions changed with them.

This points to a harder question than "what should we build?" The contemporary institutional reform conversation tends to frame the problem as one of design: identify the ideal structure, then construct it. But design capacity is only part of what determines whether institutional diversity persists. The deeper question is what conditions allow different institutional forms to coexist over time, so that when external demands shift — new problem types emerge, geopolitical configurations change, new technology paradigms arrive — the research ecosystem has the variety to respond. An ecosystem that has converged on a single dominant form cannot do this. The goal is not a single better model. It is a research and innovation system that can change fast when the world does — one that holds enough institutional variety, and can generate new forms cheaply enough, to respond when an external novelty arrives that no existing form is shaped to meet. This is Ashby's law of requisite variety applied to science: a system can only absorb as much novelty as it has internal variety to match. Maintained diversity is the means; adaptive capacity is the end.

Nadia Asparouhova, tracing the history of tech-native approaches to science funding, frames the core problem precisely: scientific progress is fundamentally a coordination problem.¹ That shift in emphasis — from talent selection and institution design toward ecosystem conditions — is closer to what the diversity-maintenance frame requires than most of the institutional design literature acknowledges.

Michael Nielsen, whose definition of metascience as the exploration of new social processes for science frames much of this resource, has written extensively about how the structure of scientific institutions shapes what scientists can and cannot do — how publication norms, funding cycles, career incentives, and collaboration tools collectively determine the frontier. His ongoing Science++ project with Kanjun Qiu treats this as a core question: not just what science produces, but what kinds of science different organisational forms make possible.

Adam Marblestone and Sam Rodriques gave one version of the institution-problem mismatch a precise name and a proposed solution. Their 2020 paper introduced the Focused Research Organisation as a formal concept: a structure matched to problems requiring sustained, team-based, tool-building work in the gap between what a single academic lab can do and what large industry or government will fund. Marblestone's organisation, Convergent Research, identifies those mismatches and builds FROs to address them — including through the Fundamental Development Gap Map, an open attempt to catalogue the structural gaps in science that new institutional forms might fill.

Ben Reinhardt has developed the most detailed anatomy of what makes research institutions work. His essay Why Does DARPA Work? maps the specific design choices that make DARPA unusually productive; his FAS paper makes the structural case for funding organisations rather than projects; and his Unbundle the University project asks which functions of the university belong together and which should be disaggregated. His organisation, Speculative Technologies, applies that analysis in practice.

Samuel Arbesman's essay Why Do Research Institutes Often Look the Same? (2026) poses the question that sits at the centre of this reframing. Despite the diversity of problems science faces, research organisations tend to converge on similar structures over time. Isomorphism — the pull toward resemblance regardless of stated purpose — is a more powerful force than deliberate design in most cases. Understanding what drives convergence, and what structural features resist it, is one of the live questions underneath the institutional reform conversation.

Patrick Collison and Tyler Cowen gave the broader intellectual project its name. Their 2018 diagnosis of diminishing scientific returns — more scientists, more funding, fewer transformative breakthroughs per dollar — and their 2019 call for a new science of progress have given the institutional reform conversation its broadest intellectual home. Progress studies asks not just how to fix academia but what conditions — organisational, cultural, economic — produce rapid discovery at all. The Atlas of Innovation maps the funding mechanism side of this, showing how "different funding approaches can radically reshape what gets invented, how quickly, and at what cost."

One significant gap in this conversation is that it has been almost entirely STEM-centric. Geoff Mulgan — professor at UCL, former CEO of Nesta, author of Exploratory Social Science (OUP, 2026) — has argued that the structural problems the institutional reform conversation identifies apply with equal or greater force to the social sciences and humanities. Social science, he argues, has become "overwhelmingly oriented towards diagnosis" with almost nothing invested in radical prescription and design — the work of imagining what institutions, democracies, and care systems could look like in twenty years. The ratio of analysis to futures work is approximately 100 to 1.² The institutional reform conversation has built FROs for neuroscience and genomics; nobody has built the equivalent for democracy design or social care.

What these writers share is a conviction that organisational form is not neutral — that institutions are not merely containers for research but active forces shaping what research is possible. The fuller version of that conviction is that no form is right for all time: the question is always one of fit between institutional structure, problem type, and the external conditions that make certain forms viable. The sections below try to map what that looks like in the current moment.

