SPB Git
spboucher.ai

spb/wp44_uqo

Public

WP44: semantic k-NN edges (5.67 M vectors, 56.7 M edges), figure layout fixes

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Simon-Pierre Boucher committed (Nov 4, 2026) parent 7446797 Browse files

49 changed files +1,392 −7

49 changed files
Unified Split
added data/metadata/semantic_edges_allenai__specter2_base.json +9 −0
@@ -0,0 +1,9 @@
1 +{
2 + "model": "allenai/specter2_base",
3 + "k": 10,
4 + "min_sim": 0.8,
5 + "n_vectors": 5668933,
6 + "n_edges": 56703396,
7 + "index": "HNSW32 efC80 efS64 inner-product",
8 + "seconds": 523
9 +}
\ No newline at end of file
modified figures/fig_area_flow_matrix.pdf +0 −0

Binary file not shown.

modified figures/fig_area_flow_matrix.png +0 −0

Binary file not shown.

modified figures/fig_area_shares.pdf +0 −0

Binary file not shown.

modified figures/fig_breaks_by_category.pdf +0 −0

Binary file not shown.

modified figures/fig_category_growth.pdf +0 −0

Binary file not shown.

modified figures/fig_category_its_panels.pdf +0 −0

Binary file not shown.

modified figures/fig_citation_inequality.pdf +0 −0

Binary file not shown.

modified figures/fig_cited_age.pdf +0 −0

Binary file not shown.

modified figures/fig_collaboration.pdf +0 −0

Binary file not shown.

modified figures/fig_country_shares.pdf +0 −0

Binary file not shown.

modified figures/fig_cross_area.pdf +0 −0

Binary file not shown.

modified figures/fig_cross_area.png +0 −0

Binary file not shown.

modified figures/fig_emerging_metrics.pdf +0 −0

Binary file not shown.

modified figures/fig_field_inflows.pdf +0 −0

Binary file not shown.

modified figures/fig_its_total.pdf +0 −0

Binary file not shown.

modified figures/fig_known_set.pdf +0 −0

Binary file not shown.

modified figures/fig_llm_cooccurrence.pdf +0 −0

Binary file not shown.

modified figures/fig_placebo_total.pdf +0 −0

Binary file not shown.

modified figures/fig_prisma_flow.pdf +0 −0

Binary file not shown.

modified figures/fig_types_openness.pdf +0 −0

Binary file not shown.

modified figures/fig_venue_concentration.pdf +0 −0

Binary file not shown.

modified figures/fig_volume_monthly_raw.pdf +0 −0

Binary file not shown.

modified figures/fig_volume_yearly.pdf +0 −0

Binary file not shown.

