WP44: verified bibliography (581 entries, Agent B-lit), semantic edges script, register
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
8 changed files +8,458 −2
modified
agents/REGISTER.md
+3 −1
@@ -5,4 +5,6 @@ | ||
| 5 | 5 | | 2026-11-02 | A (director, this session) | CLAUDE.md | RESEARCH_PLAN.md, protocol/PROTOCOL.md, reports/PHASE0_FEASIBILITY.md | PI checkpoint H-0 pending | delivered | |
| 6 | 6 | | 2026-11-02/03 | B+C (acquisition & data engineering, this session) | RESEARCH_PLAN §2–3 | ingestion/*.py, ledgers on M3U96b, data/metadata/linkage_report.json | N (audit, Phase 8) pending | delivered | |
| 7 | 7 | | 2026-11-03 | D (taxonomy architect, subagent, 2nd attempt) | agents/briefs/AGENT_D_TAXONOMY.md | taxonomy/ontology.yaml (v1.0, 40 leaves, 8 areas, 646 topic weights, 771 DBLP streams, 273 regexes, 62 unmapped CS topics listed), topic_map.csv, ONTOLOGY.md, src/ | A (build re-run, classify smoke test 2026-11-03) + validation sample (Phase 2) | delivered 2026-11-03 | |
| 8 | −| 2026-11-03 | B-lit (reference acquisition, subagent) | agents/briefs/AGENT_LIT_REFERENCES.md | manuscript/references.bib, references_verification.csv, literature_synthesis/REFERENCE_MAP.md | M (compile) + N (re-verify sample) | running | | |
| 8 | +| 2026-11-03 | B-lit (reference acquisition, subagent) | agents/briefs/AGENT_LIT_REFERENCES.md | manuscript/references.bib (581 verified entries: Crossref 404, arXiv 165, OpenAlex 1, URL 11; 7 FAILED kept in CSV), references_verification.csv, literature_synthesis/REFERENCE_MAP.md | M (compile) + N (re-verify sample, Phase 8) | delivered 2026-11-03 | | |
| 9 | +| 2026-11-03 | L×3 (writers Parts II, III, IV, subagents) | agents/briefs/WRITER_PART0{2,3,4}.md + WRITER_COMMON.md | manuscript/sections/part0{2,3,4}_*.tex, references_part0N.bib, references_verification_part0N.csv | K (critic) + M (compile) | running | | |
| 10 | +| 2026-11-03 | N (code auditor, subagent) | prompt in session | reports/audit/CODE_AUDIT_2026-11-03.md, code_issues.csv | A (fixes applied + logged) | running | | |
added
knowledge_graph/04_semantic_edges.py
+46 −0
@@ -0,0 +1,46 @@ | ||
| 1 | +#!/usr/bin/env python3 | |
| 2 | +"""WP44 — semantic k-NN edges between works (Phase 3, semantic layer) with FAISS HNSW on the SPECTER2 embeddings. | |
| 3 | +Edges: (work_id, neighbour_id, cosine_similarity, rank) for k = --k neighbours, similarity >= --min-sim; provenance = model tag + | |
| 4 | +index parameters; verification_status = 'machine-verified' (deterministic from embeddings). Also computes, per work, the share | |
| 5 | +of its k neighbours that belong to a different primary category (semantic boundary-crossing, used in RQ5) and writes | |
| 6 | +results/graph/semantic_crossing_yearly.csv. Output: data/processed/graph/semantic_edges_<tag>.parquet.""" | |
| 7 | +import argparse, glob, json, os, time, numpy as np, pandas as pd | |
| 8 | + | |
| 9 | + | |
| 10 | +def main(): | |
| 11 | + ap = argparse.ArgumentParser(); ap.add_argument("--root", required=True); ap.add_argument("--model", default="allenai/specter2_base"); ap.add_argument("--k", type=int, default=10) | |
| 12 | + ap.add_argument("--min-sim", type=float, default=0.80); ap.add_argument("--sample", type=int, default=0, help="query only a random sample of works (0 = all)") | |
| 13 | + a = ap.parse_args(); R = a.root; tag = a.model.replace("/", "__"); out = f"{R}/data/processed/graph"; os.makedirs(out, exist_ok=True); os.makedirs(f"{R}/results/graph", exist_ok=True) | |
| 14 | + import faiss, duckdb | |
| 15 | + E, I = [], [] | |
| 16 | + for f in sorted(glob.glob(f"{R}/data/processed/embeddings/{tag}/shard_*.npy")): | |
| 17 | + if f.endswith(".ids.npy"): continue | |
| 18 | + E.append(np.load(f)); I.append(np.load(f.replace(".npy", ".ids.npy"))) | |
| 19 | + X = np.concatenate(E).astype(np.float32); ids = np.concatenate(I); faiss.normalize_L2(X); n, d = X.shape; print("vectors", n, d, flush=True) | |
| 20 | + t0 = time.time(); index = faiss.IndexHNSWFlat(d, 32, faiss.METRIC_INNER_PRODUCT); index.hnsw.efConstruction = 80; index.add(X); print(f"index built in {time.time()-t0:.0f}s", flush=True) | |
| 21 | + index.hnsw.efSearch = 64 | |
| 22 | + q_idx = np.arange(n) if not a.sample else np.random.default_rng(44).choice(n, size=min(a.sample, n), replace=False) | |
| 23 | + rows = []; B = 50000 | |
| 24 | + for s in range(0, len(q_idx), B): | |
| 25 | + qi = q_idx[s:s + B]; D, N = index.search(X[qi], a.k + 1) | |
| 26 | + for r, (qq, dd, nn) in enumerate(zip(qi, D, N)): | |
| 27 | + rk = 0 | |
| 28 | + for sim, j in zip(dd, nn): | |
| 29 | + if j == qq or j < 0: continue | |
| 30 | + rk += 1 | |
| 31 | + if sim >= a.min_sim: | |
| 32 | + rows.append((int(ids[qq]), int(ids[j]), float(sim), rk)) | |
| 33 | + print(f"{s+len(qi)}/{len(q_idx)} queried, {len(rows)} edges, {time.time()-t0:.0f}s", flush=True) | |
