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};48\texttt{literature\_synthesis/agent\_E/contributions.csv}).\footnote{The49flags are a human judgement by the specialist agent from identifiers, citation50counts, venue, title and abstract; they were not produced by a script and51must be audited (Part~III, Phase~8). The counts in this paragraph are those of52the flag file, reproduced in Table~\ref{tab:ai_landmark_flags}.} Thirteen are53\emph{identifier mismatches}: the title and abstract belong to an obscure work54while the arXiv identifier and the 36-month citation count belong to a famous55one. A 2026 article on translating natural language into MongoDB queries56carries the arXiv identifier of the retrieval-augmented-generation paper57\citep{lewis2020rag} and its citations; a Portuguese-language essay on58copyright carries that of GPT-3 \citep{brown2020gpt3}; an essay on59``theatrical compliance'' carries that of chain-of-thought prompting60\citep{wei2022cot}; a preprint on an intensive-care simulation carries that of61Llama~2 \citep{touvron2023llama2}; a ``reproducible experiment runner'' for the62Adam optimiser carries more than forty thousand citations. Three records are63\emph{abstract mismatches}, in which a correct title (the GPT-4 technical64report, \citealp{openai2023gpt4}; ``GPTs are GPTs'',65\citealp{eloundou2024gpts}) is paired with the abstract of another work.66Eleven are \emph{implausible years}: mixup, the variational recurrent network67and several 2013--2019 bandit papers appear as 2024--2026 works, four of them68with April--June 2026 arXiv identifiers, so that their field-normalised scores69are computed against the wrong cohort. Seven are \emph{normalisation70artefacts}, dissertations and unindexed articles with two to twenty citations71and NCS36 above five hundred, because the reference mean for a rare document72type in a young cohort is close to zero. Ten are \emph{off scope}:73control-theoretic consensus papers assigned to AI agents, an essay on74educational-technology ``solutionism'' assigned to reinforcement learning.75% [E-02, E-03]7677The defect is not uniformly distributed. Identifier and abstract problems78concentrate in the large-language-model category, where eight of79thirty-three records are affected, and misdated classics concentrate in80machine-learning theory and deep-learning architectures. The mechanism lies81upstream of this review, in the merging of work records by the source82database and in the second-pass arXiv and title linkage of Phase~183(Section~\ref{sec:m-linkage}); it is reported here because the AI area is where84it bites hardest, the arXiv identifiers of the most-cited papers of the decade85being exactly the ones most likely to be attached to the wrong record. Two86consequences are drawn. First, the clean count of citation-ranked AI landmarks87is at most 190, and every ``most influential papers'' statement of Part~IX88that concerns AI must be computed on records whose DOI, arXiv identifier and89title are mutually consistent. Second, the field-normalised score should be90recomputed with document-type-specific reference sets and with the 202691cohort excluded; this is a recommendation to the director, not something done92here. % [E-04]9394What the set does contain is instructive in its own right95(Table~\ref{tab:ai_contribution_types}). Judged from the abstract for 19096records and from title, venue and type for the 44 without one, 84 of the 23497candidates are surveys, 67 methods, 33 empirical studies, 29 theory or position98papers, 17 engineering systems and only 4 benchmarks. % [E-05]99Within the LLM and generative-AI categories the citation-ranked top is100occupied by education and commentary papers written in the months after101November 2022 \citep{kasneci2023chatgptgood, lo2023rapidreview,102rudolph2023assessment, chan2023students, baidooanu2023chatgptedu,103dwivedi2023sowhat, vandis2023priorities, lim2023ragnarok}, not by the papers104that introduced the models. % [E-06]105This is a property of ranking by field-normalised impact inside a multi-label106category that spans computer science, education and information systems: a107commentary in an education journal is normalised against education108journals, and the first wave of ChatGPT commentary was cited at rates that109no technical paper of its cohort matched within thirty-six months. The110technical lineage---GPT-4, LLaMA, InstructGPT, chain-of-thought, DPO, QLoRA,111BLIP-2, Mamba---is present in the set only where an arXiv record was linked,112and sometimes, as just shown, linked to the wrong record. The synthesis113therefore grounds its technical claims on the verified reference list and uses114the landmark set to characterise what the corpus \emph{rewards} rather than115what the field \emph{invented}. Table~\ref{tab:ai_landmarks} lists all116234 candidates with their flag and contribution type.117118\input{tables/tab_ai_landmark_flags}119\input{tables/tab_ai_contribution_types}120121\subsection{The corpus picture of the nine categories}122\label{sec:ai-corpus}123124Table~\ref{tab:ai_growth} gives the primary-category volume, the compound125annual growth rates of the baseline (2018--2022) and post (2023--2025) periods126and the share of eligible works for each category; Table~\ref{tab:ai_counts}127gives the yearly multi-label counts, which are several times larger because128the AI categories overlap heavily and a work counts in every category it is129assigned to. The ordering is the same under either definition. Large language130models grow fastest in the post period, generative AI, multimodal AI,131foundation models and AI safety follow, reinforcement learning and132machine-learning theory accelerate modestly, deep-learning architectures grow133slightly more slowly than in the baseline, and AI agents reverse a negative134baseline growth rate. % [E-07]135The share of eligible works carrying the LLM label as primary category rises by136more than an order of magnitude between 2018 and 2025. % [E-08]137Two cautions apply. The 2026 column of Table~\ref{tab:ai_counts} covers138publications dated up to the snapshot of 14~October 2026 and is incomplete;139and the categories are defined by rules over topics, arXiv classes, venues and140keyword patterns (Section~\ref{sec:m-classification}), so that a work on141reward modelling for language models is an LLM work and not a reinforcement-learning142work, a point that matters for Section~\ref{sec:ai-rl}.143144\input{tables/tab_ai_growth}145\input{tables/tab_ai_counts}146147The structural-change estimates are in Table~\ref{tab:ai_breaks} and, for the148alternative event dates, Table~\ref{tab:ai_altdates}. For the LLM series the149interrupted time series of eq.