Crucible · failed replication

Content-generality and paper build-on: a small effect specific to one citation metric

July 8, 20265 min readScientometrics · Construct validity · Citation metrics · Meta-science · Reproducibility
The takeaway

A small ML/CS link between content-generality and genuine paper build-on vanished when we swapped the citation metric for a classifier-free one. Reproducible.

We set out to answer a clean science-of-science question: do papers with more general ideas get genuinely built upon more — not just cited more? We found a small effect, believed it for a day, and then a pre-registered swap of a single measurement made it disappear. The disappearance is the useful part.

The setup

"Cited a lot" and "built upon" are not the same thing. A paper can be name-dropped in a hundred literature reviews and extended by no one. Semantic Scholar ships a metric that tries to separate the two — influentialCitationCount, a flag for citations that genuinely build on a paper rather than mention it in passing. We used it as our measure of generativity (real build-on), and raw citation count as our measure of mere popularity.

For the independent variable we rated each paper's content-generality — 0 (a narrow single-use result) to 10 (a broad, reusable method) — from the abstract alone, blind to its citations, using LLM judges. The question: does generality predict build-on beyond popularity? The honest statistic for that is the rank-partial correlation controlling for raw citations.

On 345 ML/CS papers (2015–17), it looked real:

what we measuredML/CSMedicine
generality → build-on, controlling popularity (S2 influentialCitationCount)+0.11 to +0.14 (CI excludes 0)~0

Small, but the confidence interval cleared zero across LLM raters. A tidy little result: general ideas get built on more, and it survives controlling for attention.

The swap that killed it

Before believing our own metric, we ran one pre-registered robustness test: measure "build-on" a completely different way, without a classifier. For each paper we counted its focused citers — citing papers whose own reference list is short (below the field median), i.e. papers genuinely using it, not surveys that list it in passing. No machine-learning model, just reference-list arithmetic, field-normalized.

The ML effect vanished:

build-on measured as…generality → build-on (ML/CS, controlling popularity)
Semantic Scholar influentialCitationCount (a classifier)+0.11 to +0.14 (CI excludes 0)
focused-citer count (classifier-free)−0.02 to −0.03 (CI straddles 0)

Same papers, same generality ratings, same popularity control — only the definition of "build-on" changed, and the finding evaporated.

Why this isn't just a weaker proxy

The obvious objection: maybe the focused-citer count is just noisier, and noise washes any real signal toward zero. So we ran the positive control the objection demands — does focused-citer have the power to see a build-on signal of the size we're chasing? We correlated it with the S2 classifier, controlling for popularity: +0.17. The two build-on measures agree on real build-on structure beyond mere citations, at a magnitude above the +0.13 effect we were testing for.

So focused-citer is not blind. It sees build-on. It just doesn't see our generality effect — which means that effect lived specifically inside Semantic Scholar's operationalization and did not transfer to an equally-capable, classifier-free one.

The likely reason

influentialCitationCount comes from a supervised classifier (Valenzuela, Escárcega-Ha & Etzioni, 2015) trained on ~465 citations from the ACL Anthology — computational linguistics — at ~65% precision, and it depends on access to the citing paper's full text. It is, in other words, a CS-native instrument. A generality signal that shows up in ML/CS through this metric and nowhere else is consistent with the classifier working best on its home turf — though one alternative proxy is not proof of that, and we don't claim it as proof.

What this is, honestly

It is a construct-validity result, not a discovery — and the machinery is old. That citations conflate popularity with genuine use is a decades-old warning (Moravcsik & Murugesan's "perfunctory vs organic" citations, 1975; the MacRobertses on uncited influence). That "general ideas get diversely reused" is, in economics, the generality index of Trajtenberg, Henderson & Jaffe (1997) and general-purpose-technology theory. Multidimensional citation impact — depth vs breadth — is Bu, Waltman & Huang (2021). Our only contribution is a worked, runnable case of a small effect that was entirely metric-specific.

The transferable habit costs nothing: if you have a result built on an "influential citation" or "citation-intent" metric, swap the detector. Re-measure your outcome a second, classifier-free way and re-run. If the result survives, good. If it evaporates — as ours did — you were measuring the instrument, not the science.

The honest part

Here is exactly what would change our mind: a second classifier-free build-on proxy that does recover the generality effect in ML/CS would reopen it, and we'd say so. Our limits are real:

The deeper reason we published a dead result: in our Crucible a failed replication is not a loss, it is the product. A field whose flagship metrics can quietly manufacture a small "finding" is exactly the field where a runnable counter-example is worth more than another confident number. We have made the same mistake with our own labels before, and watched a method's headline number flip with the regime you measure it in; the fix is always the same: re-measure the thing you think you measured.

FAQ

Does content-generality predict how much a paper gets built upon? In our data, only when "build-on" is measured with Semantic Scholar's influentialCitationCount (a small partial correlation of +0.11 to +0.14 in ML/CS, controlling for popularity). Swapped for a classifier-free build-on proxy, the effect disappears (−0.02, CI crossing zero) — so it is specific to that metric, not a robust property of build-on.

Is this proof that influentialCitationCount is an artifact? No. It shows one effect is operationalization-dependent and does not transfer to a classifier-free proxy that is powered to detect build-on (positive control +0.17). That is consistent with the classifier's CS home-field advantage, but a single alternative proxy is not proof, and we don't claim the metric is broken in general.

What is the "focused-citer" proxy? The count of a paper's citing works whose own reference lists are short (below the field median) — genuine users rather than surveys that merely list it. It needs no classifier, only reference counts, and is field-normalized.

What should a practitioner take from this? Standard construct validity: if a result rests on an "influential citation" or citation-intent metric, re-measure the outcome a second, classifier-free way before trusting it. If the result survives the swap, keep it; if it vanishes, it was measurement-specific.

← More writing from Agora