1% ============================================================================2% Part V — The Artificial Intelligence Transformation3% Written by Agent E (AI Research Specialist), 2026-11-03, from4% literature_synthesis/agent_E/{review.md, claims.csv, contributions.csv,5% open_problems.csv}. Evidence rules: landmark papers are known from metadata6% and abstracts only (never full text); preprints are named as preprints; corpus7% numbers enter only through the macros of tables/numbers.tex or the generated8% tables tab_ai_*.tex (tables/make_tables.py, tables_part05). Claim identifiers9% [E-nn] in comments refer to claims.csv. No causal claim about ChatGPT.10% ============================================================================1112\section{Scope, evidence base and reading rules}13\label{sec:ai-scope}1415This part synthesises the nine categories of the taxonomy that the ontology16groups under \emph{AI and machine learning}: large language models, foundation17models, multimodal AI, generative AI, AI agents and autonomous systems,18reinforcement learning, machine-learning theory, deep-learning architectures,19and AI safety and alignment. Three kinds of evidence are combined, and the20reader should keep them apart. The first is the corpus itself: the yearly and21monthly counts, growth rates, structural-change estimates and citation flows of22Part~IV, read here only through the generated tables of this part23(Tables~\ref{tab:ai_growth}--\ref{tab:ai_landmarks}) and the macros of24\texttt{tables/numbers.tex}. The second is the automatically selected landmark25set of the area, \SelectionByAreaAiMl{} records chosen by field-normalised26citation impact within each primary category (Part~IV,27Section~\ref{sec:m-score}), of which the agent read the title, venue, type,28identifiers, citation windows and, where OpenAlex stores one, the abstract.29The third is the verified reference list assembled in Phase~530(\texttt{references.bib} and \texttt{references\_part05.bib}), which carries31the technical lineage of the field whether or not the corpus ranking surfaced32it. Nothing in this part describes the content of a paper beyond what its33abstract or bibliographic record states; a result known only from an abstract34is introduced as what ``the abstract reports''. Works that exist only as35arXiv, SSRN or medRxiv deposits are called preprints throughout, and a36preprint's claim is never upgraded to a peer-reviewed finding by the number of37times it has been cited. The boundary of 30~November 2022 is a reference date38for the time-series models; the estimates below describe when the series39changed, not why.4041\subsection{What the landmark set is, and what it is not}42\label{sec:ai-landmarks}4344The landmark candidates of the area were screened record by record before any45of them was used, and the outcome conditions everything that follows. Of the46\SelectionByAreaAiMl{} records, the screening flagged 44 as unreliable for the47purpose of a critical synthesis (Table~\ref{tab:ai_landmark_flags}; the record-level flags are in the file48\texttt{contributions.csv} of the agent's directory under49\texttt{literature\_synthesis}).\footnote{The50flags are a human judgement by the specialist agent from identifiers, citation51counts, venue, title and abstract; they were not produced by a script and52must be audited (Part~III, Phase~8). The counts in this paragraph are those of53the flag file, reproduced in Table~\ref{tab:ai_landmark_flags}.} Thirteen are54\emph{identifier mismatches}: the title and abstract belong to an obscure work55while the arXiv identifier and the 36-month citation count belong to a famous56one. A 2026 article on translating natural language into MongoDB queries57carries the arXiv identifier of the retrieval-augmented-generation paper58\citep{lewis2020rag} and its citations; a Portuguese-language essay on59copyright carries that of GPT-3 \citep{brown2020gpt3}; an essay on60``theatrical compliance'' carries that of chain-of-thought prompting61\citep{wei2022cot}; a preprint on an intensive-care simulation carries that of62Llama~2 \citep{touvron2023llama2}; a ``reproducible experiment runner'' for the63Adam optimiser carries more than forty thousand citations. Three records are64\emph{abstract mismatches}, in which a correct title (the GPT-4 technical65report, \citealp{openai2023gpt4}; ``GPTs are GPTs'',66\citealp{eloundou2024gpts}) is paired with the abstract of another work.67Eleven are \emph{implausible years}: mixup, the variational recurrent network68and several 2013--2019 bandit papers appear as 2024--2026 works, four of them69with April--June 2026 arXiv identifiers, so that their field-normalised scores70are computed against the wrong cohort. Seven are \emph{normalisation71artefacts}, dissertations and unindexed articles with two to twenty citations72and NCS36 above five hundred, because the reference mean for a rare document73type in a young cohort is close to zero. Ten are \emph{off scope}:74control-theoretic consensus papers assigned to AI agents, an essay on75educational-technology ``solutionism'' assigned to reinforcement learning.76% [E-02, E-03]7778The defect is not uniformly distributed. Identifier and abstract problems79concentrate in the large-language-model category, where eight of80thirty-three records are affected, and misdated classics concentrate in81machine-learning theory and deep-learning architectures. The mechanism lies82upstream of this review, in the merging of work records by the source83database and in the second-pass arXiv and title linkage of Phase~184(Section~\ref{sec:m-linkage}); it is reported here because the AI area is where85it bites hardest, the arXiv identifiers of the most-cited papers of the decade86being exactly the ones most likely to be attached to the wrong record. Two87consequences are drawn. First, the clean count of citation-ranked AI landmarks88is at most 190, and every ``most influential papers'' statement of Part~IX89that concerns AI must be computed on records whose DOI, arXiv identifier and90title are mutually consistent. Second, the field-normalised score should be91recomputed with document-type-specific reference sets and with the 202692cohort excluded; this is a recommendation to the director, not something done93here. % [E-04]9495What the set does contain is instructive in its own right96(Table~\ref{tab:ai_contribution_types}). Judged from the abstract for 19097records and from title, venue and type for the 44 without one, 84 of the 23498candidates are surveys, 67 methods, 33 empirical studies, 29 theory or position99papers, 17 engineering systems and only 4 benchmarks. % [E-05]100Within the LLM and generative-AI categories the citation-ranked top is101occupied by education and commentary papers written in the months after102November 2022 \citep{kasneci2023chatgptgood, lo2023rapidreview,103rudolph2023assessment, chan2023students, baidooanu2023chatgptedu,104dwivedi2023sowhat, vandis2023priorities, lim2023ragnarok}, not by the papers105that introduced the models. % [E-06]106This is a property of ranking by field-normalised impact inside a multi-label107category that spans computer science, education and information systems: a108commentary in an education journal is normalised against education109journals, and the first wave of ChatGPT commentary was cited at rates that110no technical paper of its cohort matched within thirty-six months. The111technical lineage---GPT-4, LLaMA, InstructGPT, chain-of-thought, DPO, QLoRA,112BLIP-2, Mamba---is present in the set only where an arXiv record was linked,113and sometimes, as just shown, linked to the wrong record. The synthesis114therefore grounds its technical claims on the verified reference list and uses115the landmark set to characterise what the corpus \emph{rewards} rather than116what the field \emph{invented}. Table~\ref{tab:ai_landmarks}, placed at the end of this part, lists all117234 candidates with their flag and contribution type.118119\input{tables/tab_ai_landmark_flags}120\input{tables/tab_ai_contribution_types}121122\subsection{The corpus picture of the nine categories}123\label{sec:ai-corpus}124125Table~\ref{tab:ai_growth} gives the primary-category volume, the compound126annual growth rates of the baseline (2018--2022) and post (2023--2025) periods127and the share of eligible works for each category; Table~\ref{tab:ai_counts}128gives the yearly multi-label counts, which are several times larger because129the AI categories overlap heavily and a work counts in every category it is130assigned to. The ordering is the same under either definition. Large language131models grow fastest in the post period, generative AI, multimodal AI,132foundation models and AI safety follow, reinforcement learning and133machine-learning theory accelerate modestly, deep-learning architectures grow134slightly more slowly than in the baseline, and AI agents reverse a negative135baseline growth rate. % [E-07]136The share of eligible works carrying the LLM label as primary category rises by137more than an order of magnitude between 2018 and 2025. % [E-08]138Two cautions apply. The 2026 column of Table~\ref{tab:ai_counts} covers139publications dated up to the snapshot of 14~October 2026 and is incomplete;140and the categories are defined by rules over topics, arXiv classes, venues and141keyword patterns (Section~\ref{sec:m-classification}), so that a work on142reward modelling for language models is an LLM work and not a reinforcement-learning143work, a point that matters for Section~\ref{sec:ai-rl}.144A third caution concerns the flows of Table~\ref{tab:ai_flows}. Because a145cited work can carry several categories, the row shares of a citing category146do not sum to one across the columns shown, and a reference from an LLM paper147to a paper that is both an LLM and an NLP work counts under both destinations;148the ``self'' share of a category is therefore an upper bound on its149intellectual closure, and changes in shares are more informative than their150levels.151152\input{tables/tab_ai_growth}153\input{tables/tab_ai_counts}154155The structural-change estimates are in Table~\ref{tab:ai_breaks} and, for the156alternative event dates, Table~\ref{tab:ai_altdates}. For the LLM series the157interrupted time series of eq.