~\eqref{eq:its} fitted at150\ItsEventsChatgpt{} gives a level change of151\ItsSeriesCatLlmItsChatgptLevelChange{} log points152(HAC $p = \ItsSeriesCatLlmItsChatgptLevelP$) and a slope change of153\ItsSeriesCatLlmItsChatgptSlopeChange{} log points per month154($p = \ItsSeriesCatLlmItsChatgptSlopeP$) over a pre-period slope of155\ItsSeriesCatLlmItsChatgptPreSlope{}; the placebo distribution of156Section~\ref{sec:m-placebo} places the level statistic at a placebo157$p$-value of \ItsSeriesCatLlmPlaceboChatgptLevelTPlaceboP{} and the slope158statistic at \ItsSeriesCatLlmPlaceboChatgptSlopeTPlaceboP{}; the159Bai--Perron procedure selects \ItsSeriesCatLlmBaiPerronMBic{} break, dated160in Table~\ref{tab:ai_breaks} to the middle of 2022, and the Bayesian161on-line change-point posterior mass between December 2022 and March 2023 is162\ItsSeriesCatLlmBocpdCpProbTwoZeroTwoTwoOneTwoToTwoZeroTwoThreeZeroThree{}.163Excluding three months on either side of the event (the donut specification)164changes the level estimate to165\ItsSeriesCatLlmItsChatgptDonutLevelChange{}166($p = \ItsSeriesCatLlmItsChatgptDonutLevelP$). % [E-09]167The same configuration---no significant level jump at the reference month, a168large and significant slope change, a placebo $p$-value for the level that169does not single out December 2022, a data-driven break before or well after170it---holds for multimodal AI (slope change171\ItsSeriesCatMultimodalAiItsChatgptSlopeChange) and generative AI172(\ItsSeriesCatGenerativeAiItsChatgptSlopeChange). AI agents (level173\ItsSeriesCatAiAgentsItsChatgptLevelChange, $p =174\ItsSeriesCatAiAgentsItsChatgptLevelP$; slope175\ItsSeriesCatAiAgentsItsChatgptSlopeChange), reinforcement learning (level176\ItsSeriesCatReinforcementLearningItsChatgptLevelChange, $p =177\ItsSeriesCatReinforcementLearningItsChatgptLevelP$) and AI safety (level178\ItsSeriesCatAiSafetyItsChatgptLevelChange, $p =179\ItsSeriesCatAiSafetyItsChatgptLevelP$) show \emph{negative} level changes180followed by positive slope changes, a dip and a subsequent acceleration whose181reading is taken up category by category below. Deep-learning architectures is182the only AI category whose slope change183(\ItsSeriesCatDeepLearningArchitecturesItsChatgptSlopeChange, $p =184\ItsSeriesCatDeepLearningArchitecturesItsChatgptSlopeP$) is not185distinguishable from zero. % [E-10]186187\input{tables/tab_ai_breaks}188\input{tables/tab_ai_altdates}189190The alternative dates of Section~\ref{sec:m-altdates} order themselves191monotonically. Level effects estimated at the GPT-3 (\ItsEventsGptThree) and192DALL-E~2 (\ItsEventsDalleTwo) dates are negative for every AI category,193whereas those at the GPT-4 (\ItsEventsGptFour) and o1 (\ItsEventsOOne) dates194are positive and significant for LLM, multimodal and generative AI: for the195LLM series, \ItsSeriesCatLlmItsGptThreeLevelChange{} at the GPT-3 date196($p = \ItsSeriesCatLlmItsGptThreeLevelP$), \ItsSeriesCatLlmItsGptFourLevelChange{}197at the GPT-4 date ($p = \ItsSeriesCatLlmItsGptFourLevelP$) and198\ItsSeriesCatLlmItsOOneLevelChange{} at the o1 date199($p = \ItsSeriesCatLlmItsOOneLevelP$). % [E-11]200Each of these is a separate fit with the event moved, and a positive level201effect at a later date in a convex series is what a steepening trend produces202whether or not anything happened at that date. The reading that the evidence203supports is therefore restricted: the acceleration of the AI literature is a204change of trend that the data-driven procedures place in the middle of 2022205and that steepened through 2023--2024; the monthly series cannot attribute it206to one product release, and this part does not. % [E-12]207208\subsection{Knowledge flows and detected emergence}209\label{sec:ai-flows}210211Table~\ref{tab:ai_flows} reports, for each AI category and for 2019, 2022 and2122025, the destination of its in-corpus references, i.e.\ the row shares213$s_{c \to c'}(t)$ of eq.~\eqref{eq:flow}. The large-language-model category is214the one whose profile changed most. In 2019 and 2022 roughly seven in ten of215its in-corpus references went to natural language processing; in 2025 about a216third do, while more than a quarter go to the LLM category itself and217measurable shares go to human--computer interaction and software engineering.218% [E-13]219Generative AI moved from a vision-centred profile to one split between vision,220LLM and HCI; AI safety, which cited LLM work at a negligible rate in 2019,221directs a tenth of its references there in 2025 and slightly more to HCI.222% [E-14]223Reinforcement learning moves the other way, citing robotics more and LLM work224almost not at all, and machine-learning theory and deep-learning architectures225are essentially unchanged. % [E-15]226At the level of categories, then, the LLM literature has become a citation hub227for generative AI, safety and agents while remaining peripheral to228reinforcement learning, theory and architectures. This is the corpus-level229answer to RQ5 that this part can give; it is consistent with the convergence230thesis of \citet{bommasani2021foundation}, but a change in which category a231field cites is not evidence that its methods converged, and232Section~\ref{sec:ai-synthesis} returns to the distinction. % [E-16]233234\input{tables/tab_ai_flows}235236The preregistered known-emergence set of Section~\ref{sec:m-emerging} contains237eleven AI concepts, and their lead times are the first empirical test of238RQ10 in this area. The detector flags \emph{large language models} in239\KnownLargeLanguageModelsDetectionMonth{},240\KnownLargeLanguageModelsLeadMonths{} months before the peak-growth month241\KnownLargeLanguageModelsPeakGrowthMonth{}, on242\KnownLargeLanguageModelsTotalWorks{} matching works; \emph{reasoning models}243\KnownReasoningModelsLeadMonths{} months ahead; retrieval-augmented generation244\KnownRetrievalAugmentedGenerationLeadMonths{} months; vision-language models245\KnownVisionLanguageModelsLeadMonths{}; diffusion models246\KnownDiffusionModelsLeadMonths{}; LLM agents \KnownLlmAgentsLeadMonths{};247state-space sequence models \KnownStateSpaceSequenceModelsLeadMonths{};248mixture of experts \KnownMixtureOfExpertsLeadMonths{}; speculative decoding249\KnownSpeculativeDecodingLeadMonths{}; AI code generation250\KnownAiCodeGenerationLeadMonths{}; and mechanistic interpretability251\KnownMechanisticInterpretabilityLeadMonths{}, a negative lead in which the252peak-growth month precedes detection on only253\KnownMechanisticInterpretabilityTotalWorks{} matching works. The median lead254over the whole set is \KnownSetMedianLeadMonths{} months. % [E-17]255The long lead for language models is not foresight: the regular expression256matches the pre-2020 literature on ``language models'' and the detector fires257on its slow growth years before the 2023 surge. The short leads for258state-space models, mixture of experts and speculative decoding are the259honest result, and they say that topics driven by a single widely adopted260preprint are detected only months before their growth peaks.261262\section{Large language models and foundation models}263\label{sec:ai-fm}264265\subsection{The established lineage}266267The architecture that dominates the window is the decoder-only Transformer268\citep{vaswani2017attention} trained autoregressively at scale269\citep{radford2019gpt2, brown2020gpt3, chowdhery2022palm}, with the encoder270and encoder--decoder variants of the baseline period \citep{devlin2019bert,271raffel2020t5} surviving mainly as embedding models. Two empirical regularities272from before the window organise the technical literature inside it. The first273is the scaling law, in its original form a power law of loss in parameters $N$274and tokens $D$ \citep{kaplan2020scaling} and in the compute-optimal form of275\citet{hoffmann2022chinchilla}, which, for a fixed training budget $C \approx2766ND$, allocates roughly equal factors of growth to $N$ and $D$ and therefore277implies far more tokens per parameter than the earlier prescription. The278second is the sparse mixture-of-experts layer \citep{shazeer2017moe,279fedus2022switch}, which decouples parameter count from per-token compute. The280open-weight lineage LLaMA, Llama~2 and Llama~3 \citep{touvron2023llama,281touvron2023llama2, dubey2024llama3}---all three preprints, the first of