% ============================================================================
% Part V — The Artificial Intelligence Transformation
% Written by Agent E (AI Research Specialist), 2026-11-03, from
% literature_synthesis/agent_E/{review.md, claims.csv, contributions.csv,
% open_problems.csv}. Evidence rules: landmark papers are known from metadata
% and abstracts only (never full text); preprints are named as preprints; corpus
% numbers enter only through the macros of tables/numbers.tex or the generated
% tables tab_ai_*.tex (tables/make_tables.py, tables_part05). Claim identifiers
% [E-nn] in comments refer to claims.csv. No causal claim about ChatGPT.
% ============================================================================

\section{Scope, evidence base and reading rules}
\label{sec:ai-scope}

This part synthesises the nine categories of the taxonomy that the ontology
groups under \emph{AI and machine learning}: large language models, foundation
models, multimodal AI, generative AI, AI agents and autonomous systems,
reinforcement learning, machine-learning theory, deep-learning architectures,
and AI safety and alignment. Three kinds of evidence are combined, and the
reader should keep them apart. The first is the corpus itself: the yearly and
monthly counts, growth rates, structural-change estimates and citation flows of
Part~IV, read here only through the generated tables of this part
(Tables~\ref{tab:ai_growth}--\ref{tab:ai_landmarks}) and the macros of
\texttt{tables/numbers.tex}. The second is the automatically selected landmark
set of the area, \SelectionByAreaAiMl{} records chosen by field-normalised
citation impact within each primary category (Part~IV,
Section~\ref{sec:m-score}), of which the agent read the title, venue, type,
identifiers, citation windows and, where OpenAlex stores one, the abstract.
The third is the verified reference list assembled in Phase~5
(\texttt{references.bib} and \texttt{references\_part05.bib}), which carries
the technical lineage of the field whether or not the corpus ranking surfaced
it. Nothing in this part describes the content of a paper beyond what its
abstract or bibliographic record states; a result known only from an abstract
is introduced as what ``the abstract reports''. Works that exist only as
arXiv, SSRN or medRxiv deposits are called preprints throughout, and a
preprint's claim is never upgraded to a peer-reviewed finding by the number of
times it has been cited. The boundary of 30~November 2022 is a reference date
for the time-series models; the estimates below describe when the series
changed, not why.

\subsection{What the landmark set is, and what it is not}
\label{sec:ai-landmarks}

The landmark candidates of the area were screened record by record before any
of them was used, and the outcome conditions everything that follows. Of the
\SelectionByAreaAiMl{} records, the screening flagged 44 as unreliable for the
purpose of a critical synthesis (Table~\ref{tab:ai_landmark_flags}; the record-level flags are in the file
\texttt{contributions.csv} of the agent's directory under
\texttt{literature\_synthesis}).\footnote{The
flags are a human judgement by the specialist agent from identifiers, citation
counts, venue, title and abstract; they were not produced by a script and
must be audited (Part~III, Phase~8). The counts in this paragraph are those of
the flag file, reproduced in Table~\ref{tab:ai_landmark_flags}.} Thirteen are
\emph{identifier mismatches}: the title and abstract belong to an obscure work
while the arXiv identifier and the 36-month citation count belong to a famous
one. A 2026 article on translating natural language into MongoDB queries
carries the arXiv identifier of the retrieval-augmented-generation paper
\citep{lewis2020rag} and its citations; a Portuguese-language essay on
copyright carries that of GPT-3 \citep{brown2020gpt3}; an essay on
``theatrical compliance'' carries that of chain-of-thought prompting
\citep{wei2022cot}; a preprint on an intensive-care simulation carries that of
Llama~2 \citep{touvron2023llama2}; a ``reproducible experiment runner'' for the
Adam optimiser carries more than forty thousand citations. Three records are
\emph{abstract mismatches}, in which a correct title (the GPT-4 technical
report, \citealp{openai2023gpt4}; ``GPTs are GPTs'',
\citealp{eloundou2024gpts}) is paired with the abstract of another work.
Eleven are \emph{implausible years}: mixup, the variational recurrent network
and several 2013--2019 bandit papers appear as 2024--2026 works, four of them
with April--June 2026 arXiv identifiers, so that their field-normalised scores
are computed against the wrong cohort. Seven are \emph{normalisation
artefacts}, dissertations and unindexed articles with two to twenty citations
and NCS36 above five hundred, because the reference mean for a rare document
type in a young cohort is close to zero. Ten are \emph{off scope}:
control-theoretic consensus papers assigned to AI agents, an essay on
educational-technology ``solutionism'' assigned to reinforcement learning.
% [E-02, E-03]

The defect is not uniformly distributed. Identifier and abstract problems
concentrate in the large-language-model category, where eight of
thirty-three records are affected, and misdated classics concentrate in
machine-learning theory and deep-learning architectures. The mechanism lies
upstream of this review, in the merging of work records by the source
database and in the second-pass arXiv and title linkage of Phase~1
(Section~\ref{sec:m-linkage}); it is reported here because the AI area is where
it bites hardest, the arXiv identifiers of the most-cited papers of the decade
being exactly the ones most likely to be attached to the wrong record. Two
consequences are drawn. First, the clean count of citation-ranked AI landmarks
is at most 190, and every ``most influential papers'' statement of Part~IX
that concerns AI must be computed on records whose DOI, arXiv identifier and
title are mutually consistent. Second, the field-normalised score should be
recomputed with document-type-specific reference sets and with the 2026
cohort excluded; this is a recommendation to the director, not something done
here. % [E-04]

What the set does contain is instructive in its own right
(Table~\ref{tab:ai_contribution_types}). Judged from the abstract for 190
records and from title, venue and type for the 44 without one, 84 of the 234
candidates are surveys, 67 methods, 33 empirical studies, 29 theory or position
papers, 17 engineering systems and only 4 benchmarks. % [E-05]
Within the LLM and generative-AI categories the citation-ranked top is
occupied by education and commentary papers written in the months after
November 2022 \citep{kasneci2023chatgptgood, lo2023rapidreview,
rudolph2023assessment, chan2023students, baidooanu2023chatgptedu,
dwivedi2023sowhat, vandis2023priorities, lim2023ragnarok}, not by the papers
that introduced the models. % [E-06]
This is a property of ranking by field-normalised impact inside a multi-label
category that spans computer science, education and information systems: a
commentary in an education journal is normalised against education
journals, and the first wave of ChatGPT commentary was cited at rates that
no technical paper of its cohort matched within thirty-six months. The
technical lineage---GPT-4, LLaMA, InstructGPT, chain-of-thought, DPO, QLoRA,
BLIP-2, Mamba---is present in the set only where an arXiv record was linked,
and sometimes, as just shown, linked to the wrong record. The synthesis
therefore grounds its technical claims on the verified reference list and uses
the landmark set to characterise what the corpus \emph{rewards} rather than
what the field \emph{invented}. Table~\ref{tab:ai_landmarks}, placed at the end of this part, lists all
234 candidates with their flag and contribution type.

\input{tables/tab_ai_landmark_flags}
\input{tables/tab_ai_contribution_types}

\subsection{The corpus picture of the nine categories}
\label{sec:ai-corpus}

Table~\ref{tab:ai_growth} gives the primary-category volume, the compound
annual growth rates of the baseline (2018--2022) and post (2023--2025) periods
and the share of eligible works for each category; Table~\ref{tab:ai_counts}
gives the yearly multi-label counts, which are several times larger because
the AI categories overlap heavily and a work counts in every category it is
assigned to. The ordering is the same under either definition. Large language
models grow fastest in the post period, generative AI, multimodal AI,
foundation models and AI safety follow, reinforcement learning and
machine-learning theory accelerate modestly, deep-learning architectures grow
slightly more slowly than in the baseline, and AI agents reverse a negative
baseline growth rate. % [E-07]
The share of eligible works carrying the LLM label as primary category rises by
more than an order of magnitude between 2018 and 2025. % [E-08]
Two cautions apply. The 2026 column of Table~\ref{tab:ai_counts} covers
publications dated up to the snapshot of 14~October 2026 and is incomplete;
and the categories are defined by rules over topics, arXiv classes, venues and
keyword patterns (Section~\ref{sec:m-classification}), so that a work on
reward modelling for language models is an LLM work and not a reinforcement-learning
work, a point that matters for Section~\ref{sec:ai-rl}.
A third caution concerns the flows of Table~\ref{tab:ai_flows}. Because a
cited work can carry several categories, the row shares of a citing category
do not sum to one across the columns shown, and a reference from an LLM paper
to a paper that is both an LLM and an NLP work counts under both destinations;
the ``self'' share of a category is therefore an upper bound on its
intellectual closure, and changes in shares are more informative than their
levels.