modified figures/make_figures.py +6 −6
@@ -271,7 +271,7 @@ def fig_flow_matrix(R, out):
271 271 p = f"{R}/results/graph/area_flows_yearly.csv"
272 272 if not os.path.exists(p):
273 273 return
274 − d = pd.read_csv(p); fig, axes = plt.subplots(1, 2, figsize=(7.4, 3.6))
274 + d = pd.read_csv(p); fig, axes = plt.subplots(1, 2, figsize=(7.6, 4.6))
275 275 for ax, (lo, hi), title in [(axes[0], (2018, 2022), "Baseline 2018–2022"), (axes[1], (2023, 2026), "Post 2023–2026")]:
276 276 g = d[(d.citing_year >= lo) & (d.citing_year <= hi)].groupby(["citing_area", "cited_area"]).n.sum().unstack(fill_value=0)
277 277 g = g.reindex(index=sorted(set(g.index) | set(g.columns)), columns=sorted(set(g.index) | set(g.columns)), fill_value=0); rs = g.div(g.sum(1), axis=0) * 100
@@ -281,7 +281,7 @@ def fig_flow_matrix(R, out):
281 281 for j in range(len(rs.columns)):
282 282 v = rs.values[i, j]
283 283 if v >= 5: ax.text(j, i, f"{v:.0f}", ha="center", va="center", fontsize=5.5, color="white" if v > 0.6 * rs.values.max() else style.INK)
284 − fig.text(0.5, -0.02, "Row shares (%): references from the citing area (rows) to the cited area (columns), internal citations only", ha="center", fontsize=7, color=style.INK2)
284 + fig.subplots_adjust(wspace=0.55, bottom=0.3); fig.text(0.5, 0.01, "Row shares (%): references from the citing area (rows) to the cited area (columns), internal citations only", ha="center", fontsize=7, color=style.INK2)
285 285 style.save(fig, f"{out}/fig_area_flow_matrix.pdf"); reg("fig_area_flow_matrix", "Citation-flow matrices between areas (row shares) for the baseline and post periods.", [p])
286 286
287 287
@@ -289,14 +289,14 @@ def fig_cross_area(R, out):
289 289 p = f"{R}/results/graph/cross_area_share_yearly.csv"; q = f"{R}/results/graph/modularity_yearly.csv"
290 290 if not os.path.exists(p):
291 291 return
292 − d = pd.read_csv(p); fig, axes = plt.subplots(1, 2, figsize=(7.4, 3.2))
292 + d = pd.read_csv(p); fig, axes = plt.subplots(1, 2, figsize=(7.6, 3.4))
293 293 for i, a in enumerate(sorted(d.citing_area.unique())):
294 294 g = d[d.citing_area == a]; axes[0].plot(g.citing_year, 100 * g.share_cross, marker="o", ms=3, color=style.CATEGORICAL[i % 8], label=a, lw=1.1)
295 − axes[0].set_title("Share of references leaving the citing area (%)", loc="left"); axes[0].legend(fontsize=6, ncol=2)
295 + axes[0].set_title("References leaving the citing area (%)", loc="left", fontsize=8); axes[0].legend(fontsize=5.5, ncol=2)
296 296 if os.path.exists(q):
297 297 m = pd.read_csv(q); axes[1].plot(m.year, m.Q_areas, marker="o", color=style.CATEGORICAL[0], label="Modularity Q (areas)"); axes[1].plot(m.year, m.Q_categories, marker="o", color=style.CATEGORICAL[1], label="Modularity Q (categories)"); axes[1].plot(m.year, m.self_share, marker="o", color=style.CATEGORICAL[2], label="Within-category share")
298 − axes[1].set_title("Modularity of the yearly category citation network", loc="left"); axes[1].legend(fontsize=6)
299 − style.save(fig, f"{out}/fig_cross_area.pdf"); reg("fig_cross_area", "Cross-area reference shares by citing area and modularity of the category citation network by year.", [p, q])
298 + axes[1].set_title("Modularity of the category citation network", loc="left", fontsize=8); axes[1].legend(fontsize=6)
299 + fig.subplots_adjust(wspace=0.3); style.save(fig, f"{out}/fig_cross_area.pdf"); reg("fig_cross_area", "Cross-area reference shares by citing area and modularity of the category citation network by year.", [p, q])
300 300
301 301
302 302 def fig_emerging(R, out):
modified manuscript/figures/fig_area_flow_matrix.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_area_shares.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_breaks_by_category.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_category_growth.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_category_its_panels.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_citation_inequality.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_cited_age.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_collaboration.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_country_shares.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_cross_area.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_emerging_metrics.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_field_inflows.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_its_total.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_known_set.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_llm_cooccurrence.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_placebo_total.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_prisma_flow.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_types_openness.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_venue_concentration.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_volume_monthly_raw.pdf +0 −0

Binary file not shown.

modified manuscript/figures/fig_volume_yearly.pdf +0 −0

Binary file not shown.