which282is in the landmark set---follows the Chinchilla allocation; the LLaMA abstract283reports that a 13-billion-parameter model trained only on publicly available284data outperforms GPT-3 on most benchmarks, a claim that the abstract makes and285that this review has not verified. % [E-18]286The closed frontier releases are documented by technical reports without287training details \citep{openai2023gpt4, gemini2023, openai2024o1}; they are288cited here as the dated artefacts they are, and the GPT-4 record in the289landmark set, as noted above, carries someone else's abstract. The LLM surveys290of the window \citep{zhao2023llmsurvey, naveed2023llmoverview,291minaee2024llmsurvey, min2023plmsurvey} map a sub-literature that by 2025 had292its own citation core (Table~\ref{tab:ai_flows}); the journal version of the293first of them is in the landmark set and its abstract organises the field294into pre-training, adaptation, utilisation and evaluation.295296The most visible post-2022 development in the landmark set is not a new297general model but the diffusion of the foundation-model recipe into domains298with their own data. Self-supervised pathology encoders trained on hundreds of299thousands of whole-slide images \citep{chen2024uni, lu2024conch,300wang2024chief, huang2023plip, ding2025titan}, a radiology model trained on301two- and three-dimensional scans \citep{wu2025radfm}, a dermatology model on302two million clinical images \citep{yan2025panderm}, a brain-MRI model303\citep{tak2026brainiac}, an Earth-system model trained on more than a million304hours of geophysical data \citep{bodnar2025aurora}, a tabular model that the305abstract describes as a prior-data fitted network outperforming306gradient-boosted trees on small datasets \citep{hollmann2025tabpfn},307time-series representation models \citep{jin2023timellm, liu2024timesurl} and308a language model fine-tuned on ten million human choices to predict behaviour309across psychological experiments \citep{binz2025centaur} are, with the310exception of the time-series preprint, peer-reviewed journal or conference311papers, and each abstract reports outperforming task-specific baselines on a312battery of benchmarks. % [E-19]313What the abstracts do not report, and what the reviews in the same set list as314unresolved \citep{paschali2025fmradiology, zhang2026datacentric}, is external,315multi-site validation under distribution shift; the claim that domain316foundation models will displace task-specific pipelines in regulated settings317is speculative and is carried to Part~X as such. % [E-24]318319Below the frontier, the practical growth of the open-weight ecosystem320concentrated on adaptation. Low-rank adaptation \citep{hu2021lora}, its321four-bit variant QLoRA, whose abstract reports fine-tuning a 65-billion-parameter322model on a single 48\,GB GPU \citep{dettmers2023qlora}, the zero-initialised323adaption prompts of LLaMA-Adapter \citep{zhang2023llamaadapter} and the324parameter-efficient fine-tuning surveys \citep{han2024peftsurvey} define a325regime in which a research group without frontier compute can specialise a326model in hours. The QLoRA abstract's statement that its best model reaches327``99.3\,\% of the performance level of ChatGPT'' on the Vicuna benchmark is a328judge-based measurement, a point to which Section~\ref{sec:ai-eval} returns.329% [E-20]330331\subsection{Contested claims: emergence, sparks and pure-RL reasoning}332333Three claims about what scaled language models are have been argued largely in334preprints, and the corpus can record the argument but not settle it. The first335is emergence. \citet{wei2022emergent} report abilities that appear336discontinuously as scale increases; \citet{schaeffer2023mirage} show that for337the tasks examined the discontinuities are produced by discontinuous338metrics---exact-match accuracy on multi-token answers---and largely vanish339under continuous or token-level ones. Both are preprints, the disagreement is340about measurement rather than data, and the manuscript presents it as such.341% [E-21]342The second is ``Sparks of Artificial General Intelligence''343\citep{bubeck2023sparks}, which reports qualitative experiments on an early,344unreleased version of GPT-4 and, in the abstract's words, contends that the345model exhibits more general intelligence than previous systems. The record is346among the most cited of the landmark set, it is a preprint, its experiments347are not reproducible because the model version is unavailable, and the348critique that scale does not confer understanding \citep{bender2021parrots}349is in the same set with a comparable citation count. % [E-22]350The third is the reasoning turn: the o1 system card \citep{openai2024o1}351moved computation to inference time, and the DeepSeek-R1 paper, whose352peer-reviewed version is in the landmark set \citep{deepseek2025r1nature}353alongside the preprint \citep{deepseek2025r1}, reports that reasoning ``can be354incentivized through pure reinforcement learning, obviating the need for355human-labelled reasoning trajectories''. The claim is peer reviewed; its356generality beyond domains with verifiable rewards---mathematics, code---is not357tested by anything in the corpus, and it is the first entry of the open358problems in Section~\ref{sec:ai-synthesis}. % [E-23]359360\section{Alignment and post-training}361\label{sec:ai-align}362363The recipe that turned a pretrained language model into an assistant was in364place before the window: a reward model fitted to human preference365comparisons \citep{christiano2017rlhf} and a policy optimised against it with366a proximal objective \citep{schulman2017ppo}, applied to instruction following367in InstructGPT \citep{ouyang2022instructgpt} and extended with helpfulness and368harmlessness data \citep{bai2022hh} and with model-written critiques against a369written constitution \citep{bai2022constitutional}. Inside the window the370method's two moving parts were simplified and questioned. Direct preference371optimisation \citep{rafailov2023dpo} rewrites the RLHF objective so that the372policy is fitted directly to preference pairs through a closed-form373log-ratio loss, removing the separate reward model and the on-policy sampling;374the NeurIPS record is in the landmark set, without an abstract in the file.375The open problems of the reward-model route---reward hacking, preference376mis-specification, evaluator disagreement, the gap between stated and revealed377preferences---are catalogued by \citet{casper2023rlhfproblems}, a preprint378whose taxonomy is nevertheless the standard reference. % [E-25]379The corpus adds a sociological observation to this methodological one.380Reinforcement learning as a category directs less than one per cent of its381in-corpus references to LLM work in 2025 (Table~\ref{tab:ai_flows}), while the382LLM category's own self-citation share grew to more than a quarter. The383``return of reinforcement learning'' through post-training is therefore a384phenomenon of the LLM literature, which absorbed PPO, DPO and the group-relative385policy optimisation of DeepSeek-R1 as its own methods; the classification386separates the two communities even where the mathematics is shared, and387Section~\ref{sec:ai-rl} draws the consequence for any claim about the fate of388reinforcement learning.389390That aligned behaviour is brittle under optimisation pressure is established391by three lines of work. Universal adversarial suffixes found by gradient-guided392search transfer across models \citep{zou2023universal}; the ``jailbroken''393analysis attributes failures to competing objectives and mismatched394generalisation between pretraining and safety training395\citep{wei2023jailbroken}; and PAIR, a peer-reviewed method in the landmark396set, is described in its abstract as an attacker language model that generates397semantic jailbreaks against a target with black-box access, in about twenty398queries \citep{chao2025pair}. % [E-26]399The first two are preprints and the third was published in 2025; together they400show that every public alignment method of the window can be defeated by a401modest adaptive budget. Whether this brittleness is intrinsic to402preference-based alignment or an artefact of current scale and data is the403question that none of the landmark abstracts answers, and it is carried to404Part~X.405406\section{Reasoning, retrieval, tool use and code}407\label{sec:ai-reason}408409Chain-of-thought prompting \citep{wei2022cot, kojima2022zeroshotcot},410tree-structured search over thoughts \citep{yao2023tot} and verbal411self-reflection across episodes \citep{shinn2023reflexion} are the prompting412lineage of the reasoning literature; their evidence is benchmark gains, mostly413on arithmetic and symbolic tasks, and several of the records are preprints.414The reasoning-model turn of 2024--2025 \citep{openai2024o1,415deepseek2025r1nature} internalised the procedure by training the model to416produce long intermediate traces and by spending compute at inference rather417than at training. The detector of Section~\ref{sec:m-emerging} flagged418\emph{reasoning models} in \KnownReasoningModelsDetectionMonth{},419\KnownReasoningModelsLeadMonths{} months before the peak-growth month of420\KnownReasoningModelsPeakGrowthMonth{}; the long lead is an artefact of a421regular expression that includes chain-of-thought and therefore matched the4222023 prompting literature, and the honest reading is that the concept's423vocabulary existed well before its models. % [E-17]424425Retrieval augmentation is the most widely adopted of the window's patterns.426The dense passage retriever \citep{karpukhin2020dpr} and the427retrieval-augmented generation architecture \citep{lewis2020rag} precede the428window; their adoption inside it is surveyed by \citet{gao2023ragsurvey}, and429the known-set detector counts \KnownRetrievalAugmentedGenerationTotalWorks{}430matching works, detected in \KnownRetrievalAugmentedGenerationDetectionMonth{}431with peak growth in \KnownRetrievalAugmentedGenerationPeakGrowthMonth{}.432Tool use---ReAct's interleaving of reasoning traces and actions433\citep{yao2023react}, Toolformer's self-supervised API calls434\citep{schick2023toolformer}, ToolLLM's instruction tuning over thousands of435real APIs \citep{qin2023toolllm}, surveyed by \citet{mialon2023augmented}---is436the bridge from language models to the agents of Section~\ref{sec:ai-agents};437all four are preprints. Hallucination, the generation of fluent content438unsupported by the input or by fact, is the failure mode that these439mechanisms are meant to contain, and it remains a measurement problem before440it is a modelling problem: the surveys \citep{ji2023hallucination,441huang2025hallucination} agree on taxonomies and disagree on metrics. The442clearest statement of why is in the landmark set. The Med-PaLM paper443\citep{singhal2023clinical} built MultiMedQA from six medical444question-answering datasets and a new set of consumer questions, and its445abstract reports that a human evaluation framework along axes of factuality,446comprehension, reasoning and harm was needed because automated evaluation447against limited benchmarks did not meet the bar for clinical use. % [E-27]448449Code generation is the application in which the window's models were first450measured against professional work. The Codex evaluation451\citep{chen2021codex}, AlphaCode's competition-level results452\citep{li2022alphacode}, the open code models \citep{nijkamp2023codegen,453roziere2023codellama, li2023starcoder} and the SWE-bench task of resolving454real repository issues \citep{jimenez2024swebench} define the benchmark455sequence; the controlled productivity experiment of \citet{peng2023copilot}456and the field data of \citet{ziegler2022copilot} are the usage evidence, with457the usability and security caveats of \citet{vaithilingam2022expectation},458\citet{barke2023groundedcopilot} and \citet{pearce2022asleep}. The detector's459lead for \emph{AI code generation} is \KnownAiCodeGenerationLeadMonths{}460months. The software-engineering reading of this literature belongs to461Part~VII; here it is noted only that the LLM category's references to software462engineering in 2025 (Table~\ref{tab:ai_flows}) are the trace of this463application in the citation graph.464465\section{Multimodal and generative AI}466\label{sec:ai-multimodal}467468\subsection{Vision--language models}469470Two baseline-period results make the multimodal literature of the window471possible: contrastive image--text pre-training, which aligns an image encoder472and a text encoder on web-scale pairs \citep{radford2021clip}, and the vision473Transformer, which made the image encoder the same object as the text model474\citep{dosovitskiy2021vit}. The single-stream alternative, in which words and475image regions are tokens of one Transformer, is represented in the landmark476set by VL-BERT \citep{su2019vlbert}; the taxonomy of representation,477translation, alignment, fusion and co-learning that the field still uses is478that of \citet{baltrusaitis2019multimodal}, also in the set. Inside the window479the dominant design is a frozen or lightly adapted language model fed by a480visual encoder through a small bridge: a perceiver resampler in Flamingo481\citep{alayrac2022flamingo}; a querying Transformer trained in two stages482between frozen encoders and frozen LLMs in BLIP-2, whose abstract motivates483the design by the prohibitive cost of end-to-end pre-training484\citep{li2023blip2}; instruction tuning of the same bridge in InstructBLIP485\citep{dai2023instructblip}; a linear projection in LLaVA486\citep{liu2023llava}; shared functional modules with the language decoder as487``universal interface'' in mPLUG-Owl2 \citep{ye2024mplugowl2}. The promptable488segmentation model \citep{kirillov2023sam} is the dense-prediction analogue,489and the open image--text corpus LAION-5B \citep{schuhmann2022laion5b} is what490made CLIP-scale training reproducible outside industry. % [E-28]491Of these records, BLIP-2, InstructBLIP, LLaVA and LAION-5B are preprint or492proceedings deposits; mPLUG-Owl2 is a CVPR paper. The surveys493\citep{yin2024mllmsurvey, zhang2024mmllms} organise the design space; the494corpus-level trace is the vision-language detector, which fires495\KnownVisionLanguageModelsLeadMonths{} months before the concept's peak growth.496497Whether these models \emph{fuse} modalities or merely \emph{condition} on the498text channel is the open question that the landmark abstracts raise most499sharply. An evaluation of eight proprietary and open multimodal foundation500models on 1,090 medical cases reports, in its abstract, that image predictions501are ``largely driven by text'', that accuracy rises monotonically with the502amount of informative text, and that even mild textual suggestions of an503incorrect diagnosis degrade image-based classification504\citep{buckley2026textexploit}. % [E-30]505The scoping review of 432 medical multimodal papers \citep{schouten2025multimodalmedicine}506and