| 34 | + df = pd.DataFrame(rows, columns=["work_id", "neighbour_id", "cosine", "rank"]); df.to_parquet(f"{out}/semantic_edges_{tag}.parquet") | |
| 35 | + json.dump({"model": a.model, "k": a.k, "min_sim": a.min_sim, "n_vectors": int(n), "n_edges": int(len(df)), "index": "HNSW32 efC80 efS64 inner-product", "seconds": round(time.time() - t0)}, open(f"{out}/semantic_edges_{tag}.json", "w"), indent=1) | |
| 36 | + con = duckdb.connect(f"{R}/data/processed/wp44.duckdb", read_only=True) | |
| 37 | + con.execute(f"CREATE OR REPLACE VIEW work_primary AS SELECT * FROM read_parquet('{R}/data/processed/classification/primary/*.parquet')") | |
| 38 | + con.execute(f"CREATE OR REPLACE VIEW sem AS SELECT * FROM read_parquet('{out}/semantic_edges_{tag}.parquet')") | |
| 39 | + con.execute(f"""COPY (SELECT c.publication_year AS year, pa.primary_category, count(*) n_edges, round(avg((pa.primary_category <> pb.primary_category)::int),4) share_cross_category | |
| 40 | + FROM sem s JOIN work_primary pa ON pa.work_id = s.work_id JOIN work_primary pb ON pb.work_id = s.neighbour_id JOIN corpus_works c ON c.id = s.work_id | |
| 41 | + GROUP BY 1,2 ORDER BY 1,2) TO '{R}/results/graph/semantic_crossing_yearly.csv' (HEADER)""") | |
| 42 | + print("done", len(df), "edges") | |
| 43 | + | |
| 44 | + | |
| 45 | +if __name__ == "__main__": | |
| 46 | + main() | |
added
literature_synthesis/REFERENCE_MAP.md
+251 −0
@@ -0,0 +1,251 @@ | ||
| 1 | +# REFERENCE_MAP — which verified references each manuscript part should cite (WP44, Agent B-lit, 2026-11-03) | |
| 2 | + | |
| 3 | +Source of keys: `manuscript/references.bib` (every entry resolved against Crossref / OpenAlex / DataCite / arXiv or a | |
| 4 | +live URL; see `manuscript/references_verification.csv`). Keys are `surnameYYYYkeyword`. Entries marked **core** are the | |
| 5 | +minimum set the part needs; the others are supplementary and may be cited selectively. Group letters (A–J) refer to the | |
| 6 | +scope groups of `agents/briefs/AGENT_LIT_REFERENCES.md`. Nothing in this map asserts a finding of WP44; each line only | |
| 7 | +says what the reference *supports*. arXiv-only entries are preprints and must be described as such in the text. | |
| 8 | + | |
| 9 | +Conventions: `key — why`. Parts follow CLAUDE.md §9 (I–XI + appendices). | |
| 10 | + | |
| 11 | +## Part I — Introduction and Scientific Motivation | |
| 12 | + | |
| 13 | +**core** | |
| 14 | +- openai2022chatgpt — the dated event (30 Nov 2022) that defines the research period; cite the archived announcement, not a secondary account. | |
| 15 | +- ouyang2022instructgpt — the instruction-tuned/RLHF recipe behind ChatGPT; establishes what was technically new. | |
| 16 | +- brown2020gpt3 — in-context learning at scale; the immediate pre-ChatGPT state of the art (baseline period). | |
| 17 | +- openai2023gpt4 — first major post-ChatGPT frontier release inside the window. | |
| 18 | +- bommasani2021foundation — "foundation model" vocabulary and the cross-field research agenda the manuscript tests empirically. | |
| 19 | +- fortunato2018scienceofscience ; wang2021scienceofscience — science-of-science framing: large-scale bibliographic data as the object of study. | |
| 20 | +- vannoorden2023survey ; messeri2024illusions ; birhane2023science — motivation: documented concern about AI changing how research is done. | |
| 21 | +- liang2024mapping ; kobak2025delving — evidence that LLM use in scientific writing rose sharply after 2022 (justifies studying the field's own output). | |
| 22 | +- krenn2023forecasting — AI literature growth modelled as an exponentially growing knowledge network; motivates forecasting/emergence questions. | |
| 23 | +- bornmann2021growth ; bornmann2015growth — long-run growth rates of science; needed to separate secular growth from a post-2022 break. | |
| 24 | +- park2023disruption ; chu2021slowed — the "science is slowing" debate the mapping must engage. | |
| 25 | +**supplementary** | |
| 26 | +- maslej2024aiindex ; maslej2025aiindex — descriptive statistics on AI publication/industry trends (secondary, cite as reports). | |
| 27 | +- ahmed2023industry — growing role of industry in AI research (motivates institution-type analyses). | |
| 28 | +- bianchini2022aiscience ; cockburn2018aiinnovation — AI as a general method of invention (economics framing). | |
| 29 | +- hicks2015leiden — responsible use of metrics; signals the manuscript's evaluative stance. | |
| 30 | +- dwivedi2023sowhat — multidisciplinary perspectives on ChatGPT for research and practice. | |
| 31 | + | |
| 32 | +## Part II — The Evolution of Computer Science (conceptual and historical background) | |
| 33 | + | |
| 34 | +**core (pre-2018 lineage of the deep-learning paradigm)** | |
| 35 | +- turing1950computing — origin of the machine-intelligence question. | |
| 36 | +- rumelhart1986backprop ; hochreiter1997lstm ; lecun2015deeplearning — backpropagation, recurrent memory, deep learning synthesis. | |