~\eqref{eq:its} fitted at158\ItsEventsChatgpt{} gives a level change of159\ItsSeriesCatLlmItsChatgptLevelChange{} log points160(HAC $p = \ItsSeriesCatLlmItsChatgptLevelP$) and a slope change of161\ItsSeriesCatLlmItsChatgptSlopeChange{} log points per month162($p = \ItsSeriesCatLlmItsChatgptSlopeP$) over a pre-period slope of163\ItsSeriesCatLlmItsChatgptPreSlope{}; the placebo distribution of164Section~\ref{sec:m-placebo} places the level statistic at a placebo165$p$-value of \ItsSeriesCatLlmPlaceboChatgptLevelTPlaceboP{} and the slope166statistic at \ItsSeriesCatLlmPlaceboChatgptSlopeTPlaceboP{}; the167Bai--Perron procedure selects \ItsSeriesCatLlmBaiPerronMBic{} break, dated168in Table~\ref{tab:ai_breaks} to the middle of 2022, and the Bayesian169on-line change-point posterior mass between December 2022 and March 2023 is170\ItsSeriesCatLlmBocpdCpProbTwoZeroTwoTwoOneTwoToTwoZeroTwoThreeZeroThree{}.171Excluding three months on either side of the event (the donut specification)172changes the level estimate to173\ItsSeriesCatLlmItsChatgptDonutLevelChange{}174($p = \ItsSeriesCatLlmItsChatgptDonutLevelP$). % [E-09]175The same configuration---no significant level jump at the reference month, a176large and significant slope change, a placebo $p$-value for the level that177does not single out December 2022, a data-driven break before or well after178it---holds for multimodal AI (slope change179\ItsSeriesCatMultimodalAiItsChatgptSlopeChange) and generative AI180(\ItsSeriesCatGenerativeAiItsChatgptSlopeChange). AI agents (level181\ItsSeriesCatAiAgentsItsChatgptLevelChange, $p =182\ItsSeriesCatAiAgentsItsChatgptLevelP$; slope183\ItsSeriesCatAiAgentsItsChatgptSlopeChange), reinforcement learning (level184\ItsSeriesCatReinforcementLearningItsChatgptLevelChange, $p =185\ItsSeriesCatReinforcementLearningItsChatgptLevelP$) and AI safety (level186\ItsSeriesCatAiSafetyItsChatgptLevelChange, $p =187\ItsSeriesCatAiSafetyItsChatgptLevelP$) show \emph{negative} level changes188followed by positive slope changes, a dip and a subsequent acceleration whose189reading is taken up category by category below. Deep-learning architectures is190the only AI category whose slope change191(\ItsSeriesCatDeepLearningArchitecturesItsChatgptSlopeChange, $p =192\ItsSeriesCatDeepLearningArchitecturesItsChatgptSlopeP$) is not193distinguishable from zero. % [E-10]194195\input{tables/tab_ai_breaks}196\input{tables/tab_ai_altdates}197198The alternative dates of Section~\ref{sec:m-altdates} order themselves199monotonically. Level effects estimated at the GPT-3 (\ItsEventsGptThree) and200DALL-E~2 (\ItsEventsDalleTwo) dates are negative for every AI category,201whereas those at the GPT-4 (\ItsEventsGptFour) and o1 (\ItsEventsOOne) dates202are positive and significant for LLM, multimodal and generative AI: for the203LLM series, \ItsSeriesCatLlmItsGptThreeLevelChange{} at the GPT-3 date204($p = \ItsSeriesCatLlmItsGptThreeLevelP$), \ItsSeriesCatLlmItsGptFourLevelChange{}205at the GPT-4 date ($p = \ItsSeriesCatLlmItsGptFourLevelP$) and206\ItsSeriesCatLlmItsOOneLevelChange{} at the o1 date207($p = \ItsSeriesCatLlmItsOOneLevelP$). % [E-11]208Each of these is a separate fit with the event moved, and a positive level209effect at a later date in a convex series is what a steepening trend produces210whether or not anything happened at that date. The reading that the evidence211supports is therefore restricted: the acceleration of the AI literature is a212change of trend that the data-driven procedures place in the middle of 2022213and that steepened through 2023--2024; the monthly series cannot attribute it214to one product release, and this part does not. % [E-12]215216\subsection{Knowledge flows and detected emergence}217\label{sec:ai-flows}218219Table~\ref{tab:ai_flows} reports, for each AI category and for 2019, 2022 and2202025, the destination of its in-corpus references, i.e.\ the row shares221$s_{c \to c'}(t)$ of eq.~\eqref{eq:flow}. The large-language-model category is222the one whose profile changed most. In 2019 and 2022 roughly seven in ten of223its in-corpus references went to natural language processing; in 2025 about a224third do, while more than a quarter go to the LLM category itself and225measurable shares go to human--computer interaction and software engineering.226% [E-13]227Generative AI moved from a vision-centred profile to one split between vision,228LLM and HCI; AI safety, which cited LLM work at a negligible rate in 2019,229directs a tenth of its references there in 2025 and slightly more to HCI.230% [E-14]231Reinforcement learning moves the other way, citing robotics more and LLM work232almost not at all, and machine-learning theory and deep-learning architectures233are essentially unchanged. % [E-15]234At the level of categories, then, the LLM literature has become a citation hub235for generative AI, safety and agents while remaining peripheral to236reinforcement learning, theory and architectures. This is the corpus-level237answer to RQ5 that this part can give; it is consistent with the convergence238thesis of \citet{bommasani2021foundation}, but a change in which category a239field cites is not evidence that its methods converged, and240Section~\ref{sec:ai-synthesis} returns to the distinction. % [E-16]241242\input{tables/tab_ai_flows}243244The preregistered known-emergence set of Section~\ref{sec:m-emerging} contains245eleven AI concepts, and their lead times are the first empirical test of246RQ10 in this area. The detector flags \emph{large language models} in247\KnownLargeLanguageModelsDetectionMonth{},248\KnownLargeLanguageModelsLeadMonths{} months before the peak-growth month249\KnownLargeLanguageModelsPeakGrowthMonth{}, on250\KnownLargeLanguageModelsTotalWorks{} matching works; \emph{reasoning models}251\KnownReasoningModelsLeadMonths{} months ahead; retrieval-augmented generation252\KnownRetrievalAugmentedGenerationLeadMonths{} months; vision-language models253\KnownVisionLanguageModelsLeadMonths{}; diffusion models254\KnownDiffusionModelsLeadMonths{}; LLM agents \KnownLlmAgentsLeadMonths{};255state-space sequence models \KnownStateSpaceSequenceModelsLeadMonths{};256mixture of experts \KnownMixtureOfExpertsLeadMonths{}; speculative decoding257\KnownSpeculativeDecodingLeadMonths{}; AI code generation258\KnownAiCodeGenerationLeadMonths{}; and mechanistic interpretability259\KnownMechanisticInterpretabilityLeadMonths{}, a negative lead in which the260peak-growth month precedes detection on only261\KnownMechanisticInterpretabilityTotalWorks{} matching works. The median lead262over the whole set is \KnownSetMedianLeadMonths{} months. % [E-17]263The long lead for language models is not foresight: the regular expression264matches the pre-2020 literature on ``language models'' and the detector fires265on its slow growth years before the 2023 surge. The short leads for266state-space models, mixture of experts and speculative decoding are the267honest result, and they say that topics driven by a single widely adopted268preprint are detected only months before their growth peaks.269270\section{Large language models and foundation models}271\label{sec:ai-fm}272273\subsection{The established lineage}274275The architecture that dominates the window is the decoder-only Transformer276\citep{vaswani2017attention} trained autoregressively at scale277\citep{radford2019gpt2, brown2020gpt3, chowdhery2022palm}, with the encoder278and encoder--decoder variants of the baseline period \citep{devlin2019bert,279raffel2020t5} surviving mainly as embedding models. Two empirical regularities280from before the window organise the technical literature inside it. The first281is the scaling law, in its original form a power law of loss in parameters $N$282and tokens $D$ \citep{kaplan2020scaling} and in the compute-optimal form of283\citet{hoffmann2022chinchilla}, which, for a fixed training budget $C \approx2846ND$, allocates roughly equal factors of growth to $N$ and $D$ and therefore285implies far more tokens per parameter than the earlier prescription. The286second is the sparse mixture-of-experts layer \citep{shazeer2017moe,287fedus2022switch}, which decouples parameter count from per-token compute. The288open-weight lineage LLaMA, Llama~2 and Llama~3 \citep{touvron2023llama,289touvron2023llama2, dubey2024llama3}---all three preprints, the first of which290is in the landmark set---follows the Chinchilla allocation; the LLaMA abstract291reports that a 13-billion-parameter model trained only on publicly available292data outperforms GPT-3 on most benchmarks, a claim that the abstract makes and293that this review has not verified. % [E-18]294The closed frontier releases are documented by technical reports without295training details \citep{openai2023gpt4, gemini2023, openai2024o1}; they are296cited here as the dated artefacts they are, and the GPT-4 record in the297landmark set, as noted above, carries someone else's abstract. The LLM surveys298of the window \citep{zhao2023llmsurvey, naveed2023llmoverview,299minaee2024llmsurvey, min2023plmsurvey} map a sub-literature that by 2025 had300its own citation core (Table~\ref{tab:ai_flows}); the journal version of the301first of them is in the landmark set and its abstract organises the field302into pre-training, adaptation, utilisation and evaluation.303304The most visible post-2022 development in the landmark set is not a new305general model but the diffusion of the foundation-model recipe into domains306with their own data. Self-supervised pathology encoders trained on hundreds of307thousands of whole-slide images \citep{chen2024uni, lu2024conch,308wang2024chief, huang2023plip, ding2025titan}, a radiology model trained on309two- and three-dimensional scans \citep{wu2025radfm}, a dermatology model on310two million clinical images \citep{yan2025panderm}, a brain-MRI model311\citep{tak2026brainiac}, an Earth-system model trained on more than a million312hours of geophysical data \citep{bodnar2025aurora}, a tabular model that the313abstract describes as a prior-data fitted network outperforming314gradient-boosted trees on small datasets \citep{hollmann2025tabpfn},315time-series representation models \citep{jin2023timellm, liu2024timesurl} and316a language model fine-tuned on ten million human choices to predict behaviour317across psychological experiments \citep{binz2025centaur} are, with the318exception of the time-series preprint, peer-reviewed journal or conference319papers, and each abstract reports outperforming task-specific baselines on a320battery of benchmarks. % [E-19]321What the abstracts do not report, and what the reviews in the same set list as322unresolved \citep{paschali2025fmradiology, zhang2026datacentric}, is external,323multi-site validation under distribution shift; the claim that domain324foundation models will displace task-specific pipelines in regulated settings325is speculative and is carried to Part~X as such. % [E-24]326327Below the frontier, the practical growth of the open-weight ecosystem328concentrated on adaptation. Low-rank adaptation \citep{hu2021lora}, its329four-bit variant QLoRA, whose abstract reports fine-tuning a 65-billion-parameter330model on a single 48\,GB GPU \citep{dettmers2023qlora}, the zero-initialised331adaption prompts of LLaMA-Adapter \citep{zhang2023llamaadapter} and the332parameter-efficient fine-tuning surveys \citep{han2024peftsurvey} define a333regime in which a research group without frontier compute can specialise a334model in hours. The QLoRA abstract's statement that its best model reaches335``99.3\,\% of the performance level of ChatGPT'' on the Vicuna benchmark is a336judge-based measurement, a point to which Section~\ref{sec:ai-eval} returns.337% [E-20]338339\subsection{Data as the contested resource}340341The compute-optimal allocation of \citet{hoffmann2022chinchilla} turned342training data from an afterthought into the binding constraint, and the343landmark set records the consequences from three sides without resolving any344of them. On the supply side, LAION-5B \citep{schuhmann2022laion5b} and the345public-data training of LLaMA \citep{touvron2023llama} show that web-scale346corpora could be assembled outside the frontier laboratories; the surveys of347data-centric foundation models in healthcare \citep{zhang2026datacentric} and348of multimodal fusion on low-quality data \citep{zhang2026lowqualityfusion}349describe, from the methods side, how quality, annotation, privacy and350missingness rather than volume limit the domain models of351Section~\ref{sec:ai-fm}. On the integrity side, the poisoning study of352\citet{alber2025poisoning} reports that a vanishingly small fraction of353deliberately false tokens in a public pretraining corpus suffices to shift a354model's medical answers, which turns the openness of the data pipeline into an355attack surface. On the legal side, the one landmark that addresses the356provenance of training data is a Portuguese-language essay on copyright whose357record, as shown above, carries GPT-3's identifier and citations; the358question it raises is real and the corpus's answer is an artefact. The three359sides share a measurement gap that no abstract in the set closes: the360composition of the corpora on which the frontier models were trained is not361disclosed, so that neither the scaling-law prescriptions nor the362contamination caveats of Section~\ref{sec:ai-eval} can be checked against the363models that matter most. Data, not architecture, is where the open and closed364lineages of the window diverge, and it is the resource whose exhaustion,365curation and legal status the next part of the manuscript lists among the366frontiers.367368\subsection{Contested claims: emergence, sparks and pure-RL reasoning}369370Three claims about what scaled language models are have been argued largely in371preprints, and the corpus can record the argument but not settle it. The first372is emergence. \citet{wei2022emergent} report abilities that appear373discontinuously as scale increases; \citet{schaeffer2023mirage} show that for374the tasks examined the discontinuities are produced by discontinuous375metrics---exact-match accuracy on multi-token answers---and largely vanish376under continuous or token-level ones. Both are preprints, the disagreement is377about measurement rather than data, and the manuscript presents it as such.378% [E-21]379The second is ``Sparks of Artificial General Intelligence''380\citep{bubeck2023sparks}, which reports qualitative experiments on an early,381unreleased version of GPT-4 and, in the abstract's words, contends that the382model exhibits more general intelligence than previous systems. The record is383among the most cited of the landmark set, it is a preprint, its experiments384are not reproducible because the model version is unavailable, and the385critique that scale does not confer understanding \citep{bender2021parrots}386is in the same set with a comparable citation count. % [E-22]387The third is the reasoning turn: the o1 system card \citep{openai2024o1}388moved computation to inference time, and the DeepSeek-R1 paper, whose389peer-reviewed version is in the landmark set \citep{deepseek2025r1nature}390alongside the preprint \citep{deepseek2025r1}, reports that reasoning ``can be391incentivized through pure reinforcement learning, obviating the need for392human-labelled reasoning trajectories''. The claim is peer reviewed; its393generality beyond domains with verifiable rewards---mathematics, code---is not394tested by anything in the corpus, and it is the first entry of the open395problems in Section~\ref{sec:ai-synthesis}. % [E-23]396397\section{Alignment and post-training}398\label{sec:ai-align}399400The recipe that turned a pretrained language model into an assistant was in401place before the window: a reward model fitted to human preference402comparisons \citep{christiano2017rlhf} and a policy optimised against it with403a proximal objective \citep{schulman2017ppo}, applied to instruction following404in InstructGPT \citep{ouyang2022instructgpt} and extended with helpfulness and405harmlessness data \citep{bai2022hh} and with model-written critiques against a406written constitution \citep{bai2022constitutional}. Inside the window the407method's two moving parts were simplified and questioned. Direct preference408optimisation \citep{rafailov2023dpo} rewrites the RLHF objective so that the409policy is fitted directly to preference pairs through a closed-form410log-ratio loss, removing the separate reward model and the on-policy sampling;411the NeurIPS record is in the landmark set, without an abstract in the file.412The open problems of the reward-model route---reward hacking, preference413mis-specification, evaluator disagreement, the gap between stated and revealed414preferences---are catalogued by \citet{casper2023rlhfproblems}, a preprint415whose taxonomy is nevertheless the standard reference. % [E-25]416The corpus adds a sociological observation to this