Primer

Funding Mechanisms: How the Design of Research Money Shapes What Science Happens

David Hilmer Rex · June 2026 · Updated October 2026 · 10 min read

Summary

Most debates about science funding are about how much. Both how much and how matter: the balance between basic funding to institutions and competitive project money shapes careers and power across the whole system, and the mechanism that distributes competitive money quietly determines what questions get asked, who gets to ask them, and how much risk the system can bear. This primer covers the design of competitive mechanisms. It surveys the dominant model (project grants allocated by peer review) and its documented pathologies, then the main alternatives (lotteries, prizes and advance market commitments, person-based funding, the ARPA model, and fast low-overhead grants and regranting) before setting out the design dimensions that distinguish them, why funding mechanisms are so hard to evaluate, and the funding landscape specific to Europe and the Nordics.

Why the mechanism is the message ↑

A funding mechanism is a theory, made operational, of how good science gets selected: who decides, on what evidence, how quickly, and what they are rewarded for. Change the mechanism and you restructure the portfolio of work that gets done. This is the funding-side counterpart to the institution–problem fit argument in the diversity primer: just as no single institutional form suits every problem, no single allocation rule suits every kind of research. The danger is monoculture: routing nearly all money through one mechanism and inheriting all of its biases. Mechanism diversity helps, but mechanisms interact with careers, overheads and power, and changing one mechanism without those leaves the underlying problems in place.

The default and its pathologies ↑

Project grants awarded by ex-ante peer review are the global default, and their strengths are real: expert judgement, legitimacy, and accountability for public money. But three decades of metascience on the mechanism itself have documented serious weaknesses. When researchers re-ran the NIH process (43 reviewers scoring 25 proposals that had already been funded), they found essentially no agreement between reviewers, and no reliable ability to distinguish the merely good from the excellent.1

The mechanism is also expensive in the scarcest resource it has: researcher time. An Australian study found that preparing a single grant proposal took about 34 working days on average, and that one funding round consumed an estimated 550 working years of researchers’ time, with roughly one proposal in five funded, and more preparation buying no greater chance of success.2

The indicators it leans on also carry a measurable bias against the unconventional. Taking first-time combinations of cited journals as a marker of novelty, Wang, Veugelers and Stephan found that highly novel papers are more likely to become top-cited in the long run but less likely in the short windows that funders and committees actually use; they appear in lower-impact-factor journals and are more often cited outside their own field.3 The finding concerns bibliometric indicators: those with short citation windows penalise novel research. Evaluation that leans on such indicators inherits the bias and favours safe, incremental work over risky work, optimising the default for defensible consensus.

And concentrating money in the best-funded labs yields diminishing returns. Analysing NIH data from Michael Lauer, Wayne Wahls found that output per dollar peaks at about US$400,000 of annual support per investigator (a “sweet spot”) and falls away on either side, with each marginal dollar at twice that level returning only a quarter to two-fifths as much.4 Spreading the same budget across more labs would, on this evidence, buy more science than concentrating it, which makes the default look like one mechanism among several, with biases of its own.

The most active current reform inside the peer-review paradigm is distributed peer review (DPR), in which applicants review each other’s applications, spreading the assessment load across the applicant pool. The UK Metascience Unit at DSIT has run formal DPR pilots within UKRI calls; early results show up to a 65% reduction in assessment time, and UKRI is now rolling it out to some major funding calls across research councils. The official caution is worth registering: DPR is “not the holy grail”: it cuts overhead without resolving the prior question of whether distributed assessment actually improves which proposals get funded. A planned follow-on survey will ask whether, how and why researchers use AI in their work.5 UKRI followed on 10 September 2026 with an eight-measure overhaul of grant assessment, due to take effect in October 2026, meant to speed decisions and respond to generative AI’s effect on both grant writing and peer review. It is part of a July 2026 pledge to halve grant-processing time by 2031.6

Funding by lottery ↑

If peer review cannot reliably rank the proposals above the funding line, one honest response is to stop pretending it can. Funding lotteries screen proposals for eligibility and quality and then randomise among those that qualify. The aim is to bound judgement: cut the reviewing burden, remove the false precision of fine-grained ranking, and give unusual or risky ideas a fairer shot. The standard design is a modified or partial lottery: peer review filters out the clearly weak, and sometimes fast-tracks the clearly outstanding, while chance decides the contested middle.7