| 37 | +- krizhevsky2017alexnet ; he2016resnet — the 2012–2016 vision breakthroughs (CACM reprint of AlexNet is the verified DOI). | |
| 38 | +- goodfellow2014gan ; kingma2013vae — generative modelling before diffusion. | |
| 39 | +- mnih2015dqn ; silver2016alphago ; silver2017alphagozero ; silver2018alphazero ; schulman2017ppo ; christiano2017rlhf — deep RL lineage that leads to RLHF. | |
| 40 | +- sutskever2014seq2seq ; bahdanau2014attention ; vaswani2017attention — sequence modelling to the Transformer. | |
| 41 | +- devlin2019bert ; raffel2020t5 ; radford2019gpt2 ; brown2020gpt3 — pretrained language models 2018–2020 (baseline period). | |
| 42 | +- kaplan2020scaling ; hoffmann2022chinchilla ; sevilla2022compute ; thompson2020computationallimits — scaling laws and compute growth. | |
| 43 | +- shazeer2017moe ; fedus2022switch — sparse mixture-of-experts lineage. | |
| 44 | +- sutton2019bitterlesson — the "compute over priors" thesis often invoked to explain the paradigm shift (blog essay, cite as such). | |
| 45 | +**core (other CS areas' trajectories 2018–2022)** | |
| 46 | +- hennessy2019goldenage ; leiserson2020roomattop ; dennard1974scaling — end of Dennard scaling / domain-specific architecture narrative. | |
| 47 | +- jouppi2017tpu ; dally2021gpu ; sze2017efficientdnn — accelerator lineage. | |
| 48 | +- preskill2018nisq ; arute2019supremacy — quantum computing's NISQ era. | |
| 49 | +- shi2016edge ; satyanarayanan2017edge — edge computing. | |
| 50 | +- mckeown2008openflow ; bosshart2014p4 ; kreutz2015sdn — programmable networks. | |
| 51 | +- dean2008mapreduce ; zaharia2016spark ; corbett2013spanner ; verma2015borg ; burns2016kubernetes ; jonas2019serverless — cloud/distributed systems lineage. | |
| 52 | +- lamport1998paxos ; castro2002pbft ; gilbert2002cap — distributed consensus foundations; nakamoto2008bitcoin ; zheng2018blockchain — blockchain. | |
| 53 | +- gentry2009fhe ; dwork2006dp ; dwork2014dp ; shor1997factoring ; regev2009lwe ; bernstein2017pqc — cryptography/privacy lineage up to the PQC transition. | |
| 54 | +- leroy2009compcert ; klein2009sel4 ; jung2017rustbelt ; jung2021saferust ; matsakis2014rust — formal verification and memory-safe systems languages. | |
| 55 | +- cook1971complexity ; karp1972reducibility ; babai2016graphisomorphism ; raz2019bqpph ; ji2021mipre ; chen2022maxflow ; kelley2023threeap — landmark theory results framing "theory" as a continuing frontier. | |
| 56 | +- lattner2004llvm ; lattner2021mlir ; chen2018tvm ; paszke2019pytorch ; abadi2016tensorflow — compiler/framework infrastructure. | |
| 57 | +- mildenhall2020nerf ; karras2019stylegan ; ho2020ddpm ; rombach2022ldm ; radford2021clip ; dosovitskiy2021vit — vision/graphics turn to neural fields and diffusion. | |
| 58 | +- jumper2021alphafold ; baek2021rosettafold — AI-for-science landmark before the window. | |
| 59 | +- mcmahan2017fl ; kairouz2021fl — federated learning. | |
| 60 | +- amodei2016concrete ; hendrycks2021unsolved ; bender2021parrots — safety/ethics agenda before ChatGPT. | |
| 61 | +**supplementary** | |
| 62 | +- wu2015citeseerx ; giles1998citeseer ; ley2002dblp ; ginsparg2011arxiv — history of CS's own bibliographic infrastructure (arXiv, DBLP, CiteSeer). | |
| 63 | +- lariviere2014arxiv ; sutton2017arxivcs ; lin2020arxivcs — the rise of preprint culture in CS before 2022. | |
| 64 | +- frank2019citationgraphs ; klinger2020narrowing ; ahmed2020dedemocratization ; martinezplumed2021benchmarks — bibliometric portraits of AI research pre-2022. | |
| 65 | + | |
| 66 | +## Part III — Data and Systematic Review Protocol | |
| 67 | + | |
| 68 | +**core (sources — group A)** | |
| 69 | +- priem2022openalex ; openalex2024topics ; alperin2024openalex ; culbert2025openalexcoverage ; scheidsteger2023magopenalex — OpenAlex: description, topic classification, suitability and coverage evidence, MAG→OpenAlex continuity. | |
| 70 | +- ley2002dblp ; ley2009dblp — DBLP design, scope and curation lessons. | |
| 71 | +- ginsparg2011arxiv ; clement2019arxivdataset ; arxiv2026stats — arXiv as a source and its submission statistics. | |
| 72 | +- kinney2023s2ag ; lo2020s2orc ; ammar2018literaturegraph ; fricke2018semanticscholar — Semantic Scholar Academic Graph / S2ORC. | |
| 73 | +- hendricks2020crossref — Crossref metadata as the DOI backbone used for deduplication. | |
| 74 | +- radev2013aclanthology ; acl2024anthology — ACL Anthology. | |
| 75 | +- wang2020mag ; sinha2015mas — Microsoft Academic Graph (discontinued 2021; lineage of OpenAlex). | |
| 76 | +- giles1998citeseer ; wu2015citeseerx ; tang2008arnetminer ; herzog2020dimensions ; peroni2020opencitations — other sources considered. | |
| 77 | +- visser2021comparison ; martinmartin2021coverage — coverage comparisons between bibliographic databases (justify multi-source design). | |