modified manuscript/sections/part05_ai.tex +394 −0
@@ -568,3 +568,397 @@ intellectual map would conclude that generative AI is a topic in education
568 568 research; the correct conclusion is that field normalisation inside a
569 569 multi-label category measures discourse, and that the measurement is
570 570 informative about diffusion rather than about invention.
571 +
572 +\section{AI agents and autonomous systems}
573 +\label{sec:ai-agents}
574 +
575 +\subsection{From tool use to agents}
576 +
577 +An LLM-based agent, in the definition shared by the surveys of the window
578 +\citep{wang2024agentsurvey, xi2023rise, guo2024multiagent,
579 +cheng2026agentsdefinitions}, is a language model wrapped in a loop of
580 +perception, planning, memory and action, in which the model's outputs are
581 +parsed into calls to tools or to other agents and the results are fed back as
582 +context. The architectures that fixed the vocabulary are generative agents in
583 +a simulated town \citep{park2023generativeagents}, a peer-reviewed UIST paper;
584 +the multi-agent conversation framework AutoGen \citep{wu2023autogen}; virtual
585 +context management in MemGPT \citep{packer2023memgpt}; and the ``LLM as
586 +operating system'' proposals \citep{mei2024aios, ge2023llmos}, all of the
587 +latter preprints. The memory component is surveyed by
588 +\citet{zhang2025agentmemory}, whose abstract calls self-evolution through
589 +agent--environment interaction the feature that distinguishes agents from the
590 +models they wrap. Evaluation lags design. The benchmarks that the community
591 +uses---SWE-bench for repository-level issue resolution
592 +\citep{jimenez2024swebench}, WebArena for web tasks \citep{zhou2023webarena},
593 +AgentBench across environments \citep{liu2023agentbench}---are preprints, and
594 +the only peer-reviewed agent benchmark in the landmark set is a clinical
595 +one, an evaluation of an open-source and a proprietary agent system on three
596 +families of medical decision tasks \citep{liu2026clinicalagents}. % [E-32]
597 +The detector flags \emph{LLM agents} in \KnownLlmAgentsDetectionMonth{},
598 +\KnownLlmAgentsLeadMonths{} months before peak growth in
599 +\KnownLlmAgentsPeakGrowthMonth{}, on \KnownLlmAgentsTotalWorks{} matching
600 +works; the concept arrived and peaked within a year.
601 +
602 +The strongest peer-reviewed evidence for agents is scientific. The Virtual Lab
603 +\citep{swanson2025virtuallab}, a Nature paper whose record carries no abstract
604 +in the corpus file, reports agents designing SARS-CoV-2 nanobodies that were
605 +validated experimentally; Robin \citep{ghareeb2026robin} is described in its
606 +abstract as the first multi-agent system automating both hypothesis generation
607 +and data analysis in experimental biology; DIVE \citep{zhang2026dive} extracts
608 +experimental data from the figures of the literature for hydrogen-storage
609 +materials; AgentSociety \citep{piao2026agentsociety} simulates large
610 +populations of LLM-driven generative agents for social science. % [E-33]
611 +These are single demonstrations with experimental or observational
612 +validation, which is more than the ``AI scientist'' preprints
613 +\citep{lu2024aiscientist, si2024novelideas} offer; whether autonomous research
614 +agents produce validated discoveries at a rate that justifies their cost is a
615 +question for preregistered multi-laboratory trials, not for case reports, and
616 +it is listed in Part~X.
617 +
618 +\subsection{``Agentic AI'': a definitional literature}
619 +
620 +The landmark set of the category contains six papers from 2025--2026 whose
621 +purpose is to define agentic AI and distinguish it from AI agents, generative
622 +AI and classical multi-agent systems \citep{acharya2025agenticsurvey,
623 +sapkota2026agentsvsagentic, bandi2025agenticreview, abouali2025agenticparadigms,
624 +allmendinger2026multiagentai, collaco2026agentichealthcare}, most of them
625 +reviews of reviews, one of them a PRISMA-based review of ninety studies that
626 +proposes a symbolic-versus-neural dual-paradigm framework, and one benchmark.
627 +% [E-34]
628 +A literature that produces six taxonomies and one benchmark in eighteen
629 +months is a literature whose object is not yet stabilised, and the question
630 +of whether ``agentic AI'' denotes a distinct class of systems or a relabelling
631 +of tool-augmented LLM pipelines is treated here as open. The trust, risk and
632 +security management framework for multi-agent LLM systems
633 +\citep{raza2026trism} belongs to the same genre: a structure without an
634 +empirical test.
635 +
636 +The corpus series of the category must be read with its composition in mind.
637 +Four of its landmark records are control-theoretic papers on consensus and