This is no longer hypothetical. The Health Research Council of New Zealand allocates its Explorer Grants by lottery after a quality screen; the Volkswagen Foundation, the Swiss National Science Foundation, the Austrian Science Fund and the British Academy (for the British Academy/Leverhulme Small Research Grants) have all added randomisation to at least one funding line. The early evidence is encouraging: when the Volkswagen Foundation ran a lottery before peer review in one programme, funded projects from female applicants rose by 23% and the estimated cost of the whole allocation process fell by 68% against conventional single-stage review.8 A live scheme shows the effect on who even applies: after the British Academy switched the British Academy/Leverhulme Small Research Grants to a lottery above a quality bar in 2022, applications nearly doubled (from 643 to 1,257), and the share of awards to Asian and Asian British researchers rose from 11% to 17%, as researchers who had wrongly assumed they stood little chance against bigger-name institutions began to enter.9 Adoption keeps spreading: the Research on Research Institute has catalogued eleven partial-lottery trials since 2022, including at Wellcome and the Novo Nordisk Foundation.9 The evidence base is still young, and lotteries fit some funding lines (early-stage, exploratory) far better than others, but they are the clearest example of metascience changing how real money moves.

Paying for outcomes: prizes and advance market commitments ↑

Grants pay for effort: you are funded to attempt the work. Prizes and advance market commitments (AMCs) invert this and pay for outcomes: the reward is contingent on hitting a defined target, and the funder need not pick the winner in advance. This works when the destination is definable but the path is not, and it can attract participants (firms, startups, non-traditional teams) who would never respond to a research grant.

The landmark case is the pneumococcal AMC, launched in 2009 with US$1.5 billion from five governments and the Gates Foundation through Gavi, which guaranteed a market and is credited with accelerating vaccine supply for lower-income countries; Operation Warp Speed used related pull mechanisms at speed.10 The limits are equally clear: outcome-based funding works poorly for basic research where the destination is unknown, and rewards can be gamed when milestones are loosely specified.

Funding people, not projects ↑

A different move is to fund the person or organisation rather than the project: long-horizon, failure-tolerant support that removes the incentive to diversify across safe bets. The Howard Hughes Medical Institute funds on exactly this principle, summarised in its slogan “people, not projects.” When economists compared HHMI investigators with comparable NIH-funded scientists, the HHMI model (long, renewable appointments, currently seven years, and genuine tolerance of early failure) produced more high-impact, original work, and also more flops: a higher-variance portfolio.11 The study is a matched comparison without randomisation.

The same logic underlies the full-time employment model of Focused Research Organisations (see the diversity primer) and the block funding of contract-research institutes. The trade-off is the one that recurs across this resource: autonomy and long horizons buy ambition, at the cost of the legibility and accountability that project grants provide.

Mission-driven funding: the ARPA model ↑

Where lotteries and people-not-projects loosen central control, the ARPA model concentrates it: in a person. Modelled on DARPA, ARPA-style agencies hand an empowered, term-limited programme manager a budget and wide latitude to design a programme around a goal, actively solicit and hand-pick projects, and then actively manage the portfolio: extending and expanding what works, cutting what does not. The fixed term forces an exit strategy and injects urgency. The early evidence is promising (cleantech startups funded by ARPA-E filed patents at roughly twice the rate of comparable firms), though the model’s success appears to depend on scarce managerial talent and genuine institutional autonomy, neither of which is easy to reproduce on demand.12

The Franco-German Task Force on Breakthrough Innovation (June 2026) offers the most systematic recent account of what those conditions actually are.13 Drawing on DARPA and Germany’s SPRIND as its primary cases, it distils five invariants: empowered programme managers who make decisions themselves, mission-oriented and time-bound programmes, a genuine high-risk posture, connected science (tight links between the programme and the broader research community), and institutional independence from standard procurement and ministerial oversight. The task force also argues that France’s gap is institutional: France 2030 has already committed €54 billion, so the binding constraint is the institutional form and the mechanism that runs it. It recommends France establish a dedicated ARPA-type entity interoperable with SPRIND from day one.

Speed and the cost of the cycle ↑

A mechanism’s speed is itself a design choice with consequences. Conventional cycles run six to nine months from application to decision; for time-sensitive or exploratory work that lag is often fatal. Fast grants and regranting compress the cycle to days (a thirty-minute application, a near-immediate decision) by trading exhaustive review for trusted, low-bureaucracy judgement.14 The deeper point, developed in the enabling-infrastructure section of the diversity primer, is that the speed and overhead of the funding decision set the rate at which new ideas can even be tried.