| 78 | +**core (protocol — group B)** | |
| 79 | +- page2021prisma ; page2021prismaexplanation ; rethlefsen2021prismas ; haddaway2022prismaflow ; moher2009prisma — PRISMA 2020 reporting, search reporting (PRISMA-S), flow diagram. | |
| 80 | +- kitchenham2009slr ; kitchenham2013slrprocess ; kitchenham2015ebse — SE systematic-review guidelines (Kitchenham lineage; the 2007 EBSE technical report has no resolvable record, cite the 2015 book). | |
| 81 | +- petersen2008mapping ; petersen2015mapping — systematic mapping study method (the manuscript's design). | |
| 82 | +- tricco2018prismascr ; arksey2005scoping ; grant2009typology — scoping-review methodology and review typology. | |
| 83 | +- wohlin2014snowballing ; garousi2019multivocal — snowballing and grey-literature inclusion rules. | |
| 84 | +- elliott2014living ; elliott2017living — living-review design (daily increment pipeline). | |
| 85 | +- vandeschoot2021asreview ; omaraeide2015textmining ; marshall2019automation — screening automation (active learning). | |
| 86 | +- khraisha2024gpt4screening ; guo2024llmscreening ; syriani2023chatgptscreening ; dennstadt2024llmscreening ; alshami2023chatgptsr ; wang2024zeroshotscreening — evaluations of LLM-assisted screening 2023–2024 (justify and bound the use of LLM classifiers; report their error rates). | |
| 87 | +- zupic2015bibliometric ; donthu2021bibliometric — bibliometric-study design guidelines. | |
| 88 | +**core (integrity)** | |
| 89 | +- nosek2015top ; munafo2017manifesto ; wilkinson2016fair — open-science/TOP/FAIR standards the protocol follows. | |
| 90 | +- piwowar2018oa — OA prevalence/citation advantage (relevant to full-text access limits). | |
| 91 | + | |
| 92 | +## Part IV — Computational Methodology | |
| 93 | + | |
| 94 | +**core (indicators — group C)** | |
| 95 | +- waltman2016review — survey of citation-impact indicators (anchor for the indicator section). | |
| 96 | +- waltman2019fieldnormalization ; waltman2011crown ; purkayastha2019fwci — field normalisation (MNCS, FWCI). | |
| 97 | +- wang2013citationwindow — citation-window choice (crucial for 2023–2026 papers with short windows). | |
| 98 | +- hirsch2005hindex ; waltman2012hindex ; egghe2006gindex — h/g indices and their inconsistencies. | |
| 99 | +- nielsen2021inequality ; radicchi2008universality ; bornmann2008citations — citation inequality (Gini), universality, what citations measure. | |
| 100 | +- stirling2007diversity ; leinster2012diversity ; porter2009interdisciplinarity ; rafols2010diversity ; leydesdorff2019interdisciplinarity — Rao–Stirling diversity and interdisciplinarity measures. | |
| 101 | +- newman2001collaboration ; newman2004coauthorship ; katz1997collaboration ; wagner2017collaboration ; wagner2015cooperation ; lariviere2015teamsize ; wuchty2007teams — co-authorship network and team-size measures. | |
| 102 | +- price1965networks ; garfield1955citation — citation-network foundations. | |
| 103 | +- fire2019goodhart ; hicks2015leiden ; ioannidis2020citation ; sinatra2016impact ; fanelli2016publicationrate — caveats and author-level indicators. | |
| 104 | +**core (science mapping and embeddings — group D)** | |
| 105 | +- small1973cocitation ; kessler1963coupling ; boyack2010accuracy ; klavans2017taxonomy — co-citation, bibliographic coupling, which relation maps best. | |
| 106 | +- vaneck2010vosviewer ; vaneck2014visualizing ; waltman2010unified ; waltman2012classification ; subelj2016clustering ; sjogarde2018granularity ; chen2006citespace ; borner2003domains ; leydesdorff2009globalmap — mapping/clustering of publication networks and granularity choices. | |
| 107 | +- blondel2008louvain ; traag2019leiden ; newman2006modularity ; fortunato2010community — community detection (Leiden is the default). | |
| 108 | +- beltagy2019scibert ; cohan2020specter ; ostendorff2022scincl ; singh2023scirepeval ; reimers2019sbert ; gao2021simcse ; muennighoff2023mteb ; devlin2019bert — scientific-document embeddings and their benchmarks. | |
| 109 | +- mcinnes2018umap ; mcinnes2018umapjoss ; vandermaaten2008tsne — dimensionality reduction for semantic maps. | |
| 110 | +- campello2013hdbscan ; mcinnes2017hdbscan ; schubert2017dbscan — density clustering. | |
| 111 | +- blei2003lda ; blei2006dtm ; dieng2020etm ; grootendorst2022bertopic ; angelov2020top2vec ; greene2014stability ; roder2015coherence ; sievert2014ldavis — topic models, dynamic topics, stability and coherence diagnostics. | |
| 112 | +**core (structural change — group E)** | |
| 113 | +- bai1998breaks ; bai2003breaks ; andrews1993breaks ; andrews1994optimal ; chow1960test ; perron1989crash ; hansen2001structuralchange ; zeileis2002strucchange — multiple-break tests and dating (the 30 Nov 2022 break hypothesis). | |
| 114 | +- killick2012pelt ; fryzlewicz2014wbs ; frick2014smuce ; adams2007bocpd ; truong2020ruptures ; aminikhanghahi2017survey — change-point detection algorithms and reviews. | |
