Same Problem, Different Field: Cross-Domain Solution Import via Domain-Stripped Computational Fingerprints
cs.DL · cs.CL · cs.IR
Quick verdict
Proposes LLM-distilled domain-stripped fingerprints to find cross-field twins; strong gains on a curated benchmark, wild-corpus evidence rests on LLM judges.
Summary
The paper targets cross-domain 'solution import': finding pairs of papers in unrelated fields that solve the same computational problem, so a bespoke implementation can be replaced by another field's standard, specialized solver. Each paper is distilled once, via a single cached LLM call (Claude Haiku), into a computational fingerprint: a free-text mechanism skeleton with domain vocabulary and canonical method names stripped, plus seven controlled facets. A TF-IDF/BM25 search over fingerprint text retrieves candidates; an exact facet-agreement filter tightens precision. On a curated benchmark of 109 papers across 18 method families, skeleton+TF-IDF raises average precision over abstracts from 0.222 to 0.513 (whole fingerprint 0.557), while SPECTER, SPECTER2, SciNCL and SemCSE all fall below plain abstract+TF-IDF. An interventional re-skin/math-edit test (n=10) indicates the fingerprint tracks computation rather than field. On a 501-paper wild corpus, known twins fill 23 of the top 30; with planted pairs excluded, three blind LLM judges rate 8 of the top 30 pairs genuine. Four executed imports include an open nonparametric-MLE solver reproducing NPAG's dosing estimate on NPAG's own data. Main limitations: the benchmark draws on cores already known to be swappable, wild-corpus verification relies on LLM rather than human judges, recall is capped (0.619 within the top 1000 pairs), and the cheap-distiller-beats-frontier result may reflect coarse hand-defined family labels.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A domain-stripped LLM-distilled fingerprint, retrieved by plain TF-IDF/BM25, lifts cross-domain twin average precision from 0.222 (abstract) to 0.513 (skeleton) and 0.557 (whole fingerprint) on a 109-paper, 18-family benchmark.
- SPECTER, SPECTER2, SciNCL and SemCSE all fall below abstract+TF-IDF on this task because citation-trained signals encode topical similarity, the wrong signal for cross-domain retrieval.
- An interventional re-skin/math-edit test (n=10) shows the fingerprint tracks the computation rather than the field.
- On a 501-paper wild corpus, known twins fill 23 of the top 30, and with planted pairs excluded three blind LLM judges rate 8 of the top 30 unplanted pairs genuine (0 of 30 random).
- Four executed imports — including an open solver reproducing NPAG's dosing estimate on NPAG's own data — demonstrate end-to-end solution import.
Strengths
Controlled representation swap isolates the causal claim
Section 5.2 fixes the embedder and swaps only the input: abstract to skeleton moves AP 0.222 to 0.513 under TF-IDF, P@1 0.55 to 0.82, MRR 0.67 to 0.86, with paired-bootstrap CIs excluding zero — the representation, not the model, carries the gain.
Statistical care above venue norm
Paired bootstrap (B=2000) with Holm-Bonferroni across baseline margins (all p<0.001), plus a family-level cluster bootstrap acknowledging the 210 twins derive from only 18 cores (fingerprint over abstract +0.335 [+0.230, +0.527]); honest nulls are reported (fingerprint vs skeleton +0.044, p=0.38).
Interventional, not just correlational, validation
Honest boundary drawing throughout
End-to-end execution, checked against a non-strawman reference
Weaknesses
Wild-corpus claims rest on unvalidated LLM judges
Single-author labels, no inter-annotator agreement
In-sample tuning with no held-out split
Small benchmark, wide headline CI
Structural circularity, acknowledged but real
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (JCDL '26, single author, KU Leuven) addresses cross-domain 'solution import': surfacing pairs of papers in unrelated fields that solve the same underlying computational problem, so a bespoke implementation can be swapped for another field's standard solver (Section 1). The method distills each paper once, via one cached LLM call (Claude Haiku), into a faceted computational fingerprint — a free-text mechanism skeleton with domain vocabulary and canonical method names stripped (Figure 2's pharmacokinetics-to-nonparametric-mixture example), plus seven controlled facets grounded in published taxonomies such as the Berkeley 13 dwarfs for STRUCTURE (Section 3.2). Retrieval is deliberately plain: BM25/TF-IDF over fingerprint text, then an exact-match facet conjunction as a precision filter (Section 3.3). Central claims: (1) the representation, not the embedder or a learned similarity, carries the gain — abstract-to-skeleton under TF-IDF lifts AP 0.222 to 0.513, whole fingerprint 0.557, while SPECTER, SPECTER2, SciNCL and SemCSE's base encoder all fall below plain abstract+TF-IDF (Table 2); (2) an interventional re-skin/math-edit test shows the fingerprint tracks computation, not field (Table 5); (3) in the wild, known twins dominate the top ranks and blind judges rate unplanted top-30 pairs genuine; (4) four executed imports, headlined by an open nonparametric-MLE solver reproducing NPAG's dosing estimate on NPAG's own rifapentine example (Table 8).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does 'solution import' mean in this paper?
Replacing a field's bespoke implementation with another field's standard, specialized solver for the same underlying computation. The paper demonstrates it by reproducing a clinical pharmacokinetics dosing engine (NPAG) with a generic open-source nonparametric maximum-likelihood solver on NPAG's own example data, matching clearance within 2.5%.
Why do scientific embedders like SPECTER fail at finding cross-domain twins?
SPECTER, SPECTER2 and SciNCL are trained on citation neighborhoods, which cluster papers by field. Cross-field computational twins share no topic, vocabulary, or citations, so on the paper's 109-paper benchmark their abstract-level average precision (0.095 to 0.182) falls below even plain abstract+TF-IDF (0.222).
Is the code and data available?
Yes. The benchmark (versioned link-lists with checksums), distillation prompts, shipped skeletons, and evaluation code are released at github.com/ErykKul/same-problem-different-field, with a DOI-archived dataset (10.48804/W3B9WC). Retrieval runs on CPU; the per-paper fingerprints are cached, so re-running evaluations needs no LLM API spend.
Do I need a large model to use this method?
No. The shipped fingerprints use Claude Haiku, one cached call per paper (about a dozen papers per call), and retrieval is plain TF-IDF/BM25 plus a keyword-facet filter on an Elasticsearch stack. A local qwen3:14b also runs on one consumer GPU, though it underperforms on the faceted operator (combined AP 0.087 vs 0.330 on the wild corpus).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure bodies (the Figure 1 pipeline diagram and Figure 2 example) appear only as captions in the linear text extraction; their visual content could not be independently verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.