\input{tables/tab_ai_growth}
\input{tables/tab_ai_counts}

The structural-change estimates are in Table~\ref{tab:ai_breaks} and, for the
alternative event dates, Table~\ref{tab:ai_altdates}. For the LLM series the
interrupted time series of eq.~\eqref{eq:its} fitted at
\ItsEventsChatgpt{} gives a level change of
\ItsSeriesCatLlmItsChatgptLevelChange{} log points
(HAC $p = \ItsSeriesCatLlmItsChatgptLevelP$) and a slope change of
\ItsSeriesCatLlmItsChatgptSlopeChange{} log points per month
($p = \ItsSeriesCatLlmItsChatgptSlopeP$) over a pre-period slope of
\ItsSeriesCatLlmItsChatgptPreSlope{}; the placebo distribution of
Section~\ref{sec:m-placebo} places the level statistic at a placebo
$p$-value of \ItsSeriesCatLlmPlaceboChatgptLevelTPlaceboP{} and the slope
statistic at \ItsSeriesCatLlmPlaceboChatgptSlopeTPlaceboP{}; the
Bai--Perron procedure selects \ItsSeriesCatLlmBaiPerronMBic{} break, dated
in Table~\ref{tab:ai_breaks} to the middle of 2022, and the Bayesian
on-line change-point posterior mass between December 2022 and March 2023 is
\ItsSeriesCatLlmBocpdCpProbTwoZeroTwoTwoOneTwoToTwoZeroTwoThreeZeroThree{}.
Excluding three months on either side of the event (the donut specification)
changes the level estimate to
\ItsSeriesCatLlmItsChatgptDonutLevelChange{}
($p = \ItsSeriesCatLlmItsChatgptDonutLevelP$). % [E-09]
The same configuration---no significant level jump at the reference month, a
large and significant slope change, a placebo $p$-value for the level that
does not single out December 2022, a data-driven break before or well after
it---holds for multimodal AI (slope change
\ItsSeriesCatMultimodalAiItsChatgptSlopeChange) and generative AI
(\ItsSeriesCatGenerativeAiItsChatgptSlopeChange). AI agents (level
\ItsSeriesCatAiAgentsItsChatgptLevelChange, $p =
\ItsSeriesCatAiAgentsItsChatgptLevelP$; slope
\ItsSeriesCatAiAgentsItsChatgptSlopeChange), reinforcement learning (level
\ItsSeriesCatReinforcementLearningItsChatgptLevelChange, $p =
\ItsSeriesCatReinforcementLearningItsChatgptLevelP$) and AI safety (level
\ItsSeriesCatAiSafetyItsChatgptLevelChange, $p =
\ItsSeriesCatAiSafetyItsChatgptLevelP$) show \emph{negative} level changes
followed by positive slope changes, a dip and a subsequent acceleration whose
reading is taken up category by category below. Deep-learning architectures is
the only AI category whose slope change
(\ItsSeriesCatDeepLearningArchitecturesItsChatgptSlopeChange, $p =
\ItsSeriesCatDeepLearningArchitecturesItsChatgptSlopeP$) is not
distinguishable from zero. % [E-10]

\input{tables/tab_ai_breaks}
\input{tables/tab_ai_altdates}

The alternative dates of Section~\ref{sec:m-altdates} order themselves
monotonically. Level effects estimated at the GPT-3 (\ItsEventsGptThree) and
DALL-E~2 (\ItsEventsDalleTwo) dates are negative for every AI category,
whereas those at the GPT-4 (\ItsEventsGptFour) and o1 (\ItsEventsOOne) dates
are positive and significant for LLM, multimodal and generative AI: for the
LLM series, \ItsSeriesCatLlmItsGptThreeLevelChange{} at the GPT-3 date
($p = \ItsSeriesCatLlmItsGptThreeLevelP$), \ItsSeriesCatLlmItsGptFourLevelChange{}
at the GPT-4 date ($p = \ItsSeriesCatLlmItsGptFourLevelP$) and
\ItsSeriesCatLlmItsOOneLevelChange{} at the o1 date
($p = \ItsSeriesCatLlmItsOOneLevelP$). % [E-11]
Each of these is a separate fit with the event moved, and a positive level
effect at a later date in a convex series is what a steepening trend produces
whether or not anything happened at that date. The reading that the evidence
supports is therefore restricted: the acceleration of the AI literature is a
change of trend that the data-driven procedures place in the middle of 2022
and that steepened through 2023--2024; the monthly series cannot attribute it
to one product release, and this part does not. % [E-12]

\subsection{Knowledge flows and detected emergence}
\label{sec:ai-flows}

Table~\ref{tab:ai_flows} reports, for each AI category and for 2019, 2022 and
2025, the destination of its in-corpus references, i.e.\ the row shares
$s_{c \to c'}(t)$ of eq.~\eqref{eq:flow}. The large-language-model category is
the one whose profile changed most. In 2019 and 2022 roughly seven in ten of
its in-corpus references went to natural language processing; in 2025 about a
third do, while more than a quarter go to the LLM category itself and
measurable shares go to human--computer interaction and software engineering.
% [E-13]
Generative AI moved from a vision-centred profile to one split between vision,
LLM and HCI; AI safety, which cited LLM work at a negligible rate in 2019,
directs a tenth of its references there in 2025 and slightly more to HCI.
% [E-14]
Reinforcement learning moves the other way, citing robotics more and LLM work
almost not at all, and machine-learning theory and deep-learning architectures
are essentially unchanged. % [E-15]
At the level of categories, then, the LLM literature has become a citation hub
for generative AI, safety and agents while remaining peripheral to
reinforcement learning, theory and architectures. This is the corpus-level
answer to RQ5 that this part can give; it is consistent with the convergence
thesis of \citet{bommasani2021foundation}, but a change in which category a
field cites is not evidence that its methods converged, and
Section~\ref{sec:ai-synthesis} returns to the distinction. % [E-16]

\input{tables/tab_ai_flows}

The preregistered known-emergence set of Section~\ref{sec:m-emerging} contains
eleven AI concepts, and their lead times are the first empirical test of
RQ10 in this area. The detector flags \emph{large language models} in
\KnownLargeLanguageModelsDetectionMonth{},
\KnownLargeLanguageModelsLeadMonths{} months before the peak-growth month
\KnownLargeLanguageModelsPeakGrowthMonth{}, on
\KnownLargeLanguageModelsTotalWorks{} matching works; \emph{reasoning models}
\KnownReasoningModelsLeadMonths{} months ahead; retrieval-augmented generation
\KnownRetrievalAugmentedGenerationLeadMonths{} months; vision-language models
\KnownVisionLanguageModelsLeadMonths{}; diffusion models
\KnownDiffusionModelsLeadMonths{}; LLM agents \KnownLlmAgentsLeadMonths{};
state-space sequence models \KnownStateSpaceSequenceModelsLeadMonths{};
mixture of experts \KnownMixtureOfExpertsLeadMonths{}; speculative decoding
\KnownSpeculativeDecodingLeadMonths{}; AI code generation
\KnownAiCodeGenerationLeadMonths{}; and mechanistic interpretability
\KnownMechanisticInterpretabilityLeadMonths{}, a negative lead in which the
peak-growth month precedes detection on only
\KnownMechanisticInterpretabilityTotalWorks{} matching works. The median lead
over the whole set is \KnownSetMedianLeadMonths{} months. % [E-17]
The long lead for language models is not foresight: the regular expression
matches the pre-2020 literature on ``language models'' and the detector fires
on its slow growth years before the 2023 surge. The short leads for
state-space models, mixture of experts and speculative decoding are the
honest result, and they say that topics driven by a single widely adopted
preprint are detected only months before their growth peaks.