methodological one.417Reinforcement learning as a category directs less than one per cent of its418in-corpus references to LLM work in 2025 (Table~\ref{tab:ai_flows}), while the419LLM category's own self-citation share grew to more than a quarter. The420``return of reinforcement learning'' through post-training is therefore a421phenomenon of the LLM literature, which absorbed PPO, DPO and the group-relative422policy optimisation of DeepSeek-R1 as its own methods; the classification423separates the two communities even where the mathematics is shared, and424Section~\ref{sec:ai-rl} draws the consequence for any claim about the fate of425reinforcement learning.426427That aligned behaviour is brittle under optimisation pressure is established428by three lines of work. Universal adversarial suffixes found by gradient-guided429search transfer across models \citep{zou2023universal}; the ``jailbroken''430analysis attributes failures to competing objectives and mismatched431generalisation between pretraining and safety training432\citep{wei2023jailbroken}; and PAIR, a peer-reviewed method in the landmark433set, is described in its abstract as an attacker language model that generates434semantic jailbreaks against a target with black-box access, in about twenty435queries \citep{chao2025pair}. % [E-26]436The first two are preprints and the third was published in 2025; together they437show that every public alignment method of the window can be defeated by a438modest adaptive budget. Whether this brittleness is intrinsic to439preference-based alignment or an artefact of current scale and data is the440question that none of the landmark abstracts answers, and it is carried to441Part~X.442443\section{Reasoning, retrieval, tool use and code}444\label{sec:ai-reason}445446Chain-of-thought prompting \citep{wei2022cot, kojima2022zeroshotcot},447tree-structured search over thoughts \citep{yao2023tot} and verbal448self-reflection across episodes \citep{shinn2023reflexion} are the prompting449lineage of the reasoning literature; their evidence is benchmark gains, mostly450on arithmetic and symbolic tasks, and several of the records are preprints.451The reasoning-model turn of 2024--2025 \citep{openai2024o1,452deepseek2025r1nature} internalised the procedure by training the model to453produce long intermediate traces and by spending compute at inference rather454than at training. The detector of Section~\ref{sec:m-emerging} flagged455\emph{reasoning models} in \KnownReasoningModelsDetectionMonth{},456\KnownReasoningModelsLeadMonths{} months before the peak-growth month of457\KnownReasoningModelsPeakGrowthMonth{}; the long lead is an artefact of a458regular expression that includes chain-of-thought and therefore matched the4592023 prompting literature, and the honest reading is that the concept's460vocabulary existed well before its models. % [E-17]461462Retrieval augmentation is the most widely adopted of the window's patterns.463The dense passage retriever \citep{karpukhin2020dpr} and the464retrieval-augmented generation architecture \citep{lewis2020rag} precede the465window; their adoption inside it is surveyed by \citet{gao2023ragsurvey}, and466the known-set detector counts \KnownRetrievalAugmentedGenerationTotalWorks{}467matching works, detected in \KnownRetrievalAugmentedGenerationDetectionMonth{}468with peak growth in \KnownRetrievalAugmentedGenerationPeakGrowthMonth{}.469Tool use---ReAct's interleaving of reasoning traces and actions470\citep{yao2023react}, Toolformer's self-supervised API calls471\citep{schick2023toolformer}, ToolLLM's instruction tuning over thousands of472real APIs \citep{qin2023toolllm}, surveyed by \citet{mialon2023augmented}---is473the bridge from language models to the agents of Section~\ref{sec:ai-agents};474all four are preprints. Hallucination, the generation of fluent content475unsupported by the input or by fact, is the failure mode that these476mechanisms are meant to contain, and it remains a measurement problem before477it is a modelling problem: the surveys \citep{ji2023hallucination,478huang2025hallucination} agree on taxonomies and disagree on metrics. The479clearest statement of why is in the landmark set. The Med-PaLM paper480\citep{singhal2023clinical} built MultiMedQA from six medical481question-answering datasets and a new set of consumer questions, and its482abstract reports that a human evaluation framework along axes of factuality,483comprehension, reasoning and harm was needed because automated evaluation484against limited benchmarks did not meet the bar for clinical use. % [E-27]485486Code generation is the application in which the window's models were first487measured against professional work. The Codex evaluation488\citep{chen2021codex}, AlphaCode's competition-level results489\citep{li2022alphacode}, the open code models \citep{nijkamp2023codegen,490roziere2023codellama, li2023starcoder} and the SWE-bench task of resolving491real repository issues \citep{jimenez2024swebench} define the benchmark492sequence; the controlled productivity experiment of \citet{peng2023copilot}493and the field data of \citet{ziegler2022copilot} are the usage evidence, with494the usability and security caveats of \citet{vaithilingam2022expectation},495\citet{barke2023groundedcopilot} and \citet{pearce2022asleep}. The detector's496lead for \emph{AI code generation} is \KnownAiCodeGenerationLeadMonths{}497months. The software-engineering reading of this literature belongs to498Part~VII; here it is noted only that the LLM category's references to software499engineering in 2025 (Table~\ref{tab:ai_flows}) are the trace of this500application in the citation graph.501502\section{Multimodal and generative AI}503\label{sec:ai-multimodal}504505\subsection{Vision--language models}506507Two baseline-period results make the multimodal literature of the window508possible: contrastive image--text pre-training, which aligns an image encoder509and a text encoder on web-scale pairs \citep{radford2021clip}, and the vision510Transformer, which made the image encoder the same object as the text model511\citep{dosovitskiy2021vit}. The single-stream alternative, in which words and512image regions are tokens of one Transformer, is represented in the landmark513set by VL-BERT \citep{su2019vlbert}; the taxonomy of representation,514translation, alignment, fusion and co-learning that the field still uses is515that of \citet{baltrusaitis2019multimodal}, also in the set. Inside the window516the dominant design is a frozen or lightly adapted language model fed by a517visual encoder through a small bridge: a perceiver resampler in Flamingo518\citep{alayrac2022flamingo}; a querying Transformer trained in two stages519between frozen encoders and frozen LLMs in BLIP-2, whose abstract motivates520the design by the prohibitive cost of end-to-end pre-training521\citep{li2023blip2}; instruction tuning of the same bridge in InstructBLIP522\citep{dai2023instructblip}; a linear projection in LLaVA523\citep{liu2023llava}; shared functional modules with the language decoder as524``universal interface'' in mPLUG-Owl2 \citep{ye2024mplugowl2}. The promptable525segmentation model \citep{kirillov2023sam} is the dense-prediction analogue,526and the open image--text corpus LAION-5B \citep{schuhmann2022laion5b} is what527made CLIP-scale training reproducible outside industry. % [E-28]528Of these records, BLIP-2, InstructBLIP, LLaVA and LAION-5B are preprint or529proceedings deposits; mPLUG-Owl2 is a CVPR paper. The surveys530\citep{yin2024mllmsurvey, zhang2024mmllms} organise the design space; the531corpus-level trace is the vision-language detector, which fires532\KnownVisionLanguageModelsLeadMonths{} months before the concept's peak growth.533534Whether these models \emph{fuse} modalities or merely \emph{condition} on the535text channel is the open question that the landmark abstracts raise most536sharply. An evaluation of eight proprietary and open multimodal foundation537models on 1,090 medical cases reports, in its abstract, that image predictions538are ``largely driven by text'', that accuracy rises monotonically with the539amount of informative text, and that even mild textual suggestions of an540incorrect diagnosis degrade image-based classification541\citep{buckley2026textexploit}. % [E-30]542The scoping review of 432 medical multimodal papers \citep{schouten2025multimodalmedicine}543and the survey of fusion under noisy, missing and imbalanced modalities544\citep{zhang2026lowqualityfusion} describe the same robustness gap from the545methods side. The medical records in the set---generalist biomedical and546radiology models \citep{zhang2024biomedgpt, wu2025radfm}, reinforcement-learning547fine-tuning for medical reasoning in vision-language models and a zero-shot548evaluation of GPT-5 on medical visual question answering, the latter a549preprint \citep{lai2026medr1, wang2025gpt5medical}---report benchmark gains550but not modality ablations; the question is carried to Part~X.551552\subsection{Diffusion models and their applications}553554The generative lineage of the