638 +formation control of nonlinear multi-agent systems, and its pre-2022 volume
639 +is dominated by classical multi-agent systems, distributed artificial
640 +intelligence and robotics, which explains why the category directs a large
641 +and growing share of its references to robotics (Table~\ref{tab:ai_flows}).
642 +The negative baseline growth rate, the negative level change at the reference
643 +month and the second Bai--Perron break of 2024 (Table~\ref{tab:ai_breaks})
644 +are the signature of a category in which an older literature was flat or
645 +declining while a new one, LLM agents, grew from a small base; the level
646 +effect at the o1 date, \ItsSeriesCatAiAgentsItsOOneLevelChange{}
647 +($p = \ItsSeriesCatAiAgentsItsOOneLevelP$), is where the new literature
648 +becomes visible in the aggregate. % [E-35]
649 +The explanation is inferential; separating the two literatures would require a
650 +sub-classification that the taxonomy does not provide. Embodied agents, from
651 +vision-language-action models \citep{brohan2023rt2, kim2024openvla,
652 +black2024pi0} to the driving model in the landmark set, are classified under
653 +robotics and are reviewed there.
654 +
655 +\section{Reinforcement learning}
656 +\label{sec:ai-rl}
657 +
658 +The reinforcement-learning landmarks are, almost without exception, the canon
659 +of the baseline period. Rainbow's combination of six extensions to deep
660 +Q-learning with an ablation of each \citep{hessel2018rainbow}; the
661 +maximum-entropy actor--critic \citep{haarnoja2018sac}, whose record is the
662 +arXiv deposit; clipped double-Q learning and delayed policy updates against
663 +overestimation \citep{fujimoto2018td3}; self-play without domain knowledge
664 +beyond the rules in AlphaZero \citep{silver2018alphazero}; grandmaster-level
665 +StarCraft~II \citep{vinyals2019alphastar}, a Nature record without an abstract
666 +in the file; the finding that proximal policy optimisation is, in the words of
667 +its abstract, surprisingly strong in cooperative multi-agent benchmarks
668 +\citep{yu2021mappo}; the large-scale study of curiosity-driven learning without
669 +extrinsic rewards \citep{burda2018curiosity}; and the reproducibility warning
670 +of \citet{henderson2018drlmatters}, whose abstract notes that non-determinism
671 +and intrinsic variance make reported results hard to interpret. The theoretical
672 +side is represented by an overview of multi-agent reinforcement learning
673 +\citep{zhang2021marloverview} and a survey of the synthesis of model-predictive
674 +control and reinforcement learning (record 7128436456 of the landmark file).
675 +The single post-2022 method landmark is DreamerV3 \citep{hafner2025dreamerv3},
676 +published in Nature, whose abstract reports a world-model algorithm that
677 +outperforms specialised methods across more than 150 tasks with a single
678 +configuration. % [E-36]
679 +Several records are misdated---a 2008 inverse-reinforcement-learning paper
680 +and a 2007 Bayesian one appear as 2018 and 2025 works---and four are
681 +control-theory papers; the category's citation top is therefore a mixture of
682 +classics, control theory and one new result.
683 +
684 +The corpus places reinforcement learning among the modestly accelerating
685 +categories: a post-period growth rate only slightly above the baseline
686 +(Table~\ref{tab:ai_growth}), a slope change at the reference month of
687 +\ItsSeriesCatReinforcementLearningItsChatgptSlopeChange{} log points per
688 +month after a negative level change, and a citation profile that moves towards
689 +robotics. % [E-15]
690 +It would be a mistake to read this as the field's eclipse. Reinforcement
691 +learning's most consequential role in the window---reward modelling and policy
692 +optimisation for language models, from PPO through DPO to the group-relative
693 +method of DeepSeek-R1---is invisible in the category because the papers that
694 +carry it are classified as LLM work, and the flow of references from the RL
695 +category to the LLM category stays below one per cent. Any statement that
696 +reinforcement learning declined after 2022 would be an artefact of the
697 +taxonomy, and the correct statement is that its methods migrated into another
698 +category's literature. % [E-37]
699 +
700 +\section{Machine-learning theory, architectures and efficiency}
701 +\label{sec:ai-theory}
702 +
703 +\subsection{What ``machine-learning theory'' contains in the corpus}
704 +
705 +The category named machine-learning theory does not, in its citation-ranked
706 +top, contain a learning-theory result from the window. Its landmarks are
707 +textbooks and textbook chapters, the case for the Matthews correlation