\section{Large language models and foundation models}
\label{sec:ai-fm}

\subsection{The established lineage}

The architecture that dominates the window is the decoder-only Transformer
\citep{vaswani2017attention} trained autoregressively at scale
\citep{radford2019gpt2, brown2020gpt3, chowdhery2022palm}, with the encoder
and encoder--decoder variants of the baseline period \citep{devlin2019bert,
raffel2020t5} surviving mainly as embedding models. Two empirical regularities
from before the window organise the technical literature inside it. The first
is the scaling law, in its original form a power law of loss in parameters $N$
and tokens $D$ \citep{kaplan2020scaling} and in the compute-optimal form of
\citet{hoffmann2022chinchilla}, which, for a fixed training budget $C \approx
6ND$, allocates roughly equal factors of growth to $N$ and $D$ and therefore
implies far more tokens per parameter than the earlier prescription. The
second is the sparse mixture-of-experts layer \citep{shazeer2017moe,
fedus2022switch}, which decouples parameter count from per-token compute. The
open-weight lineage LLaMA, Llama~2 and Llama~3 \citep{touvron2023llama,
touvron2023llama2, dubey2024llama3}---all three preprints, the first of which
is in the landmark set---follows the Chinchilla allocation; the LLaMA abstract
reports that a 13-billion-parameter model trained only on publicly available
data outperforms GPT-3 on most benchmarks, a claim that the abstract makes and
that this review has not verified. % [E-18]
The closed frontier releases are documented by technical reports without
training details \citep{openai2023gpt4, gemini2023, openai2024o1}; they are
cited here as the dated artefacts they are, and the GPT-4 record in the
landmark set, as noted above, carries someone else's abstract. The LLM surveys
of the window \citep{zhao2023llmsurvey, naveed2023llmoverview,
minaee2024llmsurvey, min2023plmsurvey} map a sub-literature that by 2025 had
its own citation core (Table~\ref{tab:ai_flows}); the journal version of the
first of them is in the landmark set and its abstract organises the field
into pre-training, adaptation, utilisation and evaluation.

The most visible post-2022 development in the landmark set is not a new
general model but the diffusion of the foundation-model recipe into domains
with their own data. Self-supervised pathology encoders trained on hundreds of
thousands of whole-slide images \citep{chen2024uni, lu2024conch,
wang2024chief, huang2023plip, ding2025titan}, a radiology model trained on
two- and three-dimensional scans \citep{wu2025radfm}, a dermatology model on
two million clinical images \citep{yan2025panderm}, a brain-MRI model
\citep{tak2026brainiac}, an Earth-system model trained on more than a million
hours of geophysical data \citep{bodnar2025aurora}, a tabular model that the
abstract describes as a prior-data fitted network outperforming
gradient-boosted trees on small datasets \citep{hollmann2025tabpfn},
time-series representation models \citep{jin2023timellm, liu2024timesurl} and
a language model fine-tuned on ten million human choices to predict behaviour
across psychological experiments \citep{binz2025centaur} are, with the
exception of the time-series preprint, peer-reviewed journal or conference
papers, and each abstract reports outperforming task-specific baselines on a
battery of benchmarks. % [E-19]
What the abstracts do not report, and what the reviews in the same set list as
unresolved \citep{paschali2025fmradiology, zhang2026datacentric}, is external,
multi-site validation under distribution shift; the claim that domain
foundation models will displace task-specific pipelines in regulated settings
is speculative and is carried to Part~X as such. % [E-24]

Below the frontier, the practical growth of the open-weight ecosystem
concentrated on adaptation. Low-rank adaptation \citep{hu2021lora}, its
four-bit variant QLoRA, whose abstract reports fine-tuning a 65-billion-parameter
model on a single 48\,GB GPU \citep{dettmers2023qlora}, the zero-initialised
adaption prompts of LLaMA-Adapter \citep{zhang2023llamaadapter} and the
parameter-efficient fine-tuning surveys \citep{han2024peftsurvey} define a
regime in which a research group without frontier compute can specialise a
model in hours. The QLoRA abstract's statement that its best model reaches
``99.3\,\% of the performance level of ChatGPT'' on the Vicuna benchmark is a
judge-based measurement, a point to which Section~\ref{sec:ai-eval} returns.
% [E-20]

\subsection{Data as the contested resource}

The compute-optimal allocation of \citet{hoffmann2022chinchilla} turned
training data from an afterthought into the binding constraint, and the
landmark set records the consequences from three sides without resolving any
of them. On the supply side, LAION-5B \citep{schuhmann2022laion5b} and the
public-data training of LLaMA \citep{touvron2023llama} show that web-scale
corpora could be assembled outside the frontier laboratories; the surveys of
data-centric foundation models in healthcare \citep{zhang2026datacentric} and
of multimodal fusion on low-quality data \citep{zhang2026lowqualityfusion}
describe, from the methods side, how quality, annotation, privacy and
missingness rather than volume limit the domain models of
Section~\ref{sec:ai-fm}. On the integrity side, the poisoning study of
\citet{alber2025poisoning} reports that a vanishingly small fraction of
deliberately false tokens in a public pretraining corpus suffices to shift a
model's medical answers, which turns the openness of the data pipeline into an
attack surface. On the legal side, the one landmark that addresses the
provenance of training data is a Portuguese-language essay on copyright whose
record, as shown above, carries GPT-3's identifier and citations; the
question it raises is real and the corpus's answer is an artefact. The three
sides share a measurement gap that no abstract in the set closes: the
composition of the corpora on which the frontier models were trained is not
disclosed, so that neither the scaling-law prescriptions nor the
contamination caveats of Section~\ref{sec:ai-eval} can be checked against the
models that matter most. Data, not architecture, is where the open and closed
lineages of the window diverge, and it is the resource whose exhaustion,
curation and legal status the next part of the manuscript lists among the
frontiers.

\subsection{Contested claims: emergence, sparks and pure-RL reasoning}

Three claims about what scaled language models are have been argued largely in
preprints, and the corpus can record the argument but not settle it. The first
is emergence. \citet{wei2022emergent} report abilities that appear
discontinuously as scale increases; \citet{schaeffer2023mirage} show that for
the tasks examined the discontinuities are produced by discontinuous
metrics---exact-match accuracy on multi-token answers---and largely vanish
under continuous or token-level ones. Both are preprints, the disagreement is
about measurement rather than data, and the manuscript presents it as such.
% [E-21]
The second is ``Sparks of Artificial General Intelligence''
\citep{bubeck2023sparks}, which reports qualitative experiments on an early,
unreleased version of GPT-4 and, in the abstract's words, contends that the
model exhibits more general intelligence than previous systems. The record is
among the most cited of the landmark set, it is a preprint, its experiments
are not reproducible because the model version is unavailable, and the
critique that scale does not confer understanding \citep{bender2021parrots}
is in the same set with a comparable citation count. % [E-22]
The third is the reasoning turn: the o1 system card \citep{openai2024o1}
moved computation to inference time, and the DeepSeek-R1 paper, whose
peer-reviewed version is in the landmark set \citep{deepseek2025r1nature}
alongside the preprint \citep{deepseek2025r1}, reports that reasoning ``can be
incentivized through pure reinforcement learning, obviating the need for
human-labelled reasoning trajectories''. The claim is peer reviewed; its
generality beyond domains with verifiable rewards---mathematics, code---is not
tested by anything in the corpus, and it is the first entry of the open
problems in Section~\ref{sec:ai-synthesis}. % [E-23]