The problem ↑

The post-war research settlement — government-funded, investigator-led, university-based, peer-reviewed — encoded one institutional answer to the question of how to organise science. Over the subsequent decades, through the ordinary dynamics of institutional consolidation, that answer became effectively the only answer. Funding channels concentrated on academic institutions. Career paths narrowed through the PhD-postdoc-faculty track. Reputational systems calcified around journal publication. The alternative institutional forms that once competed with universities — corporate research labs, independent institutes, government arsenals — declined or were absorbed. The result is a near-monopoly operating across four levels, as Reinhardt's Unbundle the University identifies: physical space, funding, mindsets, and research structure, all routing almost all pre-commercial research through universities regardless of fit.

The problem is not that universities are bad at research. They are extraordinarily good at certain kinds of it: open-ended inquiry, training, rapid diffusion of results through publication and teaching. The problem is that routing everything through one institutional form creates systemic fragility. Different problems have different structural requirements. Some require sustained multi-year team effort on a single defined output. Others require rapid iteration between research and engineering at a scale no academic lab can sustain, or the capacity to build tools and platforms the whole field can use rather than publish incremental papers. An ecosystem with one dominant template will systematically underperform on problems whose requirements don't match that template — not because anyone decided to neglect them, but because the institutional machinery will not naturally produce the work. Expecting a university to cover all of this is like "expecting every coffee shop to also include a laundromat, a bookstore, and a karaoke bar." Crucially: "changes to the research ecosystem are bottlenecked by where the work is done."³ Most reform efforts — new journals, different funding mechanisms, new institutes embedded in universities — still route through academic institutions because that is where the hands are. Structural change requires changing the location of work, not just the rules around it.

External shocks periodically create the conditions for institutional diversity to re-emerge. Cold War competition produced DARPA and the national laboratories. The AIDS crisis reshaped NIH's mandate and generated new advocacy-driven research structures. The biotech revolution produced university spinouts and the venture-backed drug development model. The current moment has its own pressures: the convergence of AI with biological and materials sciences, the climate transition requiring solutions R&D at unprecedented scale and speed, geopolitical competition once again making governments willing to fund research outside normal academic channels. Each of these generates demand for institutional forms that the standard university model cannot easily supply.

This is the context in which the contemporary institutional reform conversation makes most sense — not as a deliberate design project to replace a broken system, but as a response to external conditions that are once again making institutional diversity viable, and worth building. The question is not only what new structures to construct, but what conditions need to hold for those structures to persist rather than converge back toward the dominant model over time.

The historical models — Bell Labs, DuPont, Xerox PARC — are instructive for how they ended as much as for how they worked. They succeeded while specific conditions held: monopoly capital, high-conviction technology bets, and existential threats forcing management to tolerate the diffusion of results rather than capture them. When those conditions changed, the institutions changed. The new institutions movement is attempting to reconstruct, through deliberate philanthropic effort, conditions that historically required external precipitation. That is harder than it looks. As Reinhardt and Recht have put it: "There is no best structure, but there are many bad compromises."⁴ The question is whether institutions designed against the grain of prevailing conditions can sustain the structural commitments that make them work — and what external conditions need to accompany that effort.

A map of institutional forms ↑

The institutional forms that have emerged or been deliberately revived in the current moment are best understood as responses to specific external conditions — each suited to particular problem types and particular funding structures — rather than as a menu of optimal designs to be selected and assembled. No single form is best for all problems or all contexts. What the taxonomy below offers is a map of what each form can and cannot do, and a basis for asking which external conditions are required to sustain each. The value of maintaining several forms simultaneously, rather than converging on whichever seems most promising right now, is the underlying argument of this primer.

Lowering the cost of trying ↑

There are two ways to read the argument for institutional diversity. One treats it as a problem of stock: a set of valuable forms to be designed well and then maintained against the pull of convergence. That is the frame most of this primer has used, and most of the field uses. The other treats it as a problem of flow: diversity as the standing result of a high rate of experimentation, continuously replenished, where what matters is less the quality of any single new form than how many attempts the ecosystem can afford to make.