| 115 | +- newey1987hac — HAC standard errors for monthly series. | |
| 116 | +- hamilton1989regime — regime-switching alternative. | |
| 117 | +- bernal2017its ; bernal2018itscontrol ; kontopantelis2015its ; box1975intervention ; brodersen2015causalimpact — interrupted time series / intervention analysis with controls (other fields as controls). | |
| 118 | +- abadie2010synthetic ; abadie2021synthetic — synthetic control. | |
| 119 | +- sun2021eventstudy ; callaway2021did ; goodmanbacon2021did ; roth2022pretest ; roth2023trending — event-study/DiD pitfalls (staggered adoption, pre-trends). | |
| 120 | +- taylor2018prophet ; hyndman2018fpp — forecasting baselines and seasonality handling. | |
| 121 | +**core (emergence detection — group F)** | |
| 122 | +- rotolo2015emerging ; cozzens2010emerging ; small2014emergence ; wang2018emerging ; klavans2017prominence ; boyack2022forecasting ; guo2011mixedindicators ; shibata2008fronts ; porter2019emergence ; salatino2017topicsborn ; salatino2018augur ; small2006growth — definitions and operationalisations of emerging topics (the original-contribution section builds on these). | |
| 123 | +- kleinberg2003burst ; kleinberg2002burstkdd ; mane2004bursts — burst detection. | |
| 124 | +- ke2015sleepingbeauties — delayed recognition. | |
| 125 | +- funk2017disruption ; wu2019disruption ; park2023disruption ; bornmann2020disruptionvalidity ; leibel2024disruptionreview ; holst2024disruptionartefacts — disruption index, its validity and dataset-artefact critique. | |
| 126 | +- uzzi2013novelty ; wang2017novelty ; lee2015novelty ; foster2015tradition — novelty measures and bias against novelty. | |
| 127 | +- milojevic2015cognitiveextent ; bettencourt2009discovery ; chu2021slowed ; lin2023remote — cognitive extent, topological transitions, large-field slowdown. | |
| 128 | +**core (reproducibility of the pipeline — group I)** | |
| 129 | +- gundersen2018reproducibility ; pineau2021reproducibility ; kapoor2023leakage ; kapoor2024reforms ; semmelrock2025reproducibility ; gundersen2022sources ; dodge2019showyourwork ; gebru2021datasheets ; mitchell2019modelcards ; bender2018datastatements — standards applied to WP44's own classifiers and datasets. | |
| 130 | + | |
| 131 | +## Part V — The Artificial Intelligence Transformation | |
| 132 | + | |
| 133 | +**core (foundation models and training)** | |
| 134 | +- vaswani2017attention ; devlin2019bert ; brown2020gpt3 ; raffel2020t5 ; chowdhery2022palm ; openai2023gpt4 ; gemini2023 ; touvron2023llama ; touvron2023llama2 ; dubey2024llama3 ; deepseek2025r1 ; deepseek2025r1nature ; openai2024o1 — model lineage inside the window (open vs closed; reasoning models). | |
| 135 | +- kaplan2020scaling ; hoffmann2022chinchilla ; wei2022emergent ; schaeffer2023mirage — scaling and the emergent-abilities controversy (conflicting findings to be evaluated critically). | |
| 136 | +- ouyang2022instructgpt ; christiano2017rlhf ; bai2022hh ; bai2022constitutional ; rafailov2023dpo ; casper2023rlhfproblems — alignment training methods. | |
| 137 | +- fedus2022switch ; shazeer2017moe ; gu2023mamba ; dao2022flashattention ; hu2021lora ; dettmers2023qlora ; frantar2023gptq ; dettmers2022int8 ; leviathan2023speculative — efficiency: sparsity, state-space models, attention kernels, PEFT, quantisation, decoding. | |
| 138 | +- zhao2023llmsurvey ; naveed2023llmoverview ; minaee2024llmsurvey ; min2023plmsurvey ; wan2023efficientllm ; zhou2024inferencesurvey — LLM surveys (map the sub-literature). | |
| 139 | +**core (reasoning, retrieval, tools, agents)** | |
| 140 | +- wei2022cot ; kojima2022zeroshotcot ; yao2023tot ; shinn2023reflexion — reasoning prompting. | |
| 141 | +- lewis2020rag ; gao2023ragsurvey ; karpukhin2020dpr — retrieval augmentation. | |
| 142 | +- yao2023react ; schick2023toolformer ; qin2023toolllm ; mialon2023augmented — tool use. | |
| 143 | +- wang2024agentsurvey ; xi2023rise ; guo2024multiagent ; park2023generativeagents ; wu2023autogen ; jimenez2024swebench ; zhou2023webarena ; liu2023agentbench ; packer2023memgpt ; mei2024aios ; ge2023llmos — agents: surveys, systems, benchmarks, "LLM as OS". | |
| 144 | +**core (multimodal, vision, speech, generation)** | |
| 145 | +- radford2021clip ; dosovitskiy2021vit ; kirillov2023sam ; alayrac2022flamingo ; li2023blip2 ; liu2023llava ; yin2024mllmsurvey ; zhang2024mmllms ; baltrusaitis2019multimodal — multimodal models and surveys. | |
| 146 | +- ho2020ddpm ; rombach2022ldm ; saharia2022imagen ; ramesh2022dalle2 ; peebles2023dit ; esser2024sd3 — diffusion generation. | |
| 147 | +- radford2023whisper ; baevski2020wav2vec2 ; hsu2021hubert ; mehrish2023speechreview — speech. | |
| 148 | +**core (evaluation)** | |
