yerim0210/Scientific_Slop

SciSlopBench

paper link

SciSlopBench is a paired benchmark for scientific slop: failures of global scientific reasoning in AI-generated research papers that token-level AI detectors do not see. It holds 390 pairs. Each pair is one AI-generated paper and one human-written, peer-accepted paper (its anchor) of the same contribution type on the same or a neighboring research problem. The AI papers come from two sources, and the same pairing rule is applied to both:

Half	AI papers	Anchor pool	Topic similarity tier 3 / 2 / 1 / 0
FARS	143 papers generated end to end by the FARS system	120 cited by the AI paper, 23 from accepted ICLR 2026	93 / 23 / 27 / 0
Agents4Science 2025	all 247 submissions to the Agents4Science 2025 conference	66 cited, 119 from accepted ICLR 2026, 62 nearest admissible anchor	63 / 133 / 45 / 6

Contribution types over the 390 pairs: method 286, analysis 71, framework 21, benchmark 6, survey 6. Topics: the FARS half covers eight machine-learning areas; the Agents4Science half is mostly computer and data sciences (178) with life and health sciences (17), social sciences (14), natural sciences (13), interdisciplinary (11), engineering (6), mathematics and statistics (4), humanities (2), and law, policy and business (1).

The benchmark accompanies the paper Science or Slop? Benchmarking and Mitigating Scientific Slop in AI-Generated Papers (under review, ICLR 2027). The paper's six SciSlop measures across Structure, Argument and Artifacts identify the AI paper in 85.9% of pairs, against 68.7% for Binoculars. All numbers of the paper's benchmark tables are recomputable from scores alone (see below).

What is in the release

Three tables, all Parquet, loadable with datasets:

from datasets import load_dataset
pairs  = load_dataset("yerim0210/Scientific_Slop", "pairs")["train"]    # 390 rows, one per pair
papers = load_dataset("yerim0210/Scientific_Slop", "papers")["train"]   # 773 rows, one per unique paper
scores = load_dataset("yerim0210/Scientific_Slop", "scores")["train"]   # 780 rows, one per pair side

pairs (390 rows): the pairing record

Every match can be checked independently: the candidate pool, similarity tier, the four anchor conditions and the retrieval rank are stored for each pair.

Column	Meaning
pair_id	FA0001 … (FARS) or A4S0011 … (Agents4Science)
source	fars or agents4science
ai_paper_id, human_paper_id	keys into papers (AI_<pair_id> and HU_<arxiv id>)
ai_title, human_title, human_arxiv, human_venue, human_v1_year	identity of both sides; the anchor is always an arXiv e-print
human_iclr_rating	mean public ICLR review rating of the anchor when it exists (221 anchors)
contribution_type, human_contribution_type, same_contribution_type	one of method, framework, benchmark, dataset, analysis, survey, judged from title and abstract
pool	cited (anchor cited by the AI paper), widened_iclr26 (accepted ICLR 2026 paper found by TF-IDF), filled_iclr26 (Agents4Science only: closest admissible anchor when no same-type anchor exists)
sim_tier	3 same problem and contribution type, 2 same problem, 1 same area, 0 cited only as a tool, model or dataset
relation, tier_reason	how the anchor relates to the AI paper; the judge's one-line reason (Agents4Science half)
H1_human, H1_by, H2_top_tier, H3_type_match, H4_source	outcome of the four anchor conditions (see Pairing rule); H1_by is year<=2024 or pangram
specter2_cosine, tfidf_similarity, match_rank	retrieval diagnostics (SPECTER2 cosine for FARS, TF-IDF for Agents4Science; rank of the assigned anchor among candidates)
human_pangram_ai_fraction	Pangram AI share of the anchor when H1 was checked with Pangram (≤ 0.05 by construction)
ai_body_words, human_body_words, body_length_ratio	body lengths of the two sides
topic, secondary_topic	FARS: one of eight ML areas; Agents4Science: the submission's primary and secondary topic
a4s_submission_id, a4s_openreview_url, a4s_decision, a4s_reviewer_scores, a4s_ai_involvement	Agents4Science metadata: OpenReview id and URL, accept/reject decision (48 / 199), the three AI reviewer scores and the human score as JSON, and the self-reported AI involvement per stage (A–D) as JSON. These are metadata only and were not used for pairing
ai_tex_source	shipped_tex (LaTeX provided by the system or in the supplementary material, 161 papers) or rebuilt_from_pdf (LaTeX reconstructed from the OpenReview PDF, 229 papers)