\section{Alignment and post-training}
\label{sec:ai-align}

The recipe that turned a pretrained language model into an assistant was in
place before the window: a reward model fitted to human preference
comparisons \citep{christiano2017rlhf} and a policy optimised against it with
a proximal objective \citep{schulman2017ppo}, applied to instruction following
in InstructGPT \citep{ouyang2022instructgpt} and extended with helpfulness and
harmlessness data \citep{bai2022hh} and with model-written critiques against a
written constitution \citep{bai2022constitutional}. Inside the window the
method's two moving parts were simplified and questioned. Direct preference
optimisation \citep{rafailov2023dpo} rewrites the RLHF objective so that the
policy is fitted directly to preference pairs through a closed-form
log-ratio loss, removing the separate reward model and the on-policy sampling;
the NeurIPS record is in the landmark set, without an abstract in the file.
The open problems of the reward-model route---reward hacking, preference
mis-specification, evaluator disagreement, the gap between stated and revealed
preferences---are catalogued by \citet{casper2023rlhfproblems}, a preprint
whose taxonomy is nevertheless the standard reference. % [E-25]
The corpus adds a sociological observation to this methodological one.
Reinforcement learning as a category directs less than one per cent of its
in-corpus references to LLM work in 2025 (Table~\ref{tab:ai_flows}), while the
LLM category's own self-citation share grew to more than a quarter. The
``return of reinforcement learning'' through post-training is therefore a
phenomenon of the LLM literature, which absorbed PPO, DPO and the group-relative
policy optimisation of DeepSeek-R1 as its own methods; the classification
separates the two communities even where the mathematics is shared, and
Section~\ref{sec:ai-rl} draws the consequence for any claim about the fate of
reinforcement learning.

That aligned behaviour is brittle under optimisation pressure is established
by three lines of work. Universal adversarial suffixes found by gradient-guided
search transfer across models \citep{zou2023universal}; the ``jailbroken''
analysis attributes failures to competing objectives and mismatched
generalisation between pretraining and safety training
\citep{wei2023jailbroken}; and PAIR, a peer-reviewed method in the landmark
set, is described in its abstract as an attacker language model that generates
semantic jailbreaks against a target with black-box access, in about twenty
queries \citep{chao2025pair}. % [E-26]
The first two are preprints and the third was published in 2025; together they
show that every public alignment method of the window can be defeated by a
modest adaptive budget. Whether this brittleness is intrinsic to
preference-based alignment or an artefact of current scale and data is the
question that none of the landmark abstracts answers, and it is carried to
Part~X.

\section{Reasoning, retrieval, tool use and code}
\label{sec:ai-reason}

Chain-of-thought prompting \citep{wei2022cot, kojima2022zeroshotcot},
tree-structured search over thoughts \citep{yao2023tot} and verbal
self-reflection across episodes \citep{shinn2023reflexion} are the prompting
lineage of the reasoning literature; their evidence is benchmark gains, mostly
on arithmetic and symbolic tasks, and several of the records are preprints.
The reasoning-model turn of 2024--2025 \citep{openai2024o1,
deepseek2025r1nature} internalised the procedure by training the model to
produce long intermediate traces and by spending compute at inference rather
than at training. The detector of Section~\ref{sec:m-emerging} flagged
\emph{reasoning models} in \KnownReasoningModelsDetectionMonth{},
\KnownReasoningModelsLeadMonths{} months before the peak-growth month of
\KnownReasoningModelsPeakGrowthMonth{}; the long lead is an artefact of a
regular expression that includes chain-of-thought and therefore matched the
2023 prompting literature, and the honest reading is that the concept's
vocabulary existed well before its models. % [E-17]

Retrieval augmentation is the most widely adopted of the window's patterns.
The dense passage retriever \citep{karpukhin2020dpr} and the
retrieval-augmented generation architecture \citep{lewis2020rag} precede the
window; their adoption inside it is surveyed by \citet{gao2023ragsurvey}, and
the known-set detector counts \KnownRetrievalAugmentedGenerationTotalWorks{}
matching works, detected in \KnownRetrievalAugmentedGenerationDetectionMonth{}
with peak growth in \KnownRetrievalAugmentedGenerationPeakGrowthMonth{}.
Tool use---ReAct's interleaving of reasoning traces and actions
\citep{yao2023react}, Toolformer's self-supervised API calls
\citep{schick2023toolformer}, ToolLLM's instruction tuning over thousands of
real APIs \citep{qin2023toolllm}, surveyed by \citet{mialon2023augmented}---is
the bridge from language models to the agents of Section~\ref{sec:ai-agents};
all four are preprints. Hallucination, the generation of fluent content
unsupported by the input or by fact, is the failure mode that these
mechanisms are meant to contain, and it remains a measurement problem before
it is a modelling problem: the surveys \citep{ji2023hallucination,
huang2025hallucination} agree on taxonomies and disagree on metrics. The
clearest statement of why is in the landmark set. The Med-PaLM paper
\citep{singhal2023clinical} built MultiMedQA from six medical
question-answering datasets and a new set of consumer questions, and its
abstract reports that a human evaluation framework along axes of factuality,
comprehension, reasoning and harm was needed because automated evaluation
against limited benchmarks did not meet the bar for clinical use. % [E-27]

Code generation is the application in which the window's models were first
measured against professional work. The Codex evaluation
\citep{chen2021codex}, AlphaCode's competition-level results
\citep{li2022alphacode}, the open code models \citep{nijkamp2023codegen,
roziere2023codellama, li2023starcoder} and the SWE-bench task of resolving
real repository issues \citep{jimenez2024swebench} define the benchmark
sequence; the controlled productivity experiment of \citet{peng2023copilot}
and the field data of \citet{ziegler2022copilot} are the usage evidence, with
the usability and security caveats of \citet{vaithilingam2022expectation},
\citet{barke2023groundedcopilot} and \citet{pearce2022asleep}. The detector's
lead for \emph{AI code generation} is \KnownAiCodeGenerationLeadMonths{}
months. The software-engineering reading of this literature belongs to
Part~VII; here it is noted only that the LLM category's references to software
engineering in 2025 (Table~\ref{tab:ai_flows}) are the trace of this
application in the citation graph.

\section{Multimodal and generative AI}
\label{sec:ai-multimodal}

\subsection{Vision--language models}

Two baseline-period results make the multimodal literature of the window
possible: contrastive image--text pre-training, which aligns an image encoder
and a text encoder on web-scale pairs \citep{radford2021clip}, and the vision
Transformer, which made the image encoder the same object as the text model
\citep{dosovitskiy2021vit}. The single-stream alternative, in which words and
image regions are tokens of one Transformer, is represented in the landmark
set by VL-BERT \citep{su2019vlbert}; the taxonomy of representation,
translation, alignment, fusion and co-learning that the field still uses is
that of \citet{baltrusaitis2019multimodal}, also in the set. Inside the window
the dominant design is a frozen or lightly adapted language model fed by a
visual encoder through a small bridge: a perceiver resampler in Flamingo
\citep{alayrac2022flamingo}; a querying Transformer trained in two stages
between frozen encoders and frozen LLMs in BLIP-2, whose abstract motivates
the design by the prohibitive cost of end-to-end pre-training
\citep{li2023blip2}; instruction tuning of the same bridge in InstructBLIP
\citep{dai2023instructblip}; a linear projection in LLaVA
\citep{liu2023llava}; shared functional modules with the language decoder as
``universal interface'' in mPLUG-Owl2 \citep{ye2024mplugowl2}. The promptable
segmentation model \citep{kirillov2023sam} is the dense-prediction analogue,
and the open image--text corpus LAION-5B \citep{schuhmann2022laion5b} is what
made CLIP-scale training reproducible outside industry. % [E-28]
Of these records, BLIP-2, InstructBLIP, LLaVA and LAION-5B are preprint or
proceedings deposits; mPLUG-Owl2 is a CVPR paper. The surveys
\citep{yin2024mllmsurvey, zhang2024mmllms} organise the design space; the
corpus-level trace is the vision-language detector, which fires
\KnownVisionLanguageModelsLeadMonths{} months before the concept's peak growth.