| 149 | +- hendrycks2021mmlu ; wang2018glue ; srivastava2022bigbench ; liang2022helm ; chiang2024arena ; zheng2023judge ; chang2024evalsurvey — benchmarks and LLM-as-judge; sainz2023contamination ; ott2022benchmarks ; dehghani2021lottery ; raji2021everything ; kiela2021dynabench ; recht2019imagenet — contamination, saturation, benchmark lottery. | |
| 150 | +- ji2023hallucination ; huang2025hallucination — hallucination. | |
| 151 | +**core (safety and alignment)** | |
| 152 | +- hendrycks2021unsolved ; ji2023alignment ; amodei2016concrete ; weidinger2022risks ; bengio2024extremerisks ; bengio2025safetyreport ; anwar2024foundational ; hendrycks2023catastrophic ; shevlane2023extremerisks ; zou2023universal ; wei2023jailbroken ; bender2021parrots ; bubeck2023sparks — safety surveys, risk taxonomies, adversarial robustness; "Sparks" as a contested claim. | |
| 153 | +**core (code)** | |
| 154 | +- chen2021codex ; li2022alphacode ; nijkamp2023codegen ; austin2021mbpp ; roziere2023codellama ; li2023starcoder ; jiang2024codegensurvey ; fan2023llm4se ; hou2024llm4se ; xu2022codelms ; pearce2022asleep ; ziegler2022copilot ; peng2023copilot ; barke2023groundedcopilot ; liang2024usability ; vaithilingam2022expectation — code generation models, SE surveys, productivity and usability evidence. | |
| 155 | +**supplementary** | |
| 156 | +- maslej2024aiindex ; maslej2025aiindex — descriptive context. | |
| 157 | + | |
| 158 | +## Part VI — Systems and Computing Infrastructure | |
| 159 | + | |
| 160 | +**core** | |
| 161 | +- hennessy2019goldenage ; leiserson2020roomattop ; dennard1974scaling ; jouppi2017tpu ; jouppi2023tpuv4 ; reuther2022accelerators ; sze2017efficientdnn ; dally2021gpu ; gholami2024memorywall — architecture and accelerators. | |
| 162 | +- narayanan2021megatron ; shoeybi2019megatronlm ; rajbhandari2020zero ; zhao2023fsdp ; barham2022pathways ; abadi2016tensorflow ; paszke2019pytorch ; ansel2024pytorch2 ; kwon2023vllm ; pope2023inference ; sheng2023flexgen — training/inference systems for large models. | |
| 163 | +- lattner2021mlir ; lattner2004llvm ; chen2018tvm ; tillet2019triton ; cummins2023llmcompiler — compilers for ML and LLMs for compilers. | |
| 164 | +- patterson2021carbon ; patterson2022plateau ; strubell2019energy ; luccioni2023bloom ; luccioni2024powerhungry ; schwartz2020greenai ; henderson2020carbon ; masanet2020datacenters ; devries2023energy ; wu2022sustainableai — sustainable computing / energy and carbon. | |
| 165 | +- shi2016edge ; satyanarayanan2017edge ; zhou2019edgeintelligence ; deng2020edgeintelligence ; letaief2022edgeai6g ; dang2020sixg — edge computing and 6G. | |
| 166 | +- bosshart2014p4 ; mckeown2008openflow ; kreutz2015sdn ; hauser2023p4survey ; sapio2021switchml ; hoiland2018xdp ; vieira2020ebpf — programmable networks, in-network computing, XDP/eBPF. | |
| 167 | +- madhavapeddy2013unikernels ; kuenzer2021unikraft ; packer2023memgpt ; mei2024aios ; ge2023llmos — operating systems: unikernels and LLM-OS proposals. | |
| 168 | +- dean2008mapreduce ; zaharia2016spark ; corbett2013spanner ; verma2015borg ; burns2016kubernetes ; jonas2019serverless ; lamport1998paxos ; castro2002pbft ; gilbert2002cap — distributed systems canon. | |
| 169 | +- pan2024vectordb ; han2023vectordb ; taipalus2024vectordb ; johnson2021faiss ; malkov2020hnsw ; fernandez2023llmdata ; stonebraker2024goesaround ; li2021ai4db — databases: vector DBMS, ANN search, LLMs and data management. | |
| 170 | +- khattab2020colbert ; zhao2024denseretrieval ; robertson2009bm25 ; zhu2023llmir — information retrieval. | |
| 171 | +- mcmahan2017fl ; kairouz2021fl ; li2020flchallenges — federated learning as a systems topic. | |
| 172 | +- preskill2018nisq ; arute2019supremacy ; kim2023utility ; googlequantum2025belowthreshold ; cerezo2021vqa ; bharti2022nisq — quantum computing status 2018–2025. | |
| 173 | +- brohan2023rt2 ; openx2023embodiment ; firoozi2024fmrobotics ; hu2023generalrobots ; kim2024openvla ; black2024pi0 — robotics foundation models (cyber-physical systems). | |
| 174 | +- kerbl2023gaussian ; muller2022instantngp ; barron2021mipnerf ; mildenhall2020nerf ; poole2022dreamfusion ; chen2024gssurvey ; fei2024gssurvey ; tewari2022neuralrendering — graphics/neural rendering (placed here or in Part V as the taxonomy decides). | |
| 175 | + | |
| 176 | +## Part VII — Theory, Algorithms, and Security | |
| 177 | + | |
| 178 | +**core** | |
| 179 | +- cook1971complexity ; karp1972reducibility ; babai2016graphisomorphism ; raz2019bqpph ; ji2021mipre ; chen2022maxflow ; kelley2023threeap ; shor1997factoring ; regev2009lwe — theory landmarks and the quantum/lattice connection. | |
| 180 | +- demoura2021lean4 ; mathlib2020 ; yang2024formalmath ; first2023baldur ; leroy2009compcert ; klein2009sel4 ; jung2017rustbelt ; jung2021saferust ; matsakis2014rust — formal methods, Lean/mathlib, LLM-assisted proving, verified systems, Rust. | |
| 181 | +- nist2024fips203 ; nist2024fips204 ; nist2024fips205 ; bernstein2017pqc — post-quantum standards (2024) and rationale. | |