The flow frame points at a different lever. In an evolutionary system, variety does not come from designing better organisms; it comes from making variation cheap and selection real. Applied to institutions, the question is not only "what is the right structure?" but "what does it cost to attempt a new one, and what happens to the attempts that fail?" If the fixed cost of one experiment is high — a multi-year fundraise, a new legal entity, a full-time founding team, a career risk for everyone who joins — then only a handful of well-capitalised, expert-led attempts will ever be made, and the ecosystem's diversity is capped by the supply of those rare events. If the fixed cost is low, many more people can try, most attempts can fail cheaply, and the good ones can compound.

This reframes what may be the practical heart of the diversity problem: if a few people have an idea for a new way of doing science, how easily can they pursue it? The barrier splits in two. The first is imagination — knowing the design space exists at all. This is partly what a map like this one is for: a legible menu of forms, worked examples, and playbooks such as TIAL's lower the cost of conceiving of an alternative. The second is launch — the transaction cost of actually standing the thing up. That is what the enabling-infrastructure layer above addresses: fast, lightweight funding so the first dollar does not require a nine-month grant cycle; fiscal sponsorship so a project need not become a corporation; regranting so trusted people can back ideas from their own networks; open dissemination so results are not held hostage to the journal system.

The constructive implication runs against the more fatalistic conclusion this primer reaches elsewhere — that durable institutions are precipitated by external conditions effort alone cannot reproduce. That is true of the largest forms. But it is not the whole story, because the rate at which small experiments can be attempted is not given by history; it is set by how much activation energy each one requires, and that is a variable we can deliberately lower. We may not be able to manufacture the conditions that produced Bell Labs. We can make it dramatically cheaper for the next thousand people with an idea to find out whether it works — and an ecosystem that runs many cheap experiments and lets the failures die gracefully may prove more diverse, and more resilient, than one that places a few large, carefully designed bets.

The European context ↑

The institutional reform conversation has been largely US-led, which matters because the structural problems it addresses exist in Europe too — and in some respects more acutely. European universities are often more constrained by state employment law, indirect cost rules, and ministerial oversight than their US counterparts. The PI model is equally dominant; the pipeline to high-risk, long-horizon research is equally narrow.

Europe does, however, carry a different institutional history — one that complicates the framing of this as a problem of building new things. The Danish and Nordic enterprise foundations are not examples of successful new institutions; they are examples of successful old ones — forms that survived the post-war consolidation that absorbed most other alternatives into the university orbit. Their durability is the interesting feature. It points to a different set of questions from those the US conversation asks: not just how to build new institutional forms, but what legal structures, capital arrangements, and governance frameworks allow diverse forms to persist across generational timescales without being captured by the dominant model. That question is as important as the design question, and Europe is better positioned to answer it empirically.

ARIA is the most significant European development. Modelled explicitly on DARPA, it has an £800m budget over four years and has already co-created Meridial and Echo Labs — the first FROs outside the United States. Its early programme managers have set ambitious technical agendas across biology, materials, and computing. Whether ARIA can maintain its unusual autonomy as it matures, and as political conditions change, remains the open question.⁸

For smaller European countries — including the Nordic ones — the question is what scale of ambition is achievable. A full DARPA equivalent is usually waved away as too expensive, but that is the wrong objection. DARPA spends roughly $4 billion a year; the Novo Nordisk Foundation alone gave DKK 11.8 billion (€1.6 billion) in 2025, and pooled Nordic foundation capital — Wallenberg in Sweden, the Danish industrial foundations — could approach that scale without strain. (Norway's $2 trillion oil fund, often invoked here, is a red herring: it is mandate-bound to invest only abroad and to feed the state budget through the fiscal rule, not to underwrite domestic research.) Capital is not the binding constraint. What the Nordics lack is DARPA's mission customer — the military procurement budgets that pull prototypes into deployment and discipline the portfolio — and the absorptive scale of a continental innovation base. A foundation can supply the capital, and even run a private ARPA on the Reinhardt model; it cannot conjure a procurement customer or a continental talent pool. Both constraints bind less for the Nordics pooled — some 27 million people and a combined GDP near $2.2 trillion, on its own about the world's twelfth-largest economy — than for any single country, which is the logic behind pan-Nordic industrial efforts such as Nordic Compass and its industry-led defence and deep-tech tracks. A 2019 EU peer review of Denmark's own system, chaired by Christian Ketels, reached a parallel verdict from the inside: the binding issue was neither money nor science but the absence of a clear, deliberate, overarching strategic direction, a system organised to "minimise the need for coordination," and an under-used role for government as a "launching customer" for innovation. The realistic question is therefore less "can we afford a DARPA" than "can we build the demand-pull and cross-border coordination to use one." Until those exist, several smaller moves are the practical path:

• An ARPA-model unit within an existing research council — small enough to be politically defensible, large enough to fund a few genuinely unusual programmes
• Co-investment in FROs alongside US-based philanthropic funders, using an existing foundation's balance sheet as an anchor
• National prize programmes for defined technical problems, structured to attract non-academic teams
• Policy frameworks that make it easier for FROs to operate — IP ownership, employment flexibility, indirect cost reform

Denmark has specific assets here: the Novo Nordisk Foundation at a scale unusual for a country of five million, and a small but engaged science policy community with close enough ties to the UK and US ecosystems to follow how the FRO model is developing in practice.

What we're still learning ↑

The new institutions movement is young. Most FROs are less than five years old. ARIA has been operating for about three years. The field lacks systematic evidence on what works, and the organisations themselves are still developing their own operating models.

The open questions that matter most:

• Replication. Can the FRO model extend beyond biology and computing, where the "tool and platform" framing is most natural? [C]Worthy (ocean carbon) and Cultivarium (novel microorganisms) suggest expansion is possible, but the boundaries of the concept remain unclear.
• Systems research. Reinhardt's essay Whence Systems Research? identifies a category of work that falls through all three main institution types: integrative work that scales at O(N²) complexity because changes to one component propagate through an entire system. Academia rejects it because it produces too little publishable novelty per unit of effort. Companies avoid it because the ROI is unclear and the timeline long. Government struggles to justify it up accountability chains. This is a gap the current taxonomy of new institutions does not yet address well.
• Transition. As the Research Leaders' Playbook puts it: "Technology ultimately lives in people's heads — the more people who worked on the technology as part of the program you can enable to continue working on it afterwards, the more likely it is to succeed."⁹ FROs must plan not just for execution but for what comes after: who shepherds the work beyond the programme's lifetime. This is largely untested for most current FROs.
• ARPA inflation. Paschkewitz and Patt have argued that the proliferation of ARPA-labelled agencies risks producing the appearance of reform without its substance — that what makes DARPA work is not the label but specific structural commitments that are easy to describe and hard to sustain politically. The question is whether the new ARPA-model agencies can maintain those commitments as they age.
• Complementary small-scale models. Not every research gap requires a $50M FRO. Eric Gilliam's case for reviving the scrappier BBN-style contract-research firm (see the taxonomy above) is one example;¹⁰ the ecosystem of new institutions may need more variation in scale, not just in type.
• The STEM bias. The entire taxonomy above — ARPAs, FROs, venture studios, prizes — has been built around natural science and engineering. Mulgan's case for a social ARIA points to a structural gap: the same institutional failures that leave certain technology problems unfunded also leave pressing social design problems untouched — how care for the elderly might be organised in 2040, what democratic institutions could look like, how to design for disability. Career disincentives, peer review gatekeeping, and disciplinary policing punish exactly the kind of speculative, design-oriented social science that would address these. The new institutions conversation has not yet seriously engaged with this.
• Field builders. David Lang has argued for a missing category of builder — the field builder, who operates between academia, startups, and philanthropy to catalyse entire new research areas. "Money can be a vessel for belief," he writes — the role is not just grantmaking but the connective work of making it possible for a field to cohere at all.¹¹ The ecosystem of new institutions may depend as much on this catalytic human infrastructure as on the design of any particular organisation type.
• The talent question. New institutions only work if researchers are willing to leave traditional career tracks. That depends on incentive structures — tenure clocks, grant eligibility, reputational norms — that the new institutions movement alone cannot change.
• Scale. The current set of FROs and ARPA-model agencies is a rounding error relative to the overall global science budget. The question is whether these models change the mainstream over time, or remain a productive niche.
• Maintaining diversity against convergence. Arbesman's isomorphism problem is ultimately the most challenging open question. The forces that produce institutional convergence — accountability pressure, grant dependence, talent market competition, regulatory norms — operate continuously and will not stop because some new forms have been established. The current moment of institutional diversity is in part a response to external conditions. Whether it can be sustained when those conditions shift — through deliberate governance structures, legal frameworks, and capital arrangements that make diversity structurally durable rather than episodically renewed — is a question the field has barely begun to address.