Evaluation subsets used in the paper, expressed on pairs:

• all: 390 pairs;
• problem-matched: sim_tier >= 2 (116 FARS + 196 Agents4Science = 312 pairs);
• strict: source == "fars" or pool != "filled_iclr26" (143 + 185 = 328 pairs, same contribution type and a cited or widened-pool anchor).

papers (773 rows): the two document views

One row per unique paper. Seven human anchors serve one FARS pair and one Agents4Science pair each and appear once, with both pairs listed in pair_ids.

Column	Meaning
paper_id	AI_<pair_id> or HU_<arxiv id>
pair_ids	list of the pairs this paper belongs to
source	fars, agents4science, or fars+agents4science for the seven shared anchors
label, role	1 / ai for the AI-generated paper, 0 / human for the anchor
title, arxiv, venue	identity; arxiv and venue are set for anchors
body_words	words in the body view
body_tex	structure view: the LaTeX source expanded in document order (\input resolved) and cut at the first of appendix, acknowledgements or bibliography. Sections, floats, equations, labels and \ref/\cite commands are intact, which is what the Structure and Artifacts measures read
body_txt	prose view: the same body as running text, with inline and display mathematics replaced by MATH, citations by CITE and cross-references by REF, which is what the Argument measures and the token-level detectors read

The body view is the paper without its appendix so that AI and human papers are compared on the part every reader sees; the human anchors pass H4 only if the appendix is separable and the body has at least 2,500 words in four or more sections.

scores (780 rows): measures and baselines per pair side

One row per side of each pair (pair_id, paper_id, label). The SciSlop measures are deterministic per paper, so the seven shared anchors carry the same measure scores in both of their rows; the detector and LLM-reviewer baselines were run once per half and are stored as run.

Column	Direction	Meaning
scislop_macro_redund, scislop_xsec_ref	higher = more slop	Structure: Macro redundancy, later sections repeating earlier material instead of developing the argument (cross-section n-gram overlap); Cross-section references, sections and labeled objects that no other section refers to
scislop_argument_graph, scislop_citation	higher = more slop	Argument: Argument graph, claims that appear before their supporting context; Citation isolation, prior work cited without its role in the argument (builds on, differs from, compares against)
scislop_fig_exposition, scislop_evidence_gap	higher = more slop	Artifacts: Figure exposition, method diagrams cluttered with experimental details or result claims; Evidence gap, aggregate result tables with no concrete input, output or case anywhere in the paper. null where the measure does not apply
scislop_aggregate	higher = more slop	mean over the three planes of the mean of the plane's available measures
binoculars	lower = more AI-like	Binoculars score
detectgpt, fast_detectgpt, nts	higher = more AI-like	DetectGPT, Fast-DetectGPT and the NTS curvature score
pangram	higher = more AI-like	Pangram AI probability (Pangram 3, v3.3.2). Pangram is also the H1 filter for anchors written after 2024, so it is not an independent detector on this benchmark
cyclereviewer_overall, ai_scientist_overall	lower = more AI-like	overall rating given by the CycleReviewer 8B and AI Scientist reviewer baselines

Reproducing the paper's pair accuracy (the AI paper scores higher than its anchor; ties count 0.5):

import pandas as pd
s = scores.to_pandas()
def pair_accuracy(col, direction=+1):
    w = s.pivot_table(index="pair_id", columns="label", values=col).dropna()
    d = (w[1] - w[0]) * direction
    return ((d > 0).sum() + 0.5 * (d == 0).sum()) / len(d), len(d)
pair_accuracy("scislop_aggregate")        # (0.859, 390)
pair_accuracy("binoculars", direction=-1) # (0.687, 390)