| 182 | +- gentry2009fhe ; brakerski2014bgv ; cheon2017ckks ; chillotti2020tfhe ; acar2018hesurvey ; marcolla2022fhesurvey ; groth2016snark ; thaler2022proofs ; evans2018mpc — FHE, ZK proofs, MPC. | |
| 183 | +- dwork2006dp ; dwork2014dp ; abadi2016dpsgd ; shokri2017membership ; carlini2021extracting ; carlini2024poisoning — privacy and ML privacy attacks. | |
| 184 | +- greshake2023injection ; perez2022ignore ; liu2023promptinjection ; shayegani2023adversarial ; wei2023jailbroken ; zou2023universal ; yao2024llmsecurity ; das2025llmsecurity ; fang2024oneday ; deng2023pentestgpt — LLM security: prompt injection, jailbreaks, surveys, offensive use. | |
| 185 | +- szegedy2014intriguing ; goodfellow2015fgsm ; kocher2019spectre ; lipp2018meltdown — adversarial examples and microarchitectural attacks (pre-window security landmarks). | |
| 186 | +- nakamoto2008bitcoin ; zheng2018blockchain ; monrat2019blockchain ; werner2022defi — blockchain/DeFi. | |
| 187 | + | |
| 188 | +## Part VIII — Scientific Knowledge Networks | |
| 189 | + | |
| 190 | +**core** | |
| 191 | +- newman2001collaboration ; newman2004coauthorship ; katz1997collaboration ; wuchty2007teams ; lariviere2015teamsize ; wagner2015cooperation ; wagner2017collaboration ; lin2023remote — co-authorship, teams, international collaboration. | |
| 192 | +- small1973cocitation ; kessler1963coupling ; boyack2010accuracy ; klavans2017taxonomy ; price1965networks ; garfield1955citation — citation-network constructs. | |
| 193 | +- blondel2008louvain ; traag2019leiden ; newman2006modularity ; fortunato2010community — community structure. | |
| 194 | +- stirling2007diversity ; leinster2012diversity ; porter2009interdisciplinarity ; rafols2010diversity ; leydesdorff2019interdisciplinarity — interdisciplinarity/diffusion measures. | |
| 195 | +- nielsen2021inequality ; radicchi2008universality ; sinatra2016impact ; ioannidis2020citation — inequality and author-level impact. | |
| 196 | +- ahmed2023industry ; ahmed2020dedemocratization ; klinger2020narrowing ; frank2019citationgraphs ; martinezplumed2021benchmarks — industry/academia, concentration and narrowing in AI. | |
| 197 | +- piwowar2018oa ; lin2020arxivcs ; lariviere2014arxiv ; fraser2020biorxiv ; fraser2021preprints ; sutton2017arxivcs — preprints and OA effects on diffusion. | |
| 198 | +- funk2017disruption ; wu2019disruption ; park2023disruption ; uzzi2013novelty ; ke2015sleepingbeauties — disruption/novelty/delayed recognition on the citation graph. | |
| 199 | +- osborne2024huggingface ; castano2023hfcarbon ; wolf2020transformers ; lhoest2021datasets ; longpre2024dataprovenance — software/dataset ecosystems as knowledge networks (Hugging Face). | |
| 200 | + | |
| 201 | +## Part IX — Empirical Results | |
| 202 | + | |
| 203 | +Cite methods references from Part IV when reporting each result; in addition: | |
| 204 | +- wang2013citationwindow ; waltman2019fieldnormalization — when presenting short-window and normalised citation results. | |
| 205 | +- bai2003breaks ; killick2012pelt ; truong2020ruptures ; bernal2017its ; abadie2021synthetic ; roth2023trending — when presenting break dates, ITS and control-field comparisons. | |
| 206 | +- liang2024mapping ; liang2024monitoring ; kobak2025delving ; geng2024writingstyle ; gray2024contamination ; latona2024reviewlottery — external estimates of LLM-modified text/reviews to compare against the corpus's own vocabulary shifts. | |
| 207 | +- ott2022benchmarks ; martinezplumed2021benchmarks — benchmark dynamics to compare with observed benchmark-paper growth. | |
| 208 | +- bornmann2021growth ; bornmann2015growth ; fanelli2016publicationrate — reference growth rates. | |
| 209 | +- maslej2024aiindex ; maslej2025aiindex ; arxiv2026stats — external descriptive series for triangulation (reports/statistics pages, not peer-reviewed). | |
| 210 | +- greene2014stability ; roder2015coherence ; subelj2016clustering — when reporting topic/cluster validation. | |
| 211 | + | |
| 212 | +## Part X — The Next Research Frontiers | |
| 213 | + | |
| 214 | +- wang2023aiscience ; abramson2024alphafold3 ; lin2023esm2 ; hayes2025esm3 ; watson2023rfdiffusion ; dauparas2022proteinmpnn ; nguyen2024evo ; cui2024scgpt ; kochkov2024neuralgcm ; bi2023pangu ; lam2023graphcast ; davies2021mathematics ; fawzi2022alphatensor ; degrave2022tokamak ; romeraparedes2024funsearch ; trinh2024alphageometry ; merchant2023gnome — AI for science frontier. | |
| 215 | +- lu2024aiscientist ; gottweis2025coscientist ; boiko2023coscientist ; si2024novelideas — autonomous research agents (hypotheses to test, not established findings). | |
| 216 | +- bengio2025safetyreport ; anwar2024foundational ; ji2023alignment — open safety problems. | |
| 217 | +- googlequantum2025belowthreshold ; kim2023utility ; nist2024fips203 — quantum/post-quantum frontier. | |