These are not reasons for scepticism — every significant institutional reform starts as a rounding error. They are the questions that need watching as the field develops.

Critiques and counterpoints ↑

The new institutions conversation is largely self-reinforcing: a community of writers, funders, and builders who agree that the mainstream research system is structurally broken and that new organisational forms can help. That consensus is productive, but certain challenges don't receive the scrutiny they deserve.

ARPA inflation. The most developed internal critique comes from Paschkewitz and Patt, who argue that what differentiates competitive economies is not breakthrough capacity but diffusion capacity — the ability to adopt and spread innovations broadly. Proliferating ARPA-labelled agencies addresses the wrong bottleneck. The deeper problems — disconnected academic silos, the collapse of corporate research labs, a venture model that fails hardware and physical sciences — require different instruments: technology selection for modularity, human capital rotation programmes, and catalyst institutions modelled on NACA. The critique is not anti-reform; it is a challenge to get the diagnosis right before prescribing.

Failure invisibility. The people best positioned to share information about what isn't working are overwhelmingly incentivised to stay quiet. Stuart Buck and Ben Reinhardt's essay "The Paradox of Harmful Information" (Arena Magazine) identifies a market-for-lemons dynamic: major research programmes fail routinely, but insiders rarely discuss this publicly for fear of professional retaliation, while unreliable outside voices fill the vacuum.¹² This applies as much to FROs and ARPA programmes — where funders and founders have reputational stakes in claiming success — as to conventional academic labs. The new institutions movement's celebration of ambitious failure in theory needs to be matched by mechanisms for honest failure disclosure in practice.

The incentive rewriting problem. A more structural critique: philanthropic new institutions — FROs, private ARPAs, venture studios — operate within the existing career and incentive structure of science. If the underlying problems are that peer review selects for consensus, grant competition rewards conservative proposals, and career structures punish risk-taking, then building a better-funded lab with more autonomy addresses these only partially. The institutions protect the people inside them from bad incentives; they do not change the broader ecosystem those people eventually return to, or compete against for talent. This critique is weakest for institutions — like FROs with time-bounded, full-employment models — that genuinely redesign the relationship between a researcher and their work, and strongest for organisations that adopt new labels while preserving familiar incentive structures.

Accountability and democratic legitimacy. The new institutions movement has largely sidestepped the question of who decides what gets built. Programme manager discretion, independence from peer review, and philanthropic governance are features that make these institutions work — and they also concentrate decision-making in small groups of unusually well-connected people, mostly in the United States, without clear mechanisms of democratic accountability. Paschkewitz and Patt note a related issue: for this style of research investment to work at scale, it requires legitimacy structures that are not automatically exportable. The question of what should govern ambitious science investment — outside the peer review system the movement has largely rejected — remains underexplored.

Replication beyond biology. Most working examples of FROs and ARPA-model programmes come from biology, computing, and adjacent fields. Whether the model extends to materials science, chemistry, energy systems, or — as Mulgan argues — social design remains genuinely uncertain. The argument from institution–problem fit cuts both ways: if organisational form shapes what's possible, then models built for genomics and brain mapping may not transfer cleanly to other problem types without significant adaptation.

The design framing itself. Perhaps the deepest challenge to this conversation is the assumption that institutional diversity is primarily a design problem — that the right structures can be built if the right people set their minds to it. The historical record is more ambiguous. Most durable research institutions were precipitated by external conditions, not designed: what looks like successful institutional design in retrospect was often the successful exploitation of a particular historical moment. Bell Labs was made possible by a regulatory settlement; DARPA by a geopolitical shock; enterprise foundations by commercial law. The implication is uncomfortable: the external conditions that are currently making new institutional forms viable — surplus philanthropic capital, AI-driven tool-building opportunities, geopolitical competition — may not be indefinitely reproducible by effort alone. Sustained diversity requires not just good design but the maintenance of the conditions that make diverse forms worth building and possible to sustain.