708 +coefficient over accuracy and F$_1$ on imbalanced binary problems
709 +\citep{chicco2020mcc}, the argument that high-stakes decisions should use
710 +inherently interpretable models rather than post-hoc explanations of black
711 +boxes \citep{rudin2019stopexplaining}, the conceptual analysis of
712 +interpretability \citep{lipton2018mythos}, exact Shapley attributions for tree
713 +ensembles \citep{lundberg2020treeexplainer}, the review of stochastic-gradient
714 +optimisation for large-scale learning \citep{bottou2018optimization}, a
715 +transfer-learning survey \citep{zhuang2021transfersurvey}, a reprinted chapter
716 +by Verma and Pearl on the equivalence of causal models, and the misdated
717 +bandit papers of Section~\ref{sec:ai-landmarks}. % [E-38]
718 +The category as the rules define it is general machine learning, and what the
719 +citation ranking rewards within it is pedagogy and measurement convention.
720 +Theoretical work on in-context learning, on the expressivity of Transformers
721 +or on the statistical basis of scaling laws exists inside the window, but it is
722 +classified under LLM or deep-learning architectures, or is cited too little
723 +within thirty-six months to surface. The corpus therefore cannot be used to
724 +say that theory lagged practice in the AI transformation; it can only say that
725 +the theory which is cited quickly is the theory of evaluation and
726 +interpretation.
727 +
728 +\subsection{Architectures: proposals, refutations and the efficiency turn}
729 +
730 +The deep-learning-architectures landmarks are the mirror image: surveys of
731 +graph neural networks \citep{wu2021gnnsurvey, zhou2020gnnreview} and of
732 +Transformers for time series \citep{wen2023tstransformers}, the
733 +long-sequence forecasting Transformer Informer \citep{zhou2021informer}, and
734 +its refutation---the finding, reported in the abstract of
735 +\citet{zeng2023transformerstsf}, that embarrassingly simple linear models
736 +outperform the Transformer-based forecasters on the standard long-term
737 +benchmarks because the permutation-invariant attention discards the temporal
738 +ordering that forecasting depends on. The disagreement is still open within
739 +the field, and the LLM-reprogramming approach to the same task
740 +\citep{jin2023timellm} reopens it from a third side. % [E-39]
741 +Mixup \citep{zhang2017mixup}, the convex-combination regulariser, is in the
742 +set with a wrong year; the tabular foundation model \citep{hollmann2025tabpfn}
743 +is its one peer-reviewed post-2022 architectural result. The two architectural
744 +proposals of the window that the set contains are both unreviewed: Mamba
745 +\citep{gu2023mamba}, whose abstract identifies the inability of earlier
746 +subquadratic models to perform content-based reasoning as their weakness and
747 +makes the state-space parameters input-dependent to obtain linear-time
748 +sequence modelling, and Kolmogorov--Arnold networks \citep{liu2024kan}, which
749 +replace fixed node activations by learnable spline functions on edges and
750 +claim accuracy and interpretability advantages over multilayer perceptrons on
751 +small scientific problems. Neither claim has been tested at language-model
752 +scale in a peer-reviewed venue within the corpus.
753 +
754 +Where architecture research demonstrably moved is efficiency. Exact attention
755 +with IO-aware tiling \citep{dao2022flashattention}; sparse mixture of experts
756 +at trillion-parameter scale \citep{fedus2022switch}; eight-bit matrix
757 +multiplication with outlier decomposition \citep{dettmers2022int8} and
758 +post-training weight quantisation to three or four bits \citep{frantar2023gptq},
759 +with the quantisation survey in the landmark set \citep{gholami2022quantization};
760 +speculative decoding, which samples from a small draft model and verifies in
761 +parallel with the large one \citep{leviathan2023speculative}; and the
762 +efficient-LLM and inference surveys \citep{wan2023efficientllm,
763 +zhou2024inferencesurvey}. The known-set detector gives these topics the
764 +shortest leads in the area---\KnownStateSpaceSequenceModelsLeadMonths{} months
765 +for state-space models, \KnownMixtureOfExpertsLeadMonths{} for mixture of
766 +experts, \KnownSpeculativeDecodingLeadMonths{} for speculative
767 +decoding---because each was driven by one widely adopted preprint and grew
768 +from near zero to its peak within a year. % [E-40]
769 +Deep-learning architectures is also the one AI category without a significant
770 +post-2022 slope change (Table~\ref{tab:ai_breaks}). Two readings are
771 +compatible with the counts: that architecture research plateaued once the
772 +Transformer became the default, or that papers which would have been