window is diffusion. Denoising diffusion555probabilistic models \citep{ho2020ddpm} and their latent-space formulation556\citep{rombach2022ldm} precede the boundary; Imagen \citep{saharia2022imagen}557and DALL-E~2 \citep{ramesh2022dalle2} are the 2022 text-to-image systems,558and the Imagen record is in the landmark set without an abstract. Three559technical contributions in the set define how the models are controlled.560Classifier-free guidance \citep{ho2022cfg}, a preprint, trains the conditional561and unconditional score jointly and interpolates between them at sampling562time, removing the separate classifier that earlier guidance required;563textual inversion \citep{gal2022textualinversion}, also a preprint, learns a564new token embedding from three to five images of a concept so that a frozen565model can render it; T2I-Adapter \citep{mou2024t2iadapter}, an AAAI paper,566trains small adapters that align external structural or colour signals with567the internal knowledge of a frozen text-to-image model. The Transformer568backbone for diffusion \citep{peebles2023dit} and its rectified-flow569successor \citep{esser2024sd3}, a preprint, close the architectural loop by570replacing the U-Net. The detector flags \emph{diffusion models}571\KnownDiffusionModelsLeadMonths{} months before peak growth. Outside images,572the one peer-reviewed scientific application in the landmark set is573MatterGen \citep{zeni2025mattergen}, whose abstract reports a diffusion model574that generates stable, diverse inorganic materials across the periodic table575and can be fine-tuned towards property constraints. % [E-29]576Speech followed the same self-supervised path a few years earlier577\citep{baevski2020wav2vec2, hsu2021hubert} and reached general-purpose578recognition with weakly supervised training at scale \citep{radford2023whisper};579the review of \citet{mehrish2023speechreview} covers the category, which the580taxonomy places outside the nine categories of this part.581582\subsection{What the generative-AI category rewards}583584Generative AI is the category whose citation profile changed most inside the585window, from a vision-centred one to a three-way split between vision, LLM586and HCI (Table~\ref{tab:ai_flows}). Its citation-ranked landmarks are,587however, not the diffusion papers just described but education, policy and588adoption studies: a policy framework from surveys of students and staff589\citep{chan2023policyframework}, scoping and systematic reviews of assessment590and academic integrity \citep{xia2024assessmentreview, bittle2025integrity},591analyses of the guidelines of 116 and of 50 United States universities592\citep{mcdonald2025hepolicies}, a paradox-theory essay \citep{lim2023ragnarok},593an OECD outlook, and two economics papers: a preregistered field experiment,594circulated as an SSRN preprint, whose title states that generative AI can harm595learning \citep{bastani2024harmlearning}, and nationally representative596surveys whose abstract reports that 45 per cent of United States adults aged59718--64 used generative AI by late 2024 and that work adoption was faster than598that of the personal computer \citep{bick2026rapidadoption}. % [E-31]599The pattern repeats the one observed for LLMs in Section~\ref{sec:ai-landmarks}:600technical generative-AI papers are counted in the category but, normalised601against their own cohort, rarely reach its citation top within thirty-six602months, whereas the first social-science and education responses to the603technology do. A reader who takes the landmark list as the field's604intellectual map would conclude that generative AI is a topic in education605research; the correct conclusion is that field normalisation inside a606multi-label category measures discourse, and that the measurement is607informative about diffusion rather than about invention.608609\section{AI agents and autonomous systems}610\label{sec:ai-agents}611612\subsection{From tool use to agents}613614An LLM-based agent, in the definition shared by the surveys of the window615\citep{wang2024agentsurvey, xi2023rise, guo2024multiagent,616cheng2026agentsdefinitions}, is a language model wrapped in a loop of617perception, planning, memory and action, in which the model's outputs are618parsed into calls to tools or to other agents and the results are fed back as619context. The architectures that fixed the vocabulary are generative agents in620a simulated town \citep{park2023generativeagents}, a peer-reviewed UIST paper;621the multi-agent conversation framework AutoGen \citep{wu2023autogen}; virtual622context management in MemGPT \citep{packer2023memgpt}; and the ``LLM as623operating system'' proposals \citep{mei2024aios, ge2023llmos}, all of the624latter preprints. The memory component is surveyed by625\citet{zhang2025agentmemory}, whose abstract calls self-evolution through626agent--environment interaction the feature that distinguishes agents from the627models they wrap. Evaluation lags design. The benchmarks that the community628uses---SWE-bench for repository-level issue resolution629\citep{jimenez2024swebench}, WebArena for web tasks \citep{zhou2023webarena},630AgentBench across environments \citep{liu2023agentbench}---are preprints, and631the only peer-reviewed agent benchmark in the landmark set is a clinical632one, an evaluation of an open-source and a proprietary agent system on three633families of medical decision tasks \citep{liu2026clinicalagents}. % [E-32]634The detector flags \emph{LLM agents} in \KnownLlmAgentsDetectionMonth{},635\KnownLlmAgentsLeadMonths{} months before peak growth in636\KnownLlmAgentsPeakGrowthMonth{}, on \KnownLlmAgentsTotalWorks{} matching637works; the concept arrived and peaked within a year.638Three features of agent evaluation limit what the benchmarks can show. The639unit of evaluation is a trajectory rather than an answer, so that640non-determinism in the environment and in sampling compounds across steps and641the variance warning of \citet{henderson2018drlmatters} applies with greater642force than in single-turn evaluation; the cost of a run---tokens, tool calls,643wall-clock time---is rarely reported alongside the success rate, so that644rankings conflate capability with budget; and the environments are public and645static, so that the contamination problem of Section~\ref{sec:ai-eval} applies646to the tasks as much as to the models. The clinical benchmark in the landmark647set \citep{liu2026clinicalagents} is notable precisely because its abstract648reports comparing an open-source and a proprietary system under one protocol649across three task families; it is one study.650651The strongest peer-reviewed evidence for agents is scientific. The Virtual Lab652\citep{swanson2025virtuallab}, a Nature paper whose record carries no abstract653in the corpus file, reports agents designing SARS-CoV-2 nanobodies that were654validated experimentally; Robin \citep{ghareeb2026robin} is described in its655abstract as the first multi-agent system automating both hypothesis generation656and data analysis in experimental biology; DIVE \citep{zhang2026dive} extracts657experimental data from the figures of the literature for hydrogen-storage658materials; AgentSociety \citep{piao2026agentsociety} simulates large659populations of LLM-driven generative agents for social science. % [E-33]660These are single demonstrations with experimental or observational661validation, which is more than the ``AI scientist'' preprints662\citep{lu2024aiscientist, si2024novelideas} offer; whether autonomous research663agents produce validated discoveries at a rate that justifies their cost is a664question for preregistered multi-laboratory trials, not for case reports, and665it is listed in Part~X.666667\subsection{``Agentic AI'': a definitional literature}668669The landmark set of the category contains six papers from 2025--2026 whose670purpose is to define agentic AI and distinguish it from AI agents, generative671AI and classical multi-agent systems \citep{acharya2025agenticsurvey,672sapkota2026agentsvsagentic, bandi2025agenticreview, abouali2025agenticparadigms,673allmendinger2026multiagentai, collaco2026agentichealthcare}, most of them674reviews of reviews, one of them a PRISMA-based review of ninety studies that675proposes a symbolic-versus-neural dual-paradigm framework, and one benchmark.676% [E-34]677A literature that produces six taxonomies and one benchmark in eighteen678months is a literature whose object is not yet stabilised, and the question679of whether ``agentic AI'' denotes a distinct class of systems or a relabelling680of tool-augmented LLM pipelines is treated here as open. The trust, risk and681security management framework for multi-agent LLM systems682\citep{raza2026trism} belongs to the same genre: a structure without an683empirical test.684685The corpus series of the category must be read with its composition in mind.686Four of its landmark records are control-theoretic papers on consensus and687formation control of nonlinear multi-agent systems, and its pre-2022 volume688is dominated by classical