Whether these models \emph{fuse} modalities or merely \emph{condition} on the
text channel is the open question that the landmark abstracts raise most
sharply. An evaluation of eight proprietary and open multimodal foundation
models on 1,090 medical cases reports, in its abstract, that image predictions
are ``largely driven by text'', that accuracy rises monotonically with the
amount of informative text, and that even mild textual suggestions of an
incorrect diagnosis degrade image-based classification
\citep{buckley2026textexploit}. % [E-30]
The scoping review of 432 medical multimodal papers \citep{schouten2025multimodalmedicine}
and the survey of fusion under noisy, missing and imbalanced modalities
\citep{zhang2026lowqualityfusion} describe the same robustness gap from the
methods side. The medical records in the set---generalist biomedical and
radiology models \citep{zhang2024biomedgpt, wu2025radfm}, reinforcement-learning
fine-tuning for medical reasoning in vision-language models and a zero-shot
evaluation of GPT-5 on medical visual question answering, the latter a
preprint \citep{lai2026medr1, wang2025gpt5medical}---report benchmark gains
but not modality ablations; the question is carried to Part~X.

\subsection{Diffusion models and their applications}

The generative lineage of the window is diffusion. Denoising diffusion
probabilistic models \citep{ho2020ddpm} and their latent-space formulation
\citep{rombach2022ldm} precede the boundary; Imagen \citep{saharia2022imagen}
and DALL-E~2 \citep{ramesh2022dalle2} are the 2022 text-to-image systems,
and the Imagen record is in the landmark set without an abstract. Three
technical contributions in the set define how the models are controlled.
Classifier-free guidance \citep{ho2022cfg}, a preprint, trains the conditional
and unconditional score jointly and interpolates between them at sampling
time, removing the separate classifier that earlier guidance required;
textual inversion \citep{gal2022textualinversion}, also a preprint, learns a
new token embedding from three to five images of a concept so that a frozen
model can render it; T2I-Adapter \citep{mou2024t2iadapter}, an AAAI paper,
trains small adapters that align external structural or colour signals with
the internal knowledge of a frozen text-to-image model. The Transformer
backbone for diffusion \citep{peebles2023dit} and its rectified-flow
successor \citep{esser2024sd3}, a preprint, close the architectural loop by
replacing the U-Net. The detector flags \emph{diffusion models}
\KnownDiffusionModelsLeadMonths{} months before peak growth. Outside images,
the one peer-reviewed scientific application in the landmark set is
MatterGen \citep{zeni2025mattergen}, whose abstract reports a diffusion model
that generates stable, diverse inorganic materials across the periodic table
and can be fine-tuned towards property constraints. % [E-29]
Speech followed the same self-supervised path a few years earlier
\citep{baevski2020wav2vec2, hsu2021hubert} and reached general-purpose
recognition with weakly supervised training at scale \citep{radford2023whisper};
the review of \citet{mehrish2023speechreview} covers the category, which the
taxonomy places outside the nine categories of this part.

\subsection{What the generative-AI category rewards}

Generative AI is the category whose citation profile changed most inside the
window, from a vision-centred one to a three-way split between vision, LLM
and HCI (Table~\ref{tab:ai_flows}). Its citation-ranked landmarks are,
however, not the diffusion papers just described but education, policy and
adoption studies: a policy framework from surveys of students and staff
\citep{chan2023policyframework}, scoping and systematic reviews of assessment
and academic integrity \citep{xia2024assessmentreview, bittle2025integrity},
analyses of the guidelines of 116 and of 50 United States universities
\citep{mcdonald2025hepolicies}, a paradox-theory essay \citep{lim2023ragnarok},
an OECD outlook, and two economics papers: a preregistered field experiment,
circulated as an SSRN preprint, whose title states that generative AI can harm
learning \citep{bastani2024harmlearning}, and nationally representative
surveys whose abstract reports that 45 per cent of United States adults aged
18--64 used generative AI by late 2024 and that work adoption was faster than
that of the personal computer \citep{bick2026rapidadoption}. % [E-31]
The pattern repeats the one observed for LLMs in Section~\ref{sec:ai-landmarks}:
technical generative-AI papers are counted in the category but, normalised
against their own cohort, rarely reach its citation top within thirty-six
months, whereas the first social-science and education responses to the
technology do. A reader who takes the landmark list as the field's
intellectual map would conclude that generative AI is a topic in education
research; the correct conclusion is that field normalisation inside a
multi-label category measures discourse, and that the measurement is
informative about diffusion rather than about invention.

\section{AI agents and autonomous systems}
\label{sec:ai-agents}

\subsection{From tool use to agents}

An LLM-based agent, in the definition shared by the surveys of the window
\citep{wang2024agentsurvey, xi2023rise, guo2024multiagent,
cheng2026agentsdefinitions}, is a language model wrapped in a loop of
perception, planning, memory and action, in which the model's outputs are
parsed into calls to tools or to other agents and the results are fed back as
context. The architectures that fixed the vocabulary are generative agents in
a simulated town \citep{park2023generativeagents}, a peer-reviewed UIST paper;
the multi-agent conversation framework AutoGen \citep{wu2023autogen}; virtual
context management in MemGPT \citep{packer2023memgpt}; and the ``LLM as
operating system'' proposals \citep{mei2024aios, ge2023llmos}, all of the
latter preprints. The memory component is surveyed by
\citet{zhang2025agentmemory}, whose abstract calls self-evolution through
agent--environment interaction the feature that distinguishes agents from the
models they wrap. Evaluation lags design. The benchmarks that the community
uses---SWE-bench for repository-level issue resolution
\citep{jimenez2024swebench}, WebArena for web tasks \citep{zhou2023webarena},
AgentBench across environments \citep{liu2023agentbench}---are preprints, and
the only peer-reviewed agent benchmark in the landmark set is a clinical
one, an evaluation of an open-source and a proprietary agent system on three
families of medical decision tasks \citep{liu2026clinicalagents}. % [E-32]
The detector flags \emph{LLM agents} in \KnownLlmAgentsDetectionMonth{},
\KnownLlmAgentsLeadMonths{} months before peak growth in
\KnownLlmAgentsPeakGrowthMonth{}, on \KnownLlmAgentsTotalWorks{} matching
works; the concept arrived and peaked within a year.
Three features of agent evaluation limit what the benchmarks can show. The
unit of evaluation is a trajectory rather than an answer, so that
non-determinism in the environment and in sampling compounds across steps and
the variance warning of \citet{henderson2018drlmatters} applies with greater
force than in single-turn evaluation; the cost of a run---tokens, tool calls,
wall-clock time---is rarely reported alongside the success rate, so that
rankings conflate capability with budget; and the environments are public and
static, so that the contamination problem of Section~\ref{sec:ai-eval} applies
to the tasks as much as to the models. The clinical benchmark in the landmark
set \citep{liu2026clinicalagents} is notable precisely because its abstract
reports comparing an open-source and a proprietary system under one protocol
across three task families; it is one study.

The strongest peer-reviewed evidence for agents is scientific. The Virtual Lab
\citep{swanson2025virtuallab}, a Nature paper whose record carries no abstract
in the corpus file, reports agents designing SARS-CoV-2 nanobodies that were
validated experimentally; Robin \citep{ghareeb2026robin} is described in its
abstract as the first multi-agent system automating both hypothesis generation
and data analysis in experimental biology; DIVE \citep{zhang2026dive} extracts
experimental data from the figures of the literature for hydrogen-storage
materials; AgentSociety \citep{piao2026agentsociety} simulates large
populations of LLM-driven generative agents for social science. % [E-33]
These are single demonstrations with experimental or observational
validation, which is more than the ``AI scientist'' preprints
\citep{lu2024aiscientist, si2024novelideas} offer; whether autonomous research
agents produce validated discoveries at a rate that justifies their cost is a
question for preregistered multi-laboratory trials, not for case reports, and
it is listed in Part~X.