the survey of fusion under noisy, missing and imbalanced modalities507\citep{zhang2026lowqualityfusion} describe the same robustness gap from the508methods side. The medical records in the set---generalist biomedical and509radiology models \citep{zhang2024biomedgpt, wu2025radfm}, reinforcement-learning510fine-tuning for medical reasoning in vision-language models and a zero-shot511evaluation of GPT-5 on medical visual question answering, the latter a512preprint \citep{lai2026medr1, wang2025gpt5medical}---report benchmark gains513but not modality ablations; the question is carried to Part~X.514515\subsection{Diffusion models and their applications}516517The generative lineage of the window is diffusion. Denoising diffusion518probabilistic models \citep{ho2020ddpm} and their latent-space formulation519\citep{rombach2022ldm} precede the boundary; Imagen \citep{saharia2022imagen}520and DALL-E~2 \citep{ramesh2022dalle2} are the 2022 text-to-image systems,521and the Imagen record is in the landmark set without an abstract. Three522technical contributions in the set define how the models are controlled.523Classifier-free guidance \citep{ho2022cfg}, a preprint, trains the conditional524and unconditional score jointly and interpolates between them at sampling525time, removing the separate classifier that earlier guidance required;526textual inversion \citep{gal2022textualinversion}, also a preprint, learns a527new token embedding from three to five images of a concept so that a frozen528model can render it; T2I-Adapter \citep{mou2024t2iadapter}, an AAAI paper,529trains small adapters that align external structural or colour signals with530the internal knowledge of a frozen text-to-image model. The Transformer531backbone for diffusion \citep{peebles2023dit} and its rectified-flow532successor \citep{esser2024sd3}, a preprint, close the architectural loop by533replacing the U-Net. The detector flags \emph{diffusion models}534\KnownDiffusionModelsLeadMonths{} months before peak growth. Outside images,535the one peer-reviewed scientific application in the landmark set is536MatterGen \citep{zeni2025mattergen}, whose abstract reports a diffusion model537that generates stable, diverse inorganic materials across the periodic table538and can be fine-tuned towards property constraints. % [E-29]539Speech followed the same self-supervised path a few years earlier540\citep{baevski2020wav2vec2, hsu2021hubert} and reached general-purpose541recognition with weakly supervised training at scale \citep{radford2023whisper};542the review of \citet{mehrish2023speechreview} covers the category, which the543taxonomy places outside the nine categories of this part.544545\subsection{What the generative-AI category rewards}546547Generative AI is the category whose citation profile changed most inside the548window, from a vision-centred one to a three-way split between vision, LLM549and HCI (Table~\ref{tab:ai_flows}). Its citation-ranked landmarks are,550however, not the diffusion papers just described but education, policy and551adoption studies: a policy framework from surveys of students and staff552\citep{chan2023policyframework}, scoping and systematic reviews of assessment553and academic integrity \citep{xia2024assessmentreview, bittle2025integrity},554analyses of the guidelines of 116 and of 50 United States universities555\citep{mcdonald2025hepolicies}, a paradox-theory essay \citep{lim2023ragnarok},556an OECD outlook, and two economics papers: a preregistered field experiment,557circulated as an SSRN preprint, whose title states that generative AI can harm558learning \citep{bastani2024harmlearning}, and nationally representative559surveys whose abstract reports that 45 per cent of United States adults aged56018--64 used generative AI by late 2024 and that work adoption was faster than561that of the personal computer \citep{bick2026rapidadoption}. % [E-31]562The pattern repeats the one observed for LLMs in Section~\ref{sec:ai-landmarks}:563technical generative-AI papers are counted in the category but, normalised564against their own cohort, rarely reach its citation top within thirty-six565months, whereas the first social-science and education responses to the566technology do. A reader who takes the landmark list as the field's567intellectual map would conclude that generative AI is a topic in education568research; the correct conclusion is that field normalisation inside a569multi-label category measures discourse, and that the measurement is570informative about diffusion rather than about invention.571572\section{AI agents and autonomous systems}573\label{sec:ai-agents}574575\subsection{From tool use to agents}576577An LLM-based agent, in the definition shared by the surveys of the window578\citep{wang2024agentsurvey, xi2023rise, guo2024multiagent,579cheng2026agentsdefinitions}, is a language model wrapped in a loop of580perception, planning, memory and action, in which the model's outputs are581parsed into calls to tools or to other agents and the results are fed back as582context. The architectures that fixed the vocabulary are generative agents in583a simulated town \citep{park2023generativeagents}, a peer-reviewed UIST paper;584the multi-agent conversation framework AutoGen \citep{wu2023autogen}; virtual585context management in MemGPT \citep{packer2023memgpt}; and the ``LLM as586operating system'' proposals \citep{mei2024aios, ge2023llmos}, all of the587latter preprints. The memory component is surveyed by588\citet{zhang2025agentmemory}, whose abstract calls self-evolution through589agent--environment interaction the feature that distinguishes agents from the590models they wrap. Evaluation lags design. The benchmarks that the community591uses---SWE-bench for repository-level issue resolution592\citep{jimenez2024swebench}, WebArena for web tasks \citep{zhou2023webarena},593AgentBench across environments \citep{liu2023agentbench}---are preprints, and594the only peer-reviewed agent benchmark in the landmark set is a clinical595one, an evaluation of an open-source and a proprietary agent system on three596families of medical decision tasks \citep{liu2026clinicalagents}. % [E-32]597The detector flags \emph{LLM agents} in \KnownLlmAgentsDetectionMonth{},598\KnownLlmAgentsLeadMonths{} months before peak growth in599\KnownLlmAgentsPeakGrowthMonth{}, on \KnownLlmAgentsTotalWorks{} matching600works; the concept arrived and peaked within a year.601602The strongest peer-reviewed evidence for agents is scientific. The Virtual Lab603\citep{swanson2025virtuallab}, a Nature paper whose record carries no abstract604in the corpus file, reports agents designing SARS-CoV-2 nanobodies that were605validated experimentally; Robin \citep{ghareeb2026robin} is described in its606abstract as the first multi-agent system automating both hypothesis generation607and data analysis in experimental biology; DIVE \citep{zhang2026dive} extracts608experimental data from the figures of the literature for hydrogen-storage609materials; AgentSociety \citep{piao2026agentsociety} simulates large610populations of LLM-driven generative agents for social science. % [E-33]611These are single demonstrations with experimental or observational612validation, which is more than the ``AI scientist'' preprints613\citep{lu2024aiscientist, si2024novelideas} offer; whether autonomous research614agents produce validated discoveries at a rate that justifies their cost is a615question for preregistered multi-laboratory trials, not for case reports, and616it is listed in Part~X.617618\subsection{``Agentic