Table 2 of the paper, recomputed from scores (pair accuracy / AUROC over the 390 pairs):

Method	PairAcc	AUROC		Plane	Measure	PairAcc	AUROC
Binoculars (prose view)	0.687	0.683		Structure	Macro redundancy	0.723	0.727
DetectGPT (prose view)	0.638	0.623			Cross-section references	0.905	0.895
NTS (prose view)	0.626	0.607		Argument	Argument graph	0.586	0.584
CycleReviewer (full text)	0.685	0.689			Citation isolation	0.793	0.782
AI Scientist reviewer (full text)	0.615	0.613		Artifacts	Figure exposition	0.809	0.786
SciSlop, mean of planes	0.859	0.854			Evidence gap	0.764	0.765

Pairing rule

A human paper can anchor an AI paper only if it passes four conditions.

• H1 Human authorship. Its first arXiv version is dated 2024 or earlier, or it is an accepted ICLR 2026 paper whose Pangram AI share is at most 0.05.
• H2 Acceptance. It was accepted to the main track of a top-tier venue, shown by the venue record, the arXiv comment or journal reference, or OpenReview.
• H3 Contribution type. It is the same kind of paper as the AI paper (method, framework, benchmark, dataset, analysis or survey).
• H4 Usable source. Its arXiv e-print expands in document order to a body of at least 2,500 words in at least four sections, with the appendix separable from the body.

Within each contribution type, the anchor is the candidate closest in research problem. A language model assigns a topic-similarity tier from title and abstract (3, 2, 1, 0 as above); ties are broken by experiment-section citation, availability of public review scores, a human-to-AI body-length ratio within 2.5, and finally SPECTER2 cosine. Two candidate pools are searched in order: papers cited by the AI paper, then 2,690 accepted ICLR 2026 papers with an arXiv identifier and Pangram share at most 0.05 (three nearest by TF-IDF over title and abstract). Assignment is one-to-one within each half; contested anchors go to the closer match. For the 62 Agents4Science submissions whose topics have no same-type anchor in either pool (physics, astronomy, biology, medicine, finance, HCI, …), H1, H2 and H4 are kept and the closest admissible anchor is taken in a recorded order (pool = filled_iclr26, H3_type_match = relaxed).

Document processing

• FARS papers ship their LaTeX; the body view is built from the shipped source.
• Agents4Science papers were collected from OpenReview. 18 provide LaTeX in the supplementary material; for the other 229 the LaTeX was reconstructed from the PDF and accepted only if at least 85% of forty sampled PDF sentences are recovered verbatim in the rebuilt source.
• Human anchors are arXiv e-prints, expanded in document order from the e-print's root file.

Provenance and licensing

• The pairing record, the scores, the FARS-generated papers and the document views produced for this benchmark are released under CC BY 4.0.
• The Agents4Science 2025 submissions are public on OpenReview (a4s_openreview_url); copyright stays with their authors and systems. Their decisions, reviewer scores and AI-involvement declarations are reproduced as published.
• The human anchors are published arXiv e-prints. Copyright stays with their authors and the text is redistributed here, for research on AI-generated scientific writing, under the terms of each e-print's arXiv license. If you are an author and want your paper's text removed, open a discussion on this repository; the pairing record and scores will keep the arXiv identifier so the benchmark stays reproducible.

Citation

@article{oh2026scislop,
  title         = {Science or Slop?: Benchmarking and Mitigating Scientific Slop in AI-Generated Papers},
  author        = {Oh, Yerim and Lee, Young-Jun and Ahn, Jaewoo and Kim, Gunhee and Kang, Dongyeop},
  journal       = {arXiv preprint arXiv:2610.00531},
  year          = {2026},
  eprint        = {2610.00531},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2610.00531}
}