\subsection{``Agentic AI'': a definitional literature}

The landmark set of the category contains six papers from 2025--2026 whose
purpose is to define agentic AI and distinguish it from AI agents, generative
AI and classical multi-agent systems \citep{acharya2025agenticsurvey,
sapkota2026agentsvsagentic, bandi2025agenticreview, abouali2025agenticparadigms,
allmendinger2026multiagentai, collaco2026agentichealthcare}, most of them
reviews of reviews, one of them a PRISMA-based review of ninety studies that
proposes a symbolic-versus-neural dual-paradigm framework, and one benchmark.
% [E-34]
A literature that produces six taxonomies and one benchmark in eighteen
months is a literature whose object is not yet stabilised, and the question
of whether ``agentic AI'' denotes a distinct class of systems or a relabelling
of tool-augmented LLM pipelines is treated here as open. The trust, risk and
security management framework for multi-agent LLM systems
\citep{raza2026trism} belongs to the same genre: a structure without an
empirical test.

The corpus series of the category must be read with its composition in mind.
Four of its landmark records are control-theoretic papers on consensus and
formation control of nonlinear multi-agent systems, and its pre-2022 volume
is dominated by classical multi-agent systems, distributed artificial
intelligence and robotics, which explains why the category directs a large
and growing share of its references to robotics (Table~\ref{tab:ai_flows}).
The negative baseline growth rate, the negative level change at the reference
month and the second Bai--Perron break of 2024 (Table~\ref{tab:ai_breaks})
are the signature of a category in which an older literature was flat or
declining while a new one, LLM agents, grew from a small base; the level
effect at the o1 date, \ItsSeriesCatAiAgentsItsOOneLevelChange{}
($p = \ItsSeriesCatAiAgentsItsOOneLevelP$), is where the new literature
becomes visible in the aggregate. % [E-35]
The explanation is inferential; separating the two literatures would require a
sub-classification that the taxonomy does not provide. Embodied agents, from
vision-language-action models \citep{brohan2023rt2, kim2024openvla,
black2024pi0} to the driving model in the landmark set, are classified under
robotics and are reviewed there.

\section{Reinforcement learning}
\label{sec:ai-rl}

The reinforcement-learning landmarks are, almost without exception, the canon
of the baseline period. Rainbow's combination of six extensions to deep
Q-learning with an ablation of each \citep{hessel2018rainbow}; the
maximum-entropy actor--critic \citep{haarnoja2018sac}, whose record is the
arXiv deposit; clipped double-Q learning and delayed policy updates against
overestimation \citep{fujimoto2018td3}; self-play without domain knowledge
beyond the rules in AlphaZero \citep{silver2018alphazero}; grandmaster-level
StarCraft~II \citep{vinyals2019alphastar}, a Nature record without an abstract
in the file; the finding that proximal policy optimisation is, in the words of
its abstract, surprisingly strong in cooperative multi-agent benchmarks
\citep{yu2021mappo}; the large-scale study of curiosity-driven learning without
extrinsic rewards \citep{burda2018curiosity}; and the reproducibility warning
of \citet{henderson2018drlmatters}, whose abstract notes that non-determinism
and intrinsic variance make reported results hard to interpret. The theoretical
side is represented by an overview of multi-agent reinforcement learning
\citep{zhang2021marloverview} and a survey of the synthesis of model-predictive
control and reinforcement learning (record 7128436456 of the landmark file).
The single post-2022 method landmark is DreamerV3 \citep{hafner2025dreamerv3},
published in Nature, whose abstract reports a world-model algorithm that
outperforms specialised methods across more than 150 tasks with a single
configuration. % [E-36]
Several records are misdated---a 2008 inverse-reinforcement-learning paper
and a 2007 Bayesian one appear as 2018 and 2025 works---and four are
control-theory papers; the category's citation top is therefore a mixture of
classics, control theory and one new result.
What the one new result represents is the model-based turn. DreamerV3 learns
a latent world model and improves its policy on imagined trajectories, and
its abstract stresses that a single configuration transfers across domains
where earlier algorithms required per-domain tuning; the survey of
model-predictive control and reinforcement learning in the set makes the
complementary point that the two fields derive from the same principles and
differ mainly in whether an environment model is part of the algorithm. Read
together with the rising share of references from the category to robotics,
they describe a field whose centre of gravity moved towards control of
physical systems while its optimisation machinery moved, under another name,
into language-model training.

The corpus places reinforcement learning among the modestly accelerating
categories: a post-period growth rate only slightly above the baseline
(Table~\ref{tab:ai_growth}), a slope change at the reference month of
\ItsSeriesCatReinforcementLearningItsChatgptSlopeChange{} log points per
month after a negative level change, and a citation profile that moves towards
robotics. % [E-15]
It would be a mistake to read this as the field's eclipse. Reinforcement
learning's most consequential role in the window---reward modelling and policy
optimisation for language models, from PPO through DPO to the group-relative
method of DeepSeek-R1---is invisible in the category because the papers that
carry it are classified as LLM work, and the flow of references from the RL
category to the LLM category stays below one per cent. Any statement that
reinforcement learning declined after 2022 would be an artefact of the
taxonomy, and the correct statement is that its methods migrated into another
category's literature. % [E-37]

\section{Machine-learning theory, architectures and efficiency}
\label{sec:ai-theory}

\subsection{What ``machine-learning theory'' contains in the corpus}

The category named machine-learning theory does not, in its citation-ranked
top, contain a learning-theory result from the window. Its landmarks are
textbooks and textbook chapters, the case for the Matthews correlation
coefficient over accuracy and F$_1$ on imbalanced binary problems
\citep{chicco2020mcc}, the argument that high-stakes decisions should use
inherently interpretable models rather than post-hoc explanations of black
boxes \citep{rudin2019stopexplaining}, the conceptual analysis of
interpretability \citep{lipton2018mythos}, exact Shapley attributions for tree
ensembles \citep{lundberg2020treeexplainer}, the review of stochastic-gradient
optimisation for large-scale learning \citep{bottou2018optimization}, a
transfer-learning survey \citep{zhuang2021transfersurvey}, a reprinted chapter
by Verma and Pearl on the equivalence of causal models, and the misdated
bandit papers of Section~\ref{sec:ai-landmarks}. % [E-38]
The category as the rules define it is general machine learning, and what the
citation ranking rewards within it is pedagogy and measurement convention.
Theoretical work on in-context learning, on the expressivity of Transformers
or on the statistical basis of scaling laws exists inside the window, but it is
classified under LLM or deep-learning architectures, or is cited too little
within thirty-six months to surface. The corpus therefore cannot be used to
say that theory lagged practice in the AI transformation; it can only say that
the theory which is cited quickly is the theory of evaluation and
interpretation.

\subsection{Architectures: proposals, refutations and the efficiency turn}

The deep-learning-architectures landmarks are the mirror image: surveys of
graph neural networks \citep{wu2021gnnsurvey, zhou2020gnnreview} and of
Transformers for time series \citep{wen2023tstransformers}, the
long-sequence forecasting Transformer Informer \citep{zhou2021informer}, and
its refutation---the finding, reported in the abstract of
\citet{zeng2023transformerstsf}, that embarrassingly simple linear models
outperform the Transformer-based forecasters on the standard long-term
benchmarks because the permutation-invariant attention discards the temporal
ordering that forecasting depends on. The disagreement is still open within
the field, and the LLM-reprogramming approach to the same task
\citep{jin2023timellm} reopens it from a third side. % [E-39]
Mixup \citep{zhang2017mixup}, the convex-combination regulariser, is in the
set with a wrong year; the tabular foundation model \citep{hollmann2025tabpfn}
is its one peer-reviewed post-2022 architectural result. The two architectural
proposals of the window that the set contains are both unreviewed: Mamba
\citep{gu2023mamba}, whose abstract identifies the inability of earlier
subquadratic models to perform content-based reasoning as their weakness and
makes the state-space parameters input-dependent to obtain linear-time
sequence modelling, and Kolmogorov--Arnold networks \citep{liu2024kan}, which
replace fixed node activations by learnable spline functions on edges and
claim accuracy and interpretability advantages over multilayer perceptrons on
small scientific problems. Neither claim has been tested at language-model
scale in a peer-reviewed venue within the corpus.