AI'': a definitional literature}619620The landmark set of the category contains six papers from 2025--2026 whose621purpose is to define agentic AI and distinguish it from AI agents, generative622AI and classical multi-agent systems \citep{acharya2025agenticsurvey,623sapkota2026agentsvsagentic, bandi2025agenticreview, abouali2025agenticparadigms,624allmendinger2026multiagentai, collaco2026agentichealthcare}, most of them625reviews of reviews, one of them a PRISMA-based review of ninety studies that626proposes a symbolic-versus-neural dual-paradigm framework, and one benchmark.627% [E-34]628A literature that produces six taxonomies and one benchmark in eighteen629months is a literature whose object is not yet stabilised, and the question630of whether ``agentic AI'' denotes a distinct class of systems or a relabelling631of tool-augmented LLM pipelines is treated here as open. The trust, risk and632security management framework for multi-agent LLM systems633\citep{raza2026trism} belongs to the same genre: a structure without an634empirical test.635636The corpus series of the category must be read with its composition in mind.637Four of its landmark records are control-theoretic papers on consensus and638formation control of nonlinear multi-agent systems, and its pre-2022 volume639is dominated by classical multi-agent systems, distributed artificial640intelligence and robotics, which explains why the category directs a large641and growing share of its references to robotics (Table~\ref{tab:ai_flows}).642The negative baseline growth rate, the negative level change at the reference643month and the second Bai--Perron break of 2024 (Table~\ref{tab:ai_breaks})644are the signature of a category in which an older literature was flat or645declining while a new one, LLM agents, grew from a small base; the level646effect at the o1 date, \ItsSeriesCatAiAgentsItsOOneLevelChange{}647($p = \ItsSeriesCatAiAgentsItsOOneLevelP$), is where the new literature648becomes visible in the aggregate. % [E-35]649The explanation is inferential; separating the two literatures would require a650sub-classification that the taxonomy does not provide. Embodied agents, from651vision-language-action models \citep{brohan2023rt2, kim2024openvla,652black2024pi0} to the driving model in the landmark set, are classified under653robotics and are reviewed there.654655\section{Reinforcement learning}656\label{sec:ai-rl}657658The reinforcement-learning landmarks are, almost without exception, the canon659of the baseline period. Rainbow's combination of six extensions to deep660Q-learning with an ablation of each \citep{hessel2018rainbow}; the661maximum-entropy actor--critic \citep{haarnoja2018sac}, whose record is the662arXiv deposit; clipped double-Q learning and delayed policy updates against663overestimation \citep{fujimoto2018td3}; self-play without domain knowledge664beyond the rules in AlphaZero \citep{silver2018alphazero}; grandmaster-level665StarCraft~II \citep{vinyals2019alphastar}, a Nature record without an abstract666in the file; the finding that proximal policy optimisation is, in the words of667its abstract, surprisingly strong in cooperative multi-agent benchmarks668\citep{yu2021mappo}; the large-scale study of curiosity-driven learning without669extrinsic rewards \citep{burda2018curiosity}; and the reproducibility warning670of \citet{henderson2018drlmatters}, whose abstract notes that non-determinism671and intrinsic variance make reported results hard to interpret. The theoretical672side is represented by an overview of multi-agent reinforcement learning673\citep{zhang2021marloverview} and a survey of the synthesis of model-predictive674control and reinforcement learning (record 7128436456 of the landmark file).675The single post-2022 method landmark is DreamerV3 \citep{hafner2025dreamerv3},676published in Nature, whose abstract reports a world-model algorithm that677outperforms specialised methods across more than 150 tasks with a single678configuration. % [E-36]679Several records are misdated---a 2008 inverse-reinforcement-learning paper680and a 2007 Bayesian one appear as 2018 and 2025 works---and four are681control-theory papers; the category's citation top is therefore a mixture of682classics, control theory and one new result.683684The corpus places reinforcement learning among the modestly accelerating685categories: a post-period growth rate only slightly above the baseline686(Table~\ref{tab:ai_growth}), a slope change at the reference month of687\ItsSeriesCatReinforcementLearningItsChatgptSlopeChange{} log points per688month after a negative level change, and a citation profile that moves towards689robotics. % [E-15]690It would be a mistake to read this as the field's eclipse. Reinforcement691learning's most consequential role in the window---reward modelling and policy692optimisation for language models, from PPO through DPO to the group-relative693method of DeepSeek-R1---is invisible in the category because the papers that694carry it are classified as LLM work, and the flow of references from the RL695category to the LLM category stays below one per cent. Any statement that696reinforcement learning declined after 2022 would be an artefact of the697taxonomy, and the correct statement is that its methods migrated into another698category's literature. % [E-37]699700\section{Machine-learning theory, architectures and efficiency}701\label{sec:ai-theory}702703\subsection{What ``machine-learning theory'' contains in the corpus}704705The category named machine-learning theory does not, in its citation-ranked706top, contain a learning-theory result from the window. Its landmarks are707textbooks and textbook chapters, the case for the Matthews correlation708coefficient over accuracy and F$_1$ on imbalanced binary problems709\citep{chicco2020mcc}, the argument that high-stakes decisions should use710inherently interpretable models rather than post-hoc explanations of black711boxes \citep{rudin2019stopexplaining}, the conceptual analysis of712interpretability \citep{lipton2018mythos}, exact Shapley attributions for tree713ensembles \citep{lundberg2020treeexplainer}, the review of stochastic-gradient714optimisation for large-scale learning \citep{bottou2018optimization}, a715transfer-learning survey \citep{zhuang2021transfersurvey}, a reprinted chapter716by Verma and Pearl on the equivalence of causal models, and the misdated717bandit papers of Section~\ref{sec:ai-landmarks}. % [E-38]718The category as the rules define it is general machine learning, and what the719citation ranking rewards within it is pedagogy and measurement convention.720Theoretical work on in-context learning, on the expressivity of Transformers721or on the statistical basis of scaling laws exists inside the window, but it is722classified under LLM or deep-learning architectures, or is cited too little723within thirty-six months to surface. The corpus therefore cannot be used to724say that theory lagged practice in the AI transformation; it can only say that725the theory which is cited quickly is the theory of evaluation and726interpretation.727728\subsection{Architectures: proposals, refutations and the efficiency turn}729730The deep-learning-architectures landmarks are the mirror image: surveys of731graph neural networks \citep{wu2021gnnsurvey, zhou2020gnnreview} and of732Transformers for time series \citep{wen2023tstransformers}, the733long-sequence forecasting Transformer Informer \citep{zhou2021informer}, and734its refutation---the finding, reported