multi-agent systems, distributed artificial689intelligence and robotics, which explains why the category directs a large690and growing share of its references to robotics (Table~\ref{tab:ai_flows}).691The negative baseline growth rate, the negative level change at the reference692month and the second Bai--Perron break of 2024 (Table~\ref{tab:ai_breaks})693are the signature of a category in which an older literature was flat or694declining while a new one, LLM agents, grew from a small base; the level695effect at the o1 date, \ItsSeriesCatAiAgentsItsOOneLevelChange{}696($p = \ItsSeriesCatAiAgentsItsOOneLevelP$), is where the new literature697becomes visible in the aggregate. % [E-35]698The explanation is inferential; separating the two literatures would require a699sub-classification that the taxonomy does not provide. Embodied agents, from700vision-language-action models \citep{brohan2023rt2, kim2024openvla,701black2024pi0} to the driving model in the landmark set, are classified under702robotics and are reviewed there.703704\section{Reinforcement learning}705\label{sec:ai-rl}706707The reinforcement-learning landmarks are, almost without exception, the canon708of the baseline period. Rainbow's combination of six extensions to deep709Q-learning with an ablation of each \citep{hessel2018rainbow}; the710maximum-entropy actor--critic \citep{haarnoja2018sac}, whose record is the711arXiv deposit; clipped double-Q learning and delayed policy updates against712overestimation \citep{fujimoto2018td3}; self-play without domain knowledge713beyond the rules in AlphaZero \citep{silver2018alphazero}; grandmaster-level714StarCraft~II \citep{vinyals2019alphastar}, a Nature record without an abstract715in the file; the finding that proximal policy optimisation is, in the words of716its abstract, surprisingly strong in cooperative multi-agent benchmarks717\citep{yu2021mappo}; the large-scale study of curiosity-driven learning without718extrinsic rewards \citep{burda2018curiosity}; and the reproducibility warning719of \citet{henderson2018drlmatters}, whose abstract notes that non-determinism720and intrinsic variance make reported results hard to interpret. The theoretical721side is represented by an overview of multi-agent reinforcement learning722\citep{zhang2021marloverview} and a survey of the synthesis of model-predictive723control and reinforcement learning (record 7128436456 of the landmark file).724The single post-2022 method landmark is DreamerV3 \citep{hafner2025dreamerv3},725published in Nature, whose abstract reports a world-model algorithm that726outperforms specialised methods across more than 150 tasks with a single727configuration. % [E-36]728Several records are misdated---a 2008 inverse-reinforcement-learning paper729and a 2007 Bayesian one appear as 2018 and 2025 works---and four are730control-theory papers; the category's citation top is therefore a mixture of731classics, control theory and one new result.732What the one new result represents is the model-based turn. DreamerV3 learns733a latent world model and improves its policy on imagined trajectories, and734its abstract stresses that a single configuration transfers across domains735where earlier algorithms required per-domain tuning; the survey of736model-predictive control and reinforcement learning in the set makes the737complementary point that the two fields derive from the same principles and738differ mainly in whether an environment model is part of the algorithm. Read739together with the rising share of references from the category to robotics,740they describe a field whose centre of gravity moved towards control of741physical systems while its optimisation machinery moved, under another name,742into language-model training.743744The corpus places reinforcement learning among the modestly accelerating745categories: a post-period growth rate only slightly above the baseline746(Table~\ref{tab:ai_growth}), a slope change at the reference month of747\ItsSeriesCatReinforcementLearningItsChatgptSlopeChange{} log points per748month after a negative level change, and a citation profile that moves towards749robotics. % [E-15]750It would be a mistake to read this as the field's eclipse. Reinforcement751learning's most consequential role in the window---reward modelling and policy752optimisation for language models, from PPO through DPO to the group-relative753method of DeepSeek-R1---is invisible in the category because the papers that754carry it are classified as LLM work, and the flow of references from the RL755category to the LLM category stays below one per cent. Any statement that756reinforcement learning declined after 2022 would be an artefact of the757taxonomy, and the correct statement is that its methods migrated into another758category's literature. % [E-37]759760\section{Machine-learning theory, architectures and efficiency}761\label{sec:ai-theory}762763\subsection{What ``machine-learning theory'' contains in the corpus}764765The category named machine-learning theory does not, in its citation-ranked766top, contain a learning-theory result from the window. Its landmarks are767textbooks and textbook chapters, the case for the Matthews correlation768coefficient over accuracy and F$_1$ on imbalanced binary problems769\citep{chicco2020mcc}, the argument that high-stakes decisions should use770inherently interpretable models rather than post-hoc explanations of black771boxes \citep{rudin2019stopexplaining}, the conceptual analysis of772interpretability \citep{lipton2018mythos}, exact Shapley attributions for tree773ensembles \citep{lundberg2020treeexplainer}, the review of stochastic-gradient774optimisation for large-scale learning \citep{bottou2018optimization}, a775transfer-learning survey \citep{zhuang2021transfersurvey}, a reprinted chapter776by Verma and Pearl on the equivalence of causal models, and the misdated777bandit papers of Section~\ref{sec:ai-landmarks}. % [E-38]778The category as the rules define it is general machine learning, and what the779citation ranking rewards within it is pedagogy and measurement convention.780Theoretical work on in-context learning, on the expressivity of Transformers781or on the statistical basis of scaling laws exists inside the window, but it is782classified under LLM or deep-learning architectures, or is cited too little783within thirty-six months to surface. The corpus therefore cannot be used to784say that theory lagged practice in the AI transformation; it can only say that785the theory which is cited quickly is the theory of evaluation and786interpretation.787788\subsection{Architectures: proposals, refutations and the efficiency turn}789790The deep-learning-architectures landmarks are the mirror image: surveys of791graph neural networks \citep{wu2021gnnsurvey, zhou2020gnnreview} and of792Transformers for time series \citep{wen2023tstransformers}, the793long-sequence forecasting Transformer Informer \citep{zhou2021informer}, and794its refutation---the finding, reported in the abstract of795\citet{zeng2023transformerstsf}, that embarrassingly simple linear models796outperform the Transformer-based forecasters on the standard long-term797benchmarks because the permutation-invariant attention discards the temporal798ordering that forecasting depends on. The disagreement is still open within799the field, and the LLM-reprogramming approach to the same task800\citep{jin2023timellm} reopens it from a third side. % [E-39]801Mixup \citep{zhang2017mixup}, the convex-combination regulariser, is in the802set with a wrong year; the tabular foundation model \citep{hollmann2025tabpfn}803is its one peer-reviewed post-2022 architectural result. The two architectural804proposals of the window that the set contains are both unreviewed: Mamba805\citep{gu2023mamba}, whose abstract identifies the inability of earlier806subquadratic models to perform content-based reasoning as their weakness and807makes the state-space parameters input-dependent to obtain linear-time808sequence modelling, and Kolmogorov--Arnold networks \citep{liu2024kan}, which809replace fixed node activations by learnable spline functions on edges and810claim accuracy and interpretability advantages over multilayer perceptrons on811small scientific problems. Neither claim has been tested at language-model812scale in a peer-reviewed venue within the corpus.813814Where architecture research demonstrably moved is efficiency. Exact attention815with IO-aware tiling \citep{dao2022flashattention}; sparse mixture of experts816at trillion-parameter scale \citep{fedus2022switch}; eight-bit matrix817multiplication with outlier decomposition \citep{dettmers2022int8} and818post-training weight quantisation to three or four bits \citep{frantar2023gptq},819with the quantisation survey in the landmark set \citep{gholami2022quantization};820speculative decoding, which samples from a small draft model and verifies in821parallel with the large one \citep{leviathan2023speculative}; and the822efficient-LLM and inference surveys \citep{wan2023efficientllm,823zhou2024inferencesurvey}. The known-set