NSF GRFP Reference Letters: What Reviewers Score

NSF GRFP Reference Letters: Intellectual Merit and Broader Impacts, Explained for Letter Writers

If a student has just asked you to write an NSF GRFP letter of recommendation — NSF calls it a reference letter — you have agreed to write the most rule-bound letter in American graduate education. It has a hard two-page cap enforced by software, a mandatory font, its own deadline, and, unlike almost any other letter you will write this year, it is scored against two named criteria that NSF defines in print.

The stakes are real. For the 2026-27 academic year, NSF announced it "will award 2,500 Graduate Research Fellowships" from "a highly competitive pool of nearly 14,000 applicants" — three years of funding, and a program whose alumni include more than 40 Nobel laureates. At three counted letters per applicant, that pool implies on the order of 40,000 reference letters in a single cycle, read by virtual panels working quickly. A letter that ignores the scoring criteria does not get partial credit for eloquence.

This guide translates NSF's own instructions — the reference writer requirements, tips, and the current program solicitation — into a working plan for letter writers, and for the applicants preparing their writers' briefing packets.

Quick answer

• Two criteria, both required. NSF: "Please address both the Intellectual Merit and Broader Impacts of the applicant in your letter."
• Two pages, hard cap. "The GRFP Module cannot accept letters that are longer than two pages."
• Format: PDF, 8.5" x 11", 1-inch margins, 11-point Times New Roman, on letterhead if possible, signed, submitted only through the GRFP Module — no email, no mail.
• Counting rules: applicants list 3-5 writers ranked by preference; the top 3 by rank are reviewed; if fewer than 2 letters arrive, the application is not reviewed at all.
• Deadline: reference letters have their own deadline, separate from the application. Check the current solicitation and do not assume slack — NSF's FAQ says flatly there are no exceptions.

The two criteria, in plain language

Every GRFP application — essays, transcripts, and your letter — is evaluated against NSF's two Merit Review Criteria. The solicitation defines them:

"The Intellectual Merit criterion encompasses the potential to advance knowledge."

"The Broader Impacts criterion encompasses the potential to benefit society and contribute to the achievement of specific, desired societal outcomes."

For a letter writer, the translation is:

Intellectual Merit (IM) asks: can this person do excellent science? Your evidence is the research you watched them do — the question they formulated themselves, the experiment they redesigned after it failed, the analysis they got working when the standard approach did not, the pace at which they went from supervised to independent. Publications and awards help, but panels read them in the application already; what only you can supply is the eyewitness account of how the applicant actually operates as a scientist.

Broader Impacts (BI) asks: will this person's work and career benefit others? From the writer's chair, that means evidence of the applicant as a multiplier: mentoring newer students, teaching and outreach, making a team function, communicating science beyond the lab, broadening who gets to participate in it. NSF's requirements also ask you to comment specifically on "the applicant's potential for contributing to a globally-engaged United States science and engineering workforce" — a workforce framing, not a volunteering-hours checklist.

The single most common failure mode is writing a superb IM letter with a throwaway BI sentence at the end. It is common enough that we gave it its own post — the all-merit, no-Broader-Impacts letter. If you read only one companion piece, read that one.

The counting rules (applicants, this is your job)

The solicitation sets mechanics that surprise both sides:

Two implications. First, the applicant's rank order matters: a fourth letter from a famous name does nothing if it is ranked below three others, and a top-ranked weak letter displaces a stronger one. Second, listing four or five writers is cheap insurance against a ghosting recommender — with a two-letter floor and a no-exceptions deadline, a bench is not optional.

Format rules: the module enforces them

From NSF's FAQ and requirements pages:

• Maximum 2 pages — the submission module rejects longer files
• PDF, standard 8.5" x 11" page, 1-inch margins on all sides
• 11-point Times New Roman in the body
• Institutional or professional letterhead, if possible
• Signed, with your name, title, department, and institution
• Submitted online through the GRFP Module only — "Email and physical mail submissions are not accepted"

Two pages is not a suggestion to fill. As with every admissions letter genre, a dense page and a half of evidence beats two pages padded with adjectives — but unlike most genres, here the ceiling is mechanical.

What NSF says the letter must contain

The requirements page enumerates the content:

1. The relationship header — your department and institution, "how long you have known the applicant, and in what capacity." Panels calibrate everything else against this sentence; it is the same credibility header that anchors any strong letter for a student.
2. If you are the research advisor — comments on the originality of the applicant's research statement and your role, if any, in assisting with the application. This is an honesty mechanism: panels know advisors help, and they want the help disclosed rather than discovered.
3. Academic potential and prior research experiences — the IM core.
4. The applicant's proposed research — your professional judgment of the plan they are actually submitting, which means you should read their statements before you write. For doctoral-bound applicants this overlaps heavily with what PhD admissions committees want from letters: research-potential evidence over course grades.
5. Workforce potential — the "globally-engaged United States science and engineering workforce" comment, plus anything else that helps panels apply the two criteria.

What panels reward (from NSF's own tips)

NSF's tips page is unusually candid about what separates useful letters from wallpaper:

• Specifics beat generics. "Letters that offer reviewers specific information are much more effective than generic ones." Describe specific projects, the applicant's contributions, work ethic, and teamwork — as evidence for the two criteria, not as free-floating praise.
• Comparative placement is welcome. NSF says it is acceptable to "compare the relative placement of the applicant among peers" — the "strongest three of the sixty I have supervised" sentence that panels quote — but do not name other students or fellowship recipients.
• Say what the applicant is too modest to say. NSF explicitly invites "accomplishments or activities that the applicant may be too modest to mention." This is frequently where the best Broader Impacts evidence lives.
• Complement, don't repeat. Coordinate with the applicant so the letter adds to their statements instead of paraphrasing them. A letter that restates the application spends its two pages saying nothing the panel hasn't read.

Confidential means confidential — and what that implies

GRFP letters go from you to NSF through the module and are never shared with the applicant. This is structural, not courtesy: the panel is buying your unfiltered judgment. It also means the common grad-school shortcut — "just draft it and I'll sign" — has no legitimate version here; if you were asked anyway, or you are an applicant wondering how much help is too much, we wrote a separate guide to self-drafting requests in fellowship references. The clean division of labor: the applicant supplies a thorough briefing packet — CV, statements, dates, project bullet points; the writer supplies the testimony.

Before you upload it

A GRFP letter is a two-page argument scored against two published criteria, and you get no second submission. GradPilot reviews the letter you are drafting the way a panel will read it: relationship evidence, incident specificity, comparative calibration, whether both criteria actually have evidence attached — plus AI detection in the full review, since a letter that reads machine-drafted undercuts its own testimony. Run the draft through GradPilot before it goes into the module. We review letters; we never write them.

---

Sources

• NSF GRFP — Reference Writer Requirements, Reference Writer Tips, Reference Writer FAQs
• NSF — GRFP Program Solicitation NSF 25-547 (Merit Review Criteria definitions; reference letter counting rules)
• NSF — NSF announces 2026 Graduate Research Fellowship Program offers (2,500 offers; ~14,000 applicants)