773 +classified as architecture work in 2021 are classified as LLM or
774 +foundation-model work in 2024 because their vocabulary changed. The
775 +multi-label counts cannot separate relabelling from decline, and the test
776 +proposed in Part~X---re-estimating the series on works whose only AI category
777 +is architectures, with a classifier vocabulary frozen at 2022---has not been
778 +run.
779 +
780 +\section{Evaluation}
781 +\label{sec:ai-eval}
782 +
783 +The evaluation literature of the window is a cycle of benchmark creation,
784 +saturation and replacement that the landmark set documents from its first
785 +turn. The SuperGLUE abstract \citep{wang2019superglue} records, already in
786 +2019, that performance on GLUE \citep{wang2018glue} had ``surpassed the level
787 +of non-expert humans'' and that a harder suite was needed; the same sentence
788 +could be written about MMLU \citep{hendrycks2021mmlu} and BIG-bench
789 +\citep{srivastava2022bigbench} by 2024. % [E-41]
790 +Three responses coexist. Holistic evaluation across scenarios and metrics
791 +\citep{liang2022helm} multiplies the axes; crowd-sourced pairwise human
792 +preference in an arena \citep{chiang2024arena}, a preprint, replaces fixed
793 +test sets by an Elo-style ranking; and LLM-as-judge \citep{zheng2023judge},
794 +also a preprint, replaces the human by a model, with the agreement rates
795 +reported there as its justification. The QLoRA abstract's ``99.3 per cent of
796 +ChatGPT'' is an instance of the third: a number produced by GPT-4 judging
797 +Vicuna-benchmark answers, whose meaning depends on the judge's biases, which
798 +the LLM-as-judge paper itself catalogues. Against all three stand the
799 +documented problems of contamination of test sets by training data
800 +\citep{sainz2023contamination}, the dependence of rankings on the choice of
801 +tasks \citep{dehghani2021lottery}, benchmark saturation and the
802 +sociology of benchmark use \citep{ott2022benchmarks, raji2021everything}, the
803 +case for dynamic, adversarially collected benchmarks \citep{kiela2021dynabench}
804 +and the earlier demonstration that even a faithfully re-collected test set
805 +shifts accuracies \citep{recht2019imagenet}. The evaluation survey of
806 +\citet{chang2024evalsurvey} maps the space.
807 +
808 +The corpus adds two facts. The landmark set contains four benchmark
809 +contributions in 234---SuperGLUE, LAION-5B, MultiMedQA and the clinical agent
810 +benchmark---so that evaluation work is under-rewarded by thirty-six-month
811 +citation impact relative to the role the field assigns it, a known property
812 +of citation measures for infrastructure. % [E-05]
813 +And no landmark abstract reports a contamination-controlled evaluation:
814 +the models that the abstracts compare are compared on benchmarks whose
815 +overlap with their training data is not stated. For a mapping study this is
816 +the central methodological caveat of the whole area. Every ``state of the art''
817 +claim read in an abstract of this part is a claim about a benchmark whose
818 +validity for the model in question is unknown, and the manuscript's
819 +distinction between genuine breakthroughs and incremental benchmark
820 +improvements (RQ7) cannot be drawn from abstracts alone; Part~IX draws it, to
821 +the extent that it can be drawn, from the citation and adoption structure
822 +instead.
823 +
824 +\section{Safety, alignment, fairness and societal evidence}
825 +\label{sec:ai-safety}
826 +
827 +\subsection{Risk framing}
828 +
829 +The safety literature entered the window with a concrete-problems agenda
830 +\citep{amodei2016concrete, hendrycks2021unsolved} and a critique of scale
831 +\citep{bender2021parrots}, and left it with a catalogue. The AI Risk
832 +Repository, published in \emph{Patterns} and in the landmark set, is
833 +described in its abstract as a living database of 1,725 risks extracted from
834 +74 taxonomies and organised by a causal taxonomy of origin and a domain
835 +taxonomy \citep{slattery2026riskrepository}; the International AI Safety
836 +Report \citep{bengio2025safetyreport} and the \emph{Science} perspective on
837 +managing extreme risks \citep{bengio2024extremerisks} are the consensus
838 +documents, and the alignment survey of \citet{ji2023alignment} and the
839 +foundational-challenges agenda of \citet{anwar2024foundational}, both
840 +preprints, the research agendas. % [E-42]
841 +Catastrophic-risk framing \citep{hendrycks2023catastrophic,
842 +shevlane2023extremerisks}, both preprints, and the ``stochastic parrots''
843 +critique disagree about whether scale produces dangerous capabilities at all;
844 +the corpus shows the AI-safety category re-orienting towards LLM work, which
845