detector gives these topics the824shortest leads in the area---\KnownStateSpaceSequenceModelsLeadMonths{} months825for state-space models, \KnownMixtureOfExpertsLeadMonths{} for mixture of826experts, \KnownSpeculativeDecodingLeadMonths{} for speculative827decoding---because each was driven by one widely adopted preprint and grew828from near zero to its peak within a year. % [E-40]829Deep-learning architectures is also the one AI category without a significant830post-2022 slope change (Table~\ref{tab:ai_breaks}). Two readings are831compatible with the counts: that architecture research plateaued once the832Transformer became the default, or that papers which would have been833classified as architecture work in 2021 are classified as LLM or834foundation-model work in 2024 because their vocabulary changed. The835multi-label counts cannot separate relabelling from decline, and the test836proposed in Part~X---re-estimating the series on works whose only AI category837is architectures, with a classifier vocabulary frozen at 2022---has not been838run.839840\subsection{Interpretability and the science of model internals}841842Between the theory that the citation ranking rewards and the architectures843that practice adopted lies a literature that the corpus barely registers: the844attempt to explain what a trained language model computes. The research845agenda of \citet{anwar2024foundational}, a preprint, lists it among the846foundational challenges of alignment, and the known-emergence set includes847\emph{mechanistic interpretability} as one of its eleven AI concepts. The848detector's result for it is the only negative lead in the set:849\KnownMechanisticInterpretabilityTotalWorks{} matching works, a peak-growth850month that precedes the detection month, and a lead of851\KnownMechanisticInterpretabilityLeadMonths{} months. % [E-17]852Two readings are possible and both are instructive. On the one hand the853concept is small---two to three orders of magnitude fewer matching works than854language models or retrieval augmentation---so that the growth statistic on855which the detector relies is noisy and a single busy quarter sets the peak.856On the other hand the vocabulary of the field changed inside the window:857``circuits'', ``features'' and ``sparse autoencoders'' replaced ``probing''858and ``attribution'', and a regular expression written in 2026 looks for the859later vocabulary in the earlier literature. The interpretability landmarks860that the ranking does surface are the pre-LLM ones---the case for inherently861interpretable models \citep{rudin2019stopexplaining}, the conceptual critique862\citep{lipton2018mythos}, the social-science account of explanation863\citep{miller2019explanation}, the explainable-AI surveys864\citep{barredoarrieta2020xai, ali2023xaitrust}---whose question is whether a865model's decision can be explained to a person, not whether its computation866can be reverse-engineered. The corpus therefore records the demand for867explanation and is nearly silent on the supply of mechanistic understanding;868the manuscript cannot say whether that silence reflects the small size of the869sub-field, its concentration in non-indexed venues and company reports, or870the vocabulary drift just described, and it flags the measurement as871unreliable rather than the sub-field as absent.872873\section{Evaluation}874\label{sec:ai-eval}875876The evaluation literature of the window is a cycle of benchmark creation,877saturation and replacement that the landmark set documents from its first878turn. The SuperGLUE abstract \citep{wang2019superglue} records, already in8792019, that performance on GLUE \citep{wang2018glue} had ``surpassed the level880of non-expert humans'' and that a harder suite was needed; the same sentence881could be written about MMLU \citep{hendrycks2021mmlu} and BIG-bench882\citep{srivastava2022bigbench} by 2024. % [E-41]883Three responses coexist. Holistic evaluation across scenarios and metrics884\citep{liang2022helm} multiplies the axes; crowd-sourced pairwise human885preference in an arena \citep{chiang2024arena}, a preprint, replaces fixed886test sets by an Elo-style ranking; and LLM-as-judge \citep{zheng2023judge},887also a preprint, replaces the human by a model, with the agreement rates888reported there as its justification. The QLoRA abstract's ``99.3 per cent of889ChatGPT'' is an instance of the third: a number produced by GPT-4 judging890Vicuna-benchmark answers, whose meaning depends on the judge's biases, which891the LLM-as-judge paper itself catalogues. Against all three stand the892documented problems of contamination of test sets by training data893\citep{sainz2023contamination}, the dependence of rankings on the choice of894tasks \citep{dehghani2021lottery}, benchmark saturation and the895sociology of benchmark use \citep{ott2022benchmarks, raji2021everything}, the896case for dynamic, adversarially collected benchmarks \citep{kiela2021dynabench}897and the earlier demonstration that even a faithfully re-collected test set898shifts accuracies \citep{recht2019imagenet}. The evaluation survey of899\citet{chang2024evalsurvey} maps the space.900901The corpus adds two facts. The landmark set contains four benchmark902contributions in 234---SuperGLUE, LAION-5B, MultiMedQA and the clinical agent903benchmark---so that evaluation work is under-rewarded by thirty-six-month904citation impact relative to the role the field assigns it, a known property905of citation measures for infrastructure. % [E-05]906And no landmark abstract reports a contamination-controlled evaluation:907the models that the abstracts compare are compared on benchmarks whose908overlap with their training data is not stated. For a mapping study this is909the central methodological caveat of the whole area. Every ``state of the art''910claim read in an abstract of this part is a claim about a benchmark whose911validity for the model in question is unknown, and the manuscript's912distinction between genuine breakthroughs and incremental benchmark913improvements (RQ7) cannot be drawn from abstracts alone; Part~IX draws it, to914the extent that it can be drawn, from the citation and adoption structure915instead.916917\subsection{Openness and the reproducibility of claims}918919RQ9 asks how open-source availability and evaluation standards evolved, and920the landmark abstracts speak to the first half directly. The LLaMA abstract921ends with the sentence that the models are released to the research922community; LAION-5B's purpose, by its title, is an open dataset for training923image--text models; Mamba, KAN, QLoRA, LLaMA-Adapter and the open code models924\citep{touvron2023llama, schuhmann2022laion5b, gu2023mamba, liu2024kan,925dettmers2023qlora, zhang2023llamaadapter, roziere2023codellama,926li2023starcoder} are weights-and-code releases whose adoption is measurable927in the short detector leads of Section~\ref{sec:ai-theory}; and the928DeepSeek-R1 paper \citep{deepseek2025r1nature} is the first peer-reviewed929frontier-class reasoning model with released weights. Against this stands930the closed lineage \citep{openai2023gpt4, gemini2023, openai2024o1}, whose931reports omit data, scale and training details and whose evaluations can932therefore not be reproduced by anyone but the vendor, and the arena and933judge-based evaluations of Section~\ref{sec:ai-eval}, which rank closed models934on prompts that are themselves not released. The AI Index reports935\citep{maslej2024aiindex, maslej2025aiindex} document, as secondary sources,936the growing share of notable models released by industry; the corpus937measures the same thing at the level of affiliations in Part~VIII\@.938Three consequences follow for the status of the claims read in this part.939First, every comparison between an open model and a closed one in an abstract940(QLoRA against ChatGPT, LLaMA against GPT-3) is a comparison between a941reproducible artefact and a moving target, and its date is part of its942meaning. Second, the reproducibility of a result in this area depends less on943whether the paper was peer reviewed than on whether the weights, data and944evaluation prompts were released; several peer-reviewed landmarks of945Section~\ref{sec:ai-fm} train on institutional clinical data that cannot be946shared, while several preprints release everything. Third, the947distinction between peer-reviewed and preprint evidence that this manuscript948maintains throughout is therefore necessary but not sufficient, and Part~IX's949openness indicators, which are computed on the whole corpus rather than on950the landmark set, are the appropriate measure of the trend.951952\section{Safety, alignment, fairness and societal evidence}953\label{sec:ai-safety}954955\subsection{Risk framing}956957The safety literature entered the window with a concrete-problems agenda958\citep{amodei2016concrete, hendrycks2021unsolved} and a critique of scale959\citep{bender2021parrots}, and left it with a catalogue. The AI Risk960Repository, published in \emph{Patterns} and in the landmark set, is961described in