Where architecture research demonstrably moved is efficiency. Exact attention
with IO-aware tiling \citep{dao2022flashattention}; sparse mixture of experts
at trillion-parameter scale \citep{fedus2022switch}; eight-bit matrix
multiplication with outlier decomposition \citep{dettmers2022int8} and
post-training weight quantisation to three or four bits \citep{frantar2023gptq},
with the quantisation survey in the landmark set \citep{gholami2022quantization};
speculative decoding, which samples from a small draft model and verifies in
parallel with the large one \citep{leviathan2023speculative}; and the
efficient-LLM and inference surveys \citep{wan2023efficientllm,
zhou2024inferencesurvey}. The known-set detector gives these topics the
shortest leads in the area---\KnownStateSpaceSequenceModelsLeadMonths{} months
for state-space models, \KnownMixtureOfExpertsLeadMonths{} for mixture of
experts, \KnownSpeculativeDecodingLeadMonths{} for speculative
decoding---because each was driven by one widely adopted preprint and grew
from near zero to its peak within a year. % [E-40]
Deep-learning architectures is also the one AI category without a significant
post-2022 slope change (Table~\ref{tab:ai_breaks}). Two readings are
compatible with the counts: that architecture research plateaued once the
Transformer became the default, or that papers which would have been
classified as architecture work in 2021 are classified as LLM or
foundation-model work in 2024 because their vocabulary changed. The
multi-label counts cannot separate relabelling from decline, and the test
proposed in Part~X---re-estimating the series on works whose only AI category
is architectures, with a classifier vocabulary frozen at 2022---has not been
run.

\subsection{Interpretability and the science of model internals}

Between the theory that the citation ranking rewards and the architectures
that practice adopted lies a literature that the corpus barely registers: the
attempt to explain what a trained language model computes. The research
agenda of \citet{anwar2024foundational}, a preprint, lists it among the
foundational challenges of alignment, and the known-emergence set includes
\emph{mechanistic interpretability} as one of its eleven AI concepts. The
detector's result for it is the only negative lead in the set:
\KnownMechanisticInterpretabilityTotalWorks{} matching works, a peak-growth
month that precedes the detection month, and a lead of
\KnownMechanisticInterpretabilityLeadMonths{} months. % [E-17]
Two readings are possible and both are instructive. On the one hand the
concept is small---two to three orders of magnitude fewer matching works than
language models or retrieval augmentation---so that the growth statistic on
which the detector relies is noisy and a single busy quarter sets the peak.
On the other hand the vocabulary of the field changed inside the window:
``circuits'', ``features'' and ``sparse autoencoders'' replaced ``probing''
and ``attribution'', and a regular expression written in 2026 looks for the
later vocabulary in the earlier literature. The interpretability landmarks
that the ranking does surface are the pre-LLM ones---the case for inherently
interpretable models \citep{rudin2019stopexplaining}, the conceptual critique
\citep{lipton2018mythos}, the social-science account of explanation
\citep{miller2019explanation}, the explainable-AI surveys
\citep{barredoarrieta2020xai, ali2023xaitrust}---whose question is whether a
model's decision can be explained to a person, not whether its computation
can be reverse-engineered. The corpus therefore records the demand for
explanation and is nearly silent on the supply of mechanistic understanding;
the manuscript cannot say whether that silence reflects the small size of the
sub-field, its concentration in non-indexed venues and company reports, or
the vocabulary drift just described, and it flags the measurement as
unreliable rather than the sub-field as absent.

\section{Evaluation}
\label{sec:ai-eval}

The evaluation literature of the window is a cycle of benchmark creation,
saturation and replacement that the landmark set documents from its first
turn. The SuperGLUE abstract \citep{wang2019superglue} records, already in
2019, that performance on GLUE \citep{wang2018glue} had ``surpassed the level
of non-expert humans'' and that a harder suite was needed; the same sentence
could be written about MMLU \citep{hendrycks2021mmlu} and BIG-bench
\citep{srivastava2022bigbench} by 2024. % [E-41]
Three responses coexist. Holistic evaluation across scenarios and metrics
\citep{liang2022helm} multiplies the axes; crowd-sourced pairwise human
preference in an arena \citep{chiang2024arena}, a preprint, replaces fixed
test sets by an Elo-style ranking; and LLM-as-judge \citep{zheng2023judge},
also a preprint, replaces the human by a model, with the agreement rates
reported there as its justification. The QLoRA abstract's ``99.3 per cent of
ChatGPT'' is an instance of the third: a number produced by GPT-4 judging
Vicuna-benchmark answers, whose meaning depends on the judge's biases, which
the LLM-as-judge paper itself catalogues. Against all three stand the
documented problems of contamination of test sets by training data
\citep{sainz2023contamination}, the dependence of rankings on the choice of
tasks \citep{dehghani2021lottery}, benchmark saturation and the
sociology of benchmark use \citep{ott2022benchmarks, raji2021everything}, the
case for dynamic, adversarially collected benchmarks \citep{kiela2021dynabench}
and the earlier demonstration that even a faithfully re-collected test set
shifts accuracies \citep{recht2019imagenet}. The evaluation survey of
\citet{chang2024evalsurvey} maps the space.

The corpus adds two facts. The landmark set contains four benchmark
contributions in 234---SuperGLUE, LAION-5B, MultiMedQA and the clinical agent
benchmark---so that evaluation work is under-rewarded by thirty-six-month
citation impact relative to the role the field assigns it, a known property
of citation measures for infrastructure. % [E-05]
And no landmark abstract reports a contamination-controlled evaluation:
the models that the abstracts compare are compared on benchmarks whose
overlap with their training data is not stated. For a mapping study this is
the central methodological caveat of the whole area. Every ``state of the art''
claim read in an abstract of this part is a claim about a benchmark whose
validity for the model in question is unknown, and the manuscript's
distinction between genuine breakthroughs and incremental benchmark
improvements (RQ7) cannot be drawn from abstracts alone; Part~IX draws it, to
the extent that it can be drawn, from the citation and adoption structure
instead.

\subsection{Openness and the reproducibility of claims}

RQ9 asks how open-source availability and evaluation standards evolved, and
the landmark abstracts speak to the first half directly. The LLaMA abstract
ends with the sentence that the models are released to the research
community; LAION-5B's purpose, by its title, is an open dataset for training
image--text models; Mamba, KAN, QLoRA, LLaMA-Adapter and the open code models
\citep{touvron2023llama, schuhmann2022laion5b, gu2023mamba, liu2024kan,
dettmers2023qlora, zhang2023llamaadapter, roziere2023codellama,
li2023starcoder} are weights-and-code releases whose adoption is measurable
in the short detector leads of Section~\ref{sec:ai-theory}; and the
DeepSeek-R1 paper \citep{deepseek2025r1nature} is the first peer-reviewed
frontier-class reasoning model with released weights. Against this stands
the closed lineage \citep{openai2023gpt4, gemini2023, openai2024o1}, whose
reports omit data, scale and training details and whose evaluations can
therefore not be reproduced by anyone but the vendor, and the arena and
judge-based evaluations of Section~\ref{sec:ai-eval}, which rank closed models
on prompts that are themselves not released. The AI Index reports
\citep{maslej2024aiindex, maslej2025aiindex} document, as secondary sources,
the growing share of notable models released by industry; the corpus
measures the same thing at the level of affiliations in Part~VIII\@.
Three consequences follow for the status of the claims read in this part.
First, every comparison between an open model and a closed one in an abstract
(QLoRA against ChatGPT, LLaMA against GPT-3) is a comparison between a
reproducible artefact and a moving target, and its date is part of its
meaning. Second, the reproducibility of a result in this area depends less on
whether the paper was peer reviewed than on whether the weights, data and
evaluation prompts were released; several peer-reviewed landmarks of
Section~\ref{sec:ai-fm} train on institutional clinical data that cannot be
shared, while several preprints release everything. Third, the
distinction between peer-reviewed and preprint evidence that this manuscript
maintains throughout is therefore necessary but not sufficient, and Part~IX's
openness indicators, which are computed on the whole corpus rather than on
the landmark set, are the appropriate measure of the trend.