in the abstract of735\citet{zeng2023transformerstsf}, that embarrassingly simple linear models736outperform the Transformer-based forecasters on the standard long-term737benchmarks because the permutation-invariant attention discards the temporal738ordering that forecasting depends on. The disagreement is still open within739the field, and the LLM-reprogramming approach to the same task740\citep{jin2023timellm} reopens it from a third side. % [E-39]741Mixup \citep{zhang2017mixup}, the convex-combination regulariser, is in the742set with a wrong year; the tabular foundation model \citep{hollmann2025tabpfn}743is its one peer-reviewed post-2022 architectural result. The two architectural744proposals of the window that the set contains are both unreviewed: Mamba745\citep{gu2023mamba}, whose abstract identifies the inability of earlier746subquadratic models to perform content-based reasoning as their weakness and747makes the state-space parameters input-dependent to obtain linear-time748sequence modelling, and Kolmogorov--Arnold networks \citep{liu2024kan}, which749replace fixed node activations by learnable spline functions on edges and750claim accuracy and interpretability advantages over multilayer perceptrons on751small scientific problems. Neither claim has been tested at language-model752scale in a peer-reviewed venue within the corpus.753754Where architecture research demonstrably moved is efficiency. Exact attention755with IO-aware tiling \citep{dao2022flashattention}; sparse mixture of experts756at trillion-parameter scale \citep{fedus2022switch}; eight-bit matrix757multiplication with outlier decomposition \citep{dettmers2022int8} and758post-training weight quantisation to three or four bits \citep{frantar2023gptq},759with the quantisation survey in the landmark set \citep{gholami2022quantization};760speculative decoding, which samples from a small draft model and verifies in761parallel with the large one \citep{leviathan2023speculative}; and the762efficient-LLM and inference surveys \citep{wan2023efficientllm,763zhou2024inferencesurvey}. The known-set detector gives these topics the764shortest leads in the area---\KnownStateSpaceSequenceModelsLeadMonths{} months765for state-space models, \KnownMixtureOfExpertsLeadMonths{} for mixture of766experts, \KnownSpeculativeDecodingLeadMonths{} for speculative767decoding---because each was driven by one widely adopted preprint and grew768from near zero to its peak within a year. % [E-40]769Deep-learning architectures is also the one AI category without a significant770post-2022 slope change (Table~\ref{tab:ai_breaks}). Two readings are771compatible with the counts: that architecture research plateaued once the772Transformer became the default, or that papers which would have been773classified as architecture work in 2021 are classified as LLM or774foundation-model work in 2024 because their vocabulary changed. The775multi-label counts cannot separate relabelling from decline, and the test776proposed in Part~X---re-estimating the series on works whose only AI category777is architectures, with a classifier vocabulary frozen at 2022---has not been778run.779780\section{Evaluation}781\label{sec:ai-eval}782783The evaluation literature of the window is a cycle of benchmark creation,784saturation and replacement that the landmark set documents from its first785turn. The SuperGLUE abstract \citep{wang2019superglue} records, already in7862019, that performance on GLUE \citep{wang2018glue} had ``surpassed the level787of non-expert humans'' and that a harder suite was needed; the same sentence788could be written about MMLU \citep{hendrycks2021mmlu} and BIG-bench789\citep{srivastava2022bigbench} by 2024. % [E-41]790Three responses coexist. Holistic evaluation across scenarios and metrics791\citep{liang2022helm} multiplies the axes; crowd-sourced pairwise human792preference in an arena \citep{chiang2024arena}, a preprint, replaces fixed793test sets by an Elo-style ranking; and LLM-as-judge \citep{zheng2023judge},794also a preprint, replaces the human by a model, with the agreement rates795reported there as its justification. The QLoRA abstract's ``99.3 per cent of796ChatGPT'' is an instance of the third: a number produced by GPT-4 judging797Vicuna-benchmark answers, whose meaning depends on the judge's biases, which798the LLM-as-judge paper itself catalogues. Against all three stand the799documented problems of contamination of test sets by training data800\citep{sainz2023contamination}, the dependence of rankings on the choice of801tasks \citep{dehghani2021lottery}, benchmark saturation and the802sociology of benchmark use \citep{ott2022benchmarks, raji2021everything}, the803case for dynamic, adversarially collected benchmarks \citep{kiela2021dynabench}804and the earlier demonstration that even a faithfully re-collected test set805shifts accuracies \citep{recht2019imagenet}. The evaluation survey of806\citet{chang2024evalsurvey} maps the space.807808The corpus adds two facts. The landmark set contains four benchmark809contributions in 234---SuperGLUE, LAION-5B, MultiMedQA and the clinical agent810benchmark---so that evaluation work is under-rewarded by thirty-six-month811citation impact relative to the role the field assigns it, a known property812of citation measures for infrastructure. % [E-05]813And no landmark abstract reports a contamination-controlled evaluation:814the models that the abstracts compare are compared on benchmarks whose815overlap with their training data is not stated. For a mapping study this is816the central methodological caveat of the whole area. Every ``state of the art''817claim read in an abstract of this part is a claim about a benchmark whose818validity for the model in question is unknown, and the manuscript's819distinction between genuine breakthroughs and incremental benchmark820improvements (RQ7) cannot be drawn from abstracts alone; Part~IX draws it, to821the extent that it can be drawn, from the citation and adoption structure822instead.823824\section{Safety, alignment, fairness and societal evidence}825\label{sec:ai-safety}826827\subsection{Risk framing}828829The safety literature entered the window with a concrete-problems agenda830\citep{amodei2016concrete, hendrycks2021unsolved} and a critique of scale831\citep{bender2021parrots}, and left it with a catalogue. The AI Risk832Repository, published in \emph{Patterns} and in the landmark set, is833described in its abstract as a living database of 1,725 risks extracted from83474 taxonomies and organised by a causal taxonomy of origin and a domain835taxonomy \citep{slattery2026riskrepository}; the International AI Safety836Report \citep{bengio2025safetyreport} and the \emph{Science} perspective on837managing extreme risks \citep{bengio2024extremerisks} are the consensus838documents, and the alignment survey of \citet{ji2023alignment} and the839foundational-challenges agenda of \citet{anwar2024foundational}, both840preprints, the research agendas. % [E-42]841Catastrophic-risk framing \citep{hendrycks2023catastrophic,842shevlane2023extremerisks}, both preprints, and the ``stochastic parrots''843critique disagree about whether scale produces dangerous capabilities at all;844the corpus shows the AI-safety category re-orienting towards LLM work, which845it cited at a negligible rate in 2019 and at a tenth of its references in8462025 (Table~\ref{tab:ai_flows}), but no landmark abstract reports a validated847measurement of a catastrophic capability, so the disagreement is recorded, not848resolved. % [E-14]849The