| 218 | +- yang2024formalmath ; first2023baldur — formal reasoning frontier. | |
| 219 | +- gu2023mamba ; deepseek2025r1 ; openai2024o1 — architectures and reasoning models. | |
| 220 | +- brohan2023rt2 ; black2024pi0 ; kim2024openvla — embodied AI. | |
| 221 | +- luccioni2024powerhungry ; devries2023energy — sustainability constraints. | |
| 222 | +- rotolo2015emerging ; boyack2022forecasting ; salatino2018augur ; krenn2023forecasting — how emergence/forecast claims should be stated. | |
| 223 | +- shneiderman2020hcai ; amershi2019guidelines ; weisz2024genaidesign ; zamfirescu2023johnny ; subramonyam2024gulf ; lee2022coauthor ; bansal2021wholeexceeds ; vaccaro2024humanai ; morris2023scientists — human–AI interaction frontier. | |
| 224 | + | |
| 225 | +## Part XI — Discussion and Conclusion | |
| 226 | + | |
| 227 | +- hicks2015leiden ; fire2019goodhart ; bornmann2008citations ; wang2017novelty — limits of citation-based evidence. | |
| 228 | +- messeri2024illusions ; birhane2023science ; dwivedi2023sowhat ; lund2023chatgpt ; stokelwalker2023author ; thorp2023notauthor ; nature2023groundrules ; else2023abstracts ; gao2023abstracts ; conroy2023paperfromscratch ; checco2021aipeerreview ; liang2024feedback — implications for scientific practice, authorship and peer review (liang2024feedback: large-scale evaluation of LLM-generated paper feedback). | |
| 229 | +- noy2023productivity ; brynjolfsson2023genaiwork ; dellacqua2023jagged ; eloundou2024gpts ; gao2024aibenefits ; hao2024aiexpands — productivity/impact evidence (external, mostly non-CS). | |
| 230 | +- park2023disruption ; holst2024disruptionartefacts ; leibel2024disruptionreview — why disruption-based conclusions must be hedged. | |
| 231 | +- visser2021comparison ; alperin2024openalex ; culbert2025openalexcoverage — source-coverage limitations. | |
| 232 | +- roth2022pretest ; sun2021eventstudy — why temporal association is not causation. | |
| 233 | + | |
| 234 | +## Appendices | |
| 235 | + | |
| 236 | +- Appendix (taxonomy): openalex2024topics ; waltman2012classification ; sjogarde2018granularity ; klavans2017taxonomy ; leydesdorff2009globalmap. | |
| 237 | +- Appendix (search strategies): rethlefsen2021prismas ; wohlin2014snowballing ; garousi2019multivocal ; haddaway2022prismaflow. | |
| 238 | +- Appendix (mathematical derivations): bai1998breaks ; andrews1993breaks ; newey1987hac ; killick2012pelt ; adams2007bocpd ; stirling2007diversity ; leinster2012diversity ; traag2019leiden ; mcinnes2018umap ; campello2013hdbscan ; funk2017disruption ; kleinberg2003burst. | |
| 239 | +- Appendix (robustness): wang2013citationwindow ; bornmann2020disruptionvalidity ; greene2014stability ; subelj2016clustering ; roth2023trending ; abadie2021synthetic. | |
| 240 | +- Appendix (data dictionaries): priem2022openalex ; kinney2023s2ag ; ley2009dblp ; hendricks2020crossref ; wilkinson2016fair ; gebru2021datasheets. | |
| 241 | +- Appendix (reproducibility): acm2020badging ; pineau2021reproducibility ; kapoor2024reforms ; nosek2015top ; munafo2017manifesto ; collberg2016repeatability ; stodden2018journalpolicy ; hutson2018crisis ; baker2016reproducibility ; henderson2018drlmatters ; lipton2019troubling ; sambasivan2021datawork ; paullada2021discontents ; koch2021datasetlife ; bommasani2023fmti ; magnusson2023olmo. | |
| 242 | + | |
| 243 | +## Not available (verification failed — do not cite until a verifiable record is found) | |
| 244 | + | |
| 245 | +- Kitchenham & Charters 2007 EBSE technical report (no DOI/stable URL) → use kitchenham2015ebse. | |
| 246 | +- McCarthy et al. 1955/2006 Dartmouth proposal (AI Magazine DOI not resolvable via Crossref/OpenAlex). | |
| 247 | +- Sculley et al. 2018 "Winner's Curse?" (OpenReview blocks automated access; HTTP 403). | |
| 248 | +- Papers with Code (site retired; redirects to Hugging Face). | |
| 249 | +- dblp statistics page (bot challenge); use ley2009dblp and the DBLP XML dump documentation instead. | |
| 250 | +- OpenAI "Sora" technical report page (403 / empty archived page). | |
| 251 | +- Haunschild 2024 OpenAlex affiliation-quality paper and a generic "LLMs for systematic review screening" item (titles could not be matched; placeholders dropped). | |
added
manuscript/references.bib
+6641 −0
Large diff (6,641 lines) is not loaded by default to keep the page fast.Load diff
added
manuscript/references_part03.bib
+339 −0
Large diff (339 lines) is not loaded by default to keep the page fast.Load diff
added
manuscript/references_verification.csv
+589 −0
Large diff (589 lines) is not loaded by default to keep the page fast.Load diff
added
manuscript/references_verification_part03.csv
+34 −0
Large diff (34 lines) is not loaded by default to keep the page fast.Load diff
modified
manuscript/sections/part04_methodology.tex
+555 −1
Large diff (557 lines) is not loaded by default to keep the page fast.Load diff