The regrantor model generalises this. Instead of a committee, a funder hands a discrete budget to a trusted, well-networked individual who makes grants on their own judgement: fast, low-overhead, and able to surface opportunities a formal call would miss, including proactively starting projects. Philanthropic experiments such as Manifund and the FTX Future Fund’s regrantor programme have tested it. Fast Grants used a different design: a fast, centrally run programme with a short application form and decisions within 48 hours, which made more than 260 grants totalling over US$50 million for COVID-19 research in 2020. The open weakness is evaluation: regranting spends almost nothing on tracking outcomes, so its track record remains largely anecdotal.15

The design dimensions ↑

Most reform is a move along one of a few familiar dials. Five matter most: who decides (a committee, a programme manager, the crowd, or chance); speed; the unit of funding (project, person, or organisation); risk tolerance; and what behaviour the mechanism rewards. Naming the dials makes otherwise incomparable instruments easy to compare: a lottery and a fast grant both attack reviewing burden but move different dials; an FRO and an HHMI investigatorship both shift the unit of funding from project to organisation or person. One choice sits above all five dials: how much money flows as basic funding to institutions (in Denmark, the universities’ basisbevillinger) and how much through competitive calls.

Why this is hard to evaluate ↑

A caution runs underneath all of this: funding mechanisms are unusually hard to evaluate, because the value of research is heavy-tailed. A handful of breakthroughs dominate the total return, so what matters is how a mechanism affects the rare right tail, while randomised trials and conventional statistics are built to characterise the average grant. If outcomes are outlier-dominated enough, sample means and variances are barely meaningful, and two mechanisms can look identical across the bulk of grants while differing entirely in the tail that counts.16 This is partly why the field leans on case studies of extreme successes alongside controlled comparison, and why confident claims that one mechanism simply “outperforms” another deserve to be read with care.

2026: mechanism experiments go mainstream ↑

In 2026 governments began treating mechanism choice as something to test. The OSTP report Science: A New Golden Age (July 2026) recommended diversified funding and metascience units inside the science agencies.17 On 17 August NSF announced $1.5 billion for foundational research through 12 new funding opportunities and a portfolio-based approach, framing many calls as structured experiments it will assess and scale, alongside merit-review changes meant to surface unconventional proposals that consensus panels miss.18 For funders choosing among mechanisms, the Atlas of Innovation and the NBER framework behind it (see Further reading) are practical guides.

The European and Nordic angle ↑

Three features distinguish the European context. The unusual weight of private foundations as research funders, above all the Danish enterprise foundations, whose patient capital is examined in the enterprise-foundations primer. The EU framework programmes (Horizon), whose scale is matched by a heavy administrative load. And the coordination problem flagged by the 2019 EU peer review of Denmark: many capable funders, public and philanthropic, with little strategic alignment among them, a system-level inefficiency distinct from the design of any single mechanism, and one best addressed by pooling and coordination.

The Franco-German Task Force on Breakthrough Innovation (June 2026) makes this last point starkest: it recommends France create a dedicated ARPA-type entity explicitly interoperable with Germany’s SPRIND from launch, a direct proposal to shift from calls-for-proposals as the dominant mode toward a mechanism built around empowered programme managers and mission-oriented time-bound programmes. The report illustrates an observation that applies across European funders: large volumes of public money can coexist with an institutional gap, because the mechanism design and the institution that runs it are independent variables.