sociology of the category's citation top is the same as that of ML850theory: it is the pre-LLM ``responsible AI'' literature---fairness851\citep{mehrabi2021biasfairness}, explainable AI \citep{barredoarrieta2020xai,852ali2023xaitrust}, explanation as the social sciences understand it853\citep{miller2019explanation}---rather than the alignment literature of the854window, which is young and mostly unreviewed.855856\subsection{Empirical harms with peer review}857858Where the landmark set is strongest is in peer-reviewed empirical harm. The859data-poisoning threat assessment in \emph{Nature Medicine} reports, in its860abstract, that replacing 0.001 per cent of the tokens of a common pretraining861corpus with medical misinformation produces models more likely to propagate862medical errors \citep{alber2025poisoning}. A physician-led red-teaming of four863public chatbots on 222 patient-posed primary-care questions found864statistically significant differences in the safety of their answers865\citep{draelos2026unsafeanswers}. Experiments on current models as866replacement therapists report expressions of stigma and inappropriate867responses contrary to clinical guidelines \citep{moore2025stigma}. % [E-43]868These three, together with the jailbreak results of Section~\ref{sec:ai-align},869are what the corpus can offer as established, replicable harm; the870over-reliance review \citep{zhai2024overreliance}, the healthcare bias871review \citep{hasanzadeh2025biashealthcare} and the governance frameworks872\citep{papagiannidis2025raigovernance} are syntheses and proposals. Energy is873the one harm with a measurement standard: the inference-cost comparison in874the landmark set reports, in its abstract, the first systematic comparison of875task-specific and general-purpose models and finds the latter more expensive876by a wide margin \citep{luccioni2024powerhungry}, in line with the877electricity-footprint estimates of \citet{devries2023energy}. % [E-44]878879\subsection{Labour, adoption and education}880881The social-science records of the set carry the window's evidence on who uses882the models and to what effect. Occupational exposure to language models is883estimated from task decompositions in \citet{eloundou2024gpts}, published in884\emph{Science}, whose landmark record, as noted, carries the wrong abstract;885the preregistered field experiment with 758 consultants, whose886\emph{Organization Science} version is in the set, reports that assistance887improved performance on some tasks and worsened it on others of similar888apparent difficulty, the ``jagged frontier'' \citep{dellacqua2023jagged}; and889the adoption surveys report use by nearly half of United States adults within890two years of the first product \citep{bick2026rapidadoption}. % [E-45]891In education, the category where the ChatGPT commentary concentrated, the892abstracts move from perception surveys \citep{chan2023students,893kasneci2023chatgptgood} to policy analysis \citep{mcdonald2025hepolicies,894chan2023policyframework} and to one preregistered experiment reporting harm895to unassisted performance \citep{bastani2024harmlearning}, a preprint. The896direction of the effect of generative-AI use on learning is thus contested897between survey evidence of perceived benefit and experimental evidence of898harm, and the manuscript lists it among the open problems rather than taking899a side.900901\section{Synthesis: established, contested, speculative}902\label{sec:ai-synthesis}903904Four reconciliations summarise what this part can say and what it cannot.905906\emph{Growth versus discontinuity.} The nine AI categories accelerated after9072022; the acceleration is a slope change that the data-driven procedures date908to the middle of 2022 or later and that steepened in 2023--2024, not a909level jump in December 2022 (Tables~\ref{tab:ai_breaks} and910\ref{tab:ai_altdates}). The design cannot attribute it to one product, and911sentences of the form ``ChatGPT caused a surge in AI research'' are not912supported by anything in this manuscript. What is supported is weaker and913more useful: the trend of the LLM literature changed, the change propagated914through citation to generative AI, safety and agents within three years, and915it did not propagate to reinforcement learning, theory or architectures.916% [E-09--E-16]917918\emph{Impact versus invention.} The citation-ranked landmarks of the LLM and919generative-AI categories are commentary and education papers; the technical920lineage is reached through arXiv records that the bibliographic linkage921sometimes attached to the wrong work. A list of the most influential AI papers922of the window built from this corpus without the screening of923Section~\ref{sec:ai-landmarks} would be wrong in about one record in five, and924would rank the first wave of ChatGPT commentary above the papers that925introduced the models. % [E-02, E-06]926927\emph{Taxonomy versus field.} The apparent stagnation of reinforcement928learning and of deep-learning architectures is at least partly a consequence929of classification: the methods of the first migrated into LLM post-training,930and the vocabulary of the second migrated into the LLM and foundation-model931categories. % [E-37, P-08]932933\emph{Evidence tiers.} Twenty-eight of the 234 landmark candidates are934preprints and seventy-two are dated 2026 with twelve-month citation windows;935many of the window's defining technical papers---LLaMA, BLIP-2, Mamba, the936agent benchmarks, the emergent-abilities debate, the catastrophic-risk937agenda---have not passed peer review. The established results of this part938are therefore of two kinds: corpus statistics, which are reproducible from939the files named in the tables, and peer-reviewed empirical findings, of which940the landmark set contains a modest number, concentrated in medicine941(foundation models, poisoning, unsafe answers), science (MatterGen, the942Virtual Lab, Robin, DreamerV3) and the social sciences (adoption, the jagged943frontier). The contested results---emergence, sparks, pure-RL reasoning,944linear versus Transformer forecasting, fusion versus text-conditioning,945agentic AI as a category, the sign of the learning effect---are contested in946the literature itself. The speculative ones---domain foundation models947displacing pipelines, autonomous science at scale, trust frameworks for948multi-agent systems---are speculative because no landmark abstract tests949them.950951The open problems that this reading leaves are collected in952\texttt{literature\_synthesis/agent\_E/open\_problems.csv} and taken up in953Part~X: whether capabilities are discontinuous in scale; whether pure-RL954reasoning generalises beyond verifiable rewards; whether alignment brittleness955is intrinsic; how to measure hallucination without references; whether956vision-language models fuse or condition; what ``agentic'' denotes; whether957autonomous science agents pay; whether architecture research plateaued or was958relabelled; how to evaluate on benchmarks that may be in the training data;959whether catastrophic capabilities can be measured before deployment; whether960domain foundation models generalise across sites; whether Transformers suit961long-horizon time series; whether generative-AI use helps or harms learning;962and, for this manuscript's own method, whether citation-based landmark963identification can be made reliable for a field whose most-cited identifiers964are the ones most often attached to the wrong record.965