its abstract as a living database of 1,725 risks extracted from96274 taxonomies and organised by a causal taxonomy of origin and a domain963taxonomy \citep{slattery2026riskrepository}; the International AI Safety964Report \citep{bengio2025safetyreport} and the \emph{Science} perspective on965managing extreme risks \citep{bengio2024extremerisks} are the consensus966documents, and the alignment survey of \citet{ji2023alignment} and the967foundational-challenges agenda of \citet{anwar2024foundational}, both968preprints, the research agendas. % [E-42]969Catastrophic-risk framing \citep{hendrycks2023catastrophic,970shevlane2023extremerisks}, both preprints, and the ``stochastic parrots''971critique disagree about whether scale produces dangerous capabilities at all;972the corpus shows the AI-safety category re-orienting towards LLM work, which973it cited at a negligible rate in 2019 and at a tenth of its references in9742025 (Table~\ref{tab:ai_flows}), but no landmark abstract reports a validated975measurement of a catastrophic capability, so the disagreement is recorded, not976resolved. % [E-14]977The sociology of the category's citation top is the same as that of ML978theory: it is the pre-LLM ``responsible AI'' literature---fairness979\citep{mehrabi2021biasfairness}, explainable AI \citep{barredoarrieta2020xai,980ali2023xaitrust}, explanation as the social sciences understand it981\citep{miller2019explanation}---rather than the alignment literature of the982window, which is young and mostly unreviewed.983984\subsection{Empirical harms with peer review}985986Where the landmark set is strongest is in peer-reviewed empirical harm. The987data-poisoning threat assessment in \emph{Nature Medicine} reports, in its988abstract, that replacing 0.001 per cent of the tokens of a common pretraining989corpus with medical misinformation produces models more likely to propagate990medical errors \citep{alber2025poisoning}. A physician-led red-teaming of four991public chatbots on 222 patient-posed primary-care questions found992statistically significant differences in the safety of their answers993\citep{draelos2026unsafeanswers}. Experiments on current models as994replacement therapists report expressions of stigma and inappropriate995responses contrary to clinical guidelines \citep{moore2025stigma}. % [E-43]996These three, together with the jailbreak results of Section~\ref{sec:ai-align},997are what the corpus can offer as established, replicable harm; the998over-reliance review \citep{zhai2024overreliance}, the healthcare bias999review \citep{hasanzadeh2025biashealthcare} and the governance frameworks1000\citep{papagiannidis2025raigovernance} are syntheses and proposals. Energy is1001the one harm with a measurement standard: the inference-cost comparison in1002the landmark set reports, in its abstract, the first systematic comparison of1003task-specific and general-purpose models and finds the latter more expensive1004by a wide margin \citep{luccioni2024powerhungry}, in line with the1005electricity-footprint estimates of \citet{devries2023energy}. % [E-44]10061007\subsection{Labour, adoption and education}10081009The social-science records of the set carry the window's evidence on who uses1010the models and to what effect. Occupational exposure to language models is1011estimated from task decompositions in \citet{eloundou2024gpts}, published in1012\emph{Science}, whose landmark record, as noted, carries the wrong abstract;1013the preregistered field experiment with 758 consultants, whose1014\emph{Organization Science} version is in the set, reports that assistance1015improved performance on some tasks and worsened it on others of similar1016apparent difficulty, the ``jagged frontier'' \citep{dellacqua2023jagged}; and1017the adoption surveys report use by nearly half of United States adults within1018two years of the first product \citep{bick2026rapidadoption}. % [E-45]1019In education, the category where the ChatGPT commentary concentrated, the1020abstracts move from perception surveys \citep{chan2023students,1021kasneci2023chatgptgood} to policy analysis \citep{mcdonald2025hepolicies,1022chan2023policyframework} and to one preregistered experiment reporting harm1023to unassisted performance \citep{bastani2024harmlearning}, a preprint. The1024direction of the effect of generative-AI use on learning is thus contested1025between survey evidence of perceived benefit and experimental evidence of1026harm, and the manuscript lists it among the open problems rather than taking1027a side.10281029\section{Synthesis: established, contested, speculative}1030\label{sec:ai-synthesis}10311032Four reconciliations summarise what this part can say and what it cannot.10331034\emph{Growth versus discontinuity.} The nine AI categories accelerated after10352022; the acceleration is a slope change that the data-driven procedures date1036to the middle of 2022 or later and that steepened in 2023--2024, not a1037level jump in December 2022 (Tables~\ref{tab:ai_breaks}1038and~\ref{tab:ai_altdates}). The design cannot attribute it to one product, and1039sentences of the form ``ChatGPT caused a surge in AI research'' are not1040supported by anything in this manuscript. What is supported is weaker and1041more useful: the trend of the LLM literature changed, the change propagated1042through citation to generative AI, safety and agents within three years, and1043it did not propagate to reinforcement learning, theory or architectures.1044% [E-09--E-16]10451046\emph{Impact versus invention.} The citation-ranked landmarks of the LLM and1047generative-AI categories are commentary and education papers; the technical1048lineage is reached through arXiv records that the bibliographic linkage1049sometimes attached to the wrong work. A list of the most influential AI papers1050of the window built from this corpus without the screening of1051Section~\ref{sec:ai-landmarks} would be wrong in about one record in five, and1052would rank the first wave of ChatGPT commentary above the papers that1053introduced the models. % [E-02, E-06]10541055\emph{Taxonomy versus field.} The apparent stagnation of reinforcement1056learning and of deep-learning architectures is at least partly a consequence1057of classification: the methods of the first migrated into LLM post-training,1058and the vocabulary of the second migrated into the LLM and foundation-model1059categories. % [E-37, P-08]10601061\emph{Evidence tiers.} Twenty-eight of the 234 landmark candidates are1062preprints and seventy-two are dated 2026 with twelve-month citation1063windows;\footnote{Counts of the columns \texttt{type} and \texttt{year} of1064\texttt{literature\_synthesis/landmarks/ai\_ml\_landmarks.csv}, reproduced row by1065row in Table~\ref{tab:ai_landmarks}; no macro exists for them.}1066many of the window's defining technical papers---LLaMA, BLIP-2, Mamba, the1067agent benchmarks, the emergent-abilities debate, the catastrophic-risk1068agenda---have not passed peer review. The established results of this part1069are therefore of two kinds: corpus statistics, which are reproducible from1070the files named in the tables, and peer-reviewed empirical findings, of which1071the landmark set contains a modest number, concentrated in medicine1072(foundation models, poisoning, unsafe answers), science (MatterGen, the1073Virtual Lab, Robin, DreamerV3) and the social sciences (adoption, the jagged1074frontier). The contested results---emergence, sparks, pure-RL reasoning,1075linear versus Transformer forecasting, fusion versus text-conditioning,1076agentic AI as a category, the sign of the learning effect---are contested in1077the literature itself. The speculative ones---domain foundation models1078displacing pipelines, autonomous science at scale, trust frameworks for1079multi-agent systems---are speculative because no landmark abstract tests1080them.10811082The open problems that this reading leaves are collected in1083\texttt{literature\_synthesis/agent\_E/open\_problems.csv} and taken up in1084Part~X: whether capabilities are discontinuous in scale; whether pure-RL1085reasoning generalises beyond verifiable rewards; whether alignment brittleness1086is intrinsic; how to measure hallucination without references; whether1087vision-language models fuse or condition; what ``agentic'' denotes; whether1088autonomous science agents pay; whether architecture research plateaued or was1089relabelled; how to evaluate on benchmarks that may be in the training data;1090whether catastrophic capabilities can be measured before deployment; whether1091domain foundation models generalise across sites; whether Transformers suit1092long-horizon time series; whether generative-AI use helps or harms learning;1093and, for this manuscript's own method, whether citation-based landmark1094identification can be made reliable for a field whose most-cited identifiers1095are the ones most often attached to the wrong record.10961097\subsection*{Supplementary table: the screened landmark candidates}1098\label{sec:ai-supp}10991100\input{tables/tab_ai_landmarks}1101