Notes ↑

  1. Elizabeth L. Pier et al., “Low agreement among reviewers evaluating the same NIH grant applications,” PNAS 115(12), 2018. Forty-three reviewers scored 25 R01 proposals that had already been funded; inter-rater reliability was statistically indistinguishable from zero, and reviewers did not reliably differentiate good from excellent applications. ↩
  2. Danielle L. Herbert et al., “On the time spent preparing grant proposals: an observational study of Australian researchers,” BMJ Open 3 (2013). A new proposal took ~38 working days and a resubmission ~28 (mean ~34); one NHMRC round consumed an estimated 550 working years of researcher time, ~21% of proposals were funded, and more preparation did not raise the chance of success. ↩
  3. Jian Wang, Reinhilde Veugelers & Paula Stephan, “Bias against novelty in science: A cautionary tale for users of bibliometric indicators,” Research Policy 46(8), 2017, 1416–1436. Novelty is measured as first-time combinations of cited journals; highly novel papers show delayed recognition, appear in lower-impact-factor journals, and are more often cited outside their own field. ↩
  4. On diminishing marginal returns to research funding, see Wayne P. Wahls, “The NIH must reduce disparities in funding to maximize its return on investments from taxpayers,” eLife 7:e34965 (2018), a single-author opinion piece analysing NIH data from Michael Lauer. Output per dollar peaks at about US$400,000 of annual support per investigator (the “sweet spot”); marginal returns at twice that level run roughly 25–40% of the peak, and the most heavily funded institutions tend to be less productive per dollar than more modestly funded ones. ↩
  5. Isaac Barbosa, “Distributed peer review ‘good but no holy grail’, DSIT official says,” Research Professional News, June 19, 2026. Reports up to 65% reduction in assessment time from UKRI DPR pilots; quotes DSIT Metascience Unit official on the mechanism’s limits and a planned survey of how researchers use AI in their work. ↩
  6. UKRI, “UKRI Modernises Grant Assessment for the Age of AI,” September 10, 2026. An eight-measure overhaul, effective October 2026, part of a July 2026 pledge to halve grant-processing time by 2031. ↩
  7. The two-stage “modified lottery”, a peer-review screen followed by randomisation among qualifying proposals, is set out in Ferric C. Fang & Arturo Casadevall, “Research Funding: the Case for a Modified Lottery,” mBio 7(2), 2016. Funders using partial randomisation include the Health Research Council of New Zealand, the Volkswagen Foundation (from 2017), the Swiss National Science Foundation, the Austrian Science Fund, and the British Academy; see also “Science funders gamble on grant lotteries,” Nature (2019). ↩
  8. “Lottery before peer review is associated with increased female representation and reduced estimated economic cost in a German funding line,” Nature Communications 16 (2025). In a Volkswagen Foundation funding line, a lottery-first procedure was associated with a 23% increase in funded projects from female applicants and an estimated 68% reduction in the cost of the allocation process relative to single-stage peer review. ↩
  9. Max Bennett, “When the best decision is no decision: the rise of randomization in grant funding,” Nature, 2026. Reports the SNSF tie-break model and the British Academy result (applications up from 643 to 1,257; share of awards to Asian and Asian British researchers up from 11% to 17%); the Research on Research Institute catalogues eleven partial-lottery trials since 2022, including Wellcome and the Novo Nordisk Foundation. ↩
  10. The pneumococcal Advance Market Commitment was launched in 2009 with US$1.5 billion committed by five governments (Italy, the United Kingdom, Canada, Russia, and Norway) and the Bill & Melinda Gates Foundation, administered through Gavi. ↩
  11. Pierre Azoulay, Joshua S. Graff Zivin & Gustavo Manso, “Incentives and Creativity: Evidence from the Academic Life Sciences,” RAND Journal of Economics 42(3), 2011. HHMI investigators, funded on long and failure-tolerant terms, produced more highly cited and more original work than a comparable group of NIH-funded scientists, and also more flops. The design is a matched comparison without randomisation. ↩
  12. Pierre Azoulay, Erica Fuchs, Anna P. Goldstein & Michael Kearney, “Funding Breakthrough Research: Promises and Challenges of the ‘ARPA Model,’” Innovation Policy and the Economy 19, 2019. The patenting evidence is from Anna Goldstein, Claudia Doblinger, Erin Baker & Laura Diaz Anadon, “Startups supported by ARPA-E were more innovative than others,” Nature Energy 5 (2020). ↩
  13. Nicolas Dufourcq, Rafael Laguna de la Vera et al., Report of the Franco-German Task Force on Breakthrough Innovation, French Ministry of Economy / German Ministry of Research, Technology and Space, June 18, 2026. Task force members include Philippe Aghion (2025 Nobel Prize in Economics), Thomas Kalil (Renaissance Philanthropy), Fidji Simo (OpenAI Applications), and André Loesekrug-Pietri (JEDI). ↩
  14. Patrick Collison, Tyler Cowen & Patrick Hsu, “What We Learned Doing Fast Grants” (2021), future.com. Fast Grants used a 30-minute application and 48-hour decisions; the regrantor model generalises the same compression of the funding-decision cycle. ↩
  15. On Fast Grants, see Collison, Cowen & Hsu, “What We Learned Doing Fast Grants” (2021): a centrally run programme with ~48-hour decisions and more than 260 grants totalling over US$50 million in 2020. On the regrantor model, see the delegated-budget programmes run by Manifund and the FTX Future Fund (2022). Retrospective evaluation of these programmes remains thin. ↩
  16. Michael Nielsen & Kanjun Qiu, “The trouble in comparing different approaches to science funding”, Science++ (2022). The argument from heavy-tailed, outlier-dominated research value: randomised trials characterise the average grant well but are nearly uninformative about the rare right tail that carries most of the value. ↩
  17. Michael Kratsios (OSTP), “Science: A New Golden Age,” White House Office of Science and Technology Policy, July 22, 2026. A 123-page report recommending support for individual scientists, diversified funding, agile research organisations, and agency metascience units. ↩
  18. NSF, “NSF Announces $1.5B for Foundational Research to Drive Scientific Breakthroughs,” August 17, 2026. ↩

Further reading ↑

Cite as: Rex, David Hilmer (2026). Funding Mechanisms: How the Design of Research Money Shapes What Science Happens. https://davidhilmerrex.nu/metascience/funding-mechanisms.html