\section{Safety, alignment, fairness and societal evidence}
\label{sec:ai-safety}

\subsection{Risk framing}

The safety literature entered the window with a concrete-problems agenda
\citep{amodei2016concrete, hendrycks2021unsolved} and a critique of scale
\citep{bender2021parrots}, and left it with a catalogue. The AI Risk
Repository, published in \emph{Patterns} and in the landmark set, is
described in its abstract as a living database of 1,725 risks extracted from
74 taxonomies and organised by a causal taxonomy of origin and a domain
taxonomy \citep{slattery2026riskrepository}; the International AI Safety
Report \citep{bengio2025safetyreport} and the \emph{Science} perspective on
managing extreme risks \citep{bengio2024extremerisks} are the consensus
documents, and the alignment survey of \citet{ji2023alignment} and the
foundational-challenges agenda of \citet{anwar2024foundational}, both
preprints, the research agendas. % [E-42]
Catastrophic-risk framing \citep{hendrycks2023catastrophic,
shevlane2023extremerisks}, both preprints, and the ``stochastic parrots''
critique disagree about whether scale produces dangerous capabilities at all;
the corpus shows the AI-safety category re-orienting towards LLM work, which
it cited at a negligible rate in 2019 and at a tenth of its references in
2025 (Table~\ref{tab:ai_flows}), but no landmark abstract reports a validated
measurement of a catastrophic capability, so the disagreement is recorded, not
resolved. % [E-14]
The sociology of the category's citation top is the same as that of ML
theory: it is the pre-LLM ``responsible AI'' literature---fairness
\citep{mehrabi2021biasfairness}, explainable AI \citep{barredoarrieta2020xai,
ali2023xaitrust}, explanation as the social sciences understand it
\citep{miller2019explanation}---rather than the alignment literature of the
window, which is young and mostly unreviewed.

\subsection{Empirical harms with peer review}

Where the landmark set is strongest is in peer-reviewed empirical harm. The
data-poisoning threat assessment in \emph{Nature Medicine} reports, in its
abstract, that replacing 0.001 per cent of the tokens of a common pretraining
corpus with medical misinformation produces models more likely to propagate
medical errors \citep{alber2025poisoning}. A physician-led red-teaming of four
public chatbots on 222 patient-posed primary-care questions found
statistically significant differences in the safety of their answers
\citep{draelos2026unsafeanswers}. Experiments on current models as
replacement therapists report expressions of stigma and inappropriate
responses contrary to clinical guidelines \citep{moore2025stigma}. % [E-43]
These three, together with the jailbreak results of Section~\ref{sec:ai-align},
are what the corpus can offer as established, replicable harm; the
over-reliance review \citep{zhai2024overreliance}, the healthcare bias
review \citep{hasanzadeh2025biashealthcare} and the governance frameworks
\citep{papagiannidis2025raigovernance} are syntheses and proposals. Energy is
the one harm with a measurement standard: the inference-cost comparison in
the landmark set reports, in its abstract, the first systematic comparison of
task-specific and general-purpose models and finds the latter more expensive
by a wide margin \citep{luccioni2024powerhungry}, in line with the
electricity-footprint estimates of \citet{devries2023energy}. % [E-44]

\subsection{Labour, adoption and education}

The social-science records of the set carry the window's evidence on who uses
the models and to what effect. Occupational exposure to language models is
estimated from task decompositions in \citet{eloundou2024gpts}, published in
\emph{Science}, whose landmark record, as noted, carries the wrong abstract;
the preregistered field experiment with 758 consultants, whose
\emph{Organization Science} version is in the set, reports that assistance
improved performance on some tasks and worsened it on others of similar
apparent difficulty, the ``jagged frontier'' \citep{dellacqua2023jagged}; and
the adoption surveys report use by nearly half of United States adults within
two years of the first product \citep{bick2026rapidadoption}. % [E-45]
In education, the category where the ChatGPT commentary concentrated, the
abstracts move from perception surveys \citep{chan2023students,
kasneci2023chatgptgood} to policy analysis \citep{mcdonald2025hepolicies,
chan2023policyframework} and to one preregistered experiment reporting harm
to unassisted performance \citep{bastani2024harmlearning}, a preprint. The
direction of the effect of generative-AI use on learning is thus contested
between survey evidence of perceived benefit and experimental evidence of
harm, and the manuscript lists it among the open problems rather than taking
a side.

\section{Synthesis: established, contested, speculative}
\label{sec:ai-synthesis}

Four reconciliations summarise what this part can say and what it cannot.

\emph{Growth versus discontinuity.} The nine AI categories accelerated after
2022; the acceleration is a slope change that the data-driven procedures date
to the middle of 2022 or later and that steepened in 2023--2024, not a
level jump in December 2022 (Tables~\ref{tab:ai_breaks}
and~\ref{tab:ai_altdates}). The design cannot attribute it to one product, and
sentences of the form ``ChatGPT caused a surge in AI research'' are not
supported by anything in this manuscript. What is supported is weaker and
more useful: the trend of the LLM literature changed, the change propagated
through citation to generative AI, safety and agents within three years, and
it did not propagate to reinforcement learning, theory or architectures.
% [E-09--E-16]

\emph{Impact versus invention.} The citation-ranked landmarks of the LLM and
generative-AI categories are commentary and education papers; the technical
lineage is reached through arXiv records that the bibliographic linkage
sometimes attached to the wrong work. A list of the most influential AI papers
of the window built from this corpus without the screening of
Section~\ref{sec:ai-landmarks} would be wrong in about one record in five, and
would rank the first wave of ChatGPT commentary above the papers that
introduced the models. % [E-02, E-06]

\emph{Taxonomy versus field.} The apparent stagnation of reinforcement
learning and of deep-learning architectures is at least partly a consequence
of classification: the methods of the first migrated into LLM post-training,
and the vocabulary of the second migrated into the LLM and foundation-model
categories. % [E-37, P-08]

\emph{Evidence tiers.} Twenty-eight of the 234 landmark candidates are
preprints and seventy-two are dated 2026 with twelve-month citation
windows;\footnote{Counts of the columns \texttt{type} and \texttt{year} of
\texttt{literature\_synthesis/landmarks/ai\_ml\_landmarks.csv}, reproduced row by
row in Table~\ref{tab:ai_landmarks}; no macro exists for them.}
many of the window's defining technical papers---LLaMA, BLIP-2, Mamba, the
agent benchmarks, the emergent-abilities debate, the catastrophic-risk
agenda---have not passed peer review. The established results of this part
are therefore of two kinds: corpus statistics, which are reproducible from
the files named in the tables, and peer-reviewed empirical findings, of which
the landmark set contains a modest number, concentrated in medicine
(foundation models, poisoning, unsafe answers), science (MatterGen, the
Virtual Lab, Robin, DreamerV3) and the social sciences (adoption, the jagged
frontier). The contested results---emergence, sparks, pure-RL reasoning,
linear versus Transformer forecasting, fusion versus text-conditioning,
agentic AI as a category, the sign of the learning effect---are contested in
the literature itself. The speculative ones---domain foundation models
displacing pipelines, autonomous science at scale, trust frameworks for
multi-agent systems---are speculative because no landmark abstract tests
them.

The open problems that this reading leaves are collected in
\texttt{literature\_synthesis/agent\_E/open\_problems.csv} and taken up in
Part~X: whether capabilities are discontinuous in scale; whether pure-RL
reasoning generalises beyond verifiable rewards; whether alignment brittleness
is intrinsic; how to measure hallucination without references; whether
vision-language models fuse or condition; what ``agentic'' denotes; whether
autonomous science agents pay; whether architecture research plateaued or was
relabelled; how to evaluate on benchmarks that may be in the training data;
whether catastrophic capabilities can be measured before deployment; whether
domain foundation models generalise across sites; whether Transformers suit
long-horizon time series; whether generative-AI use helps or harms learning;
and, for this manuscript's own method, whether citation-based landmark
identification can be made reliable for a field whose most-cited identifiers
are the ones most often attached to the wrong record.

\subsection*{Supplementary table: the screened landmark candidates}
\label{sec:ai-supp}

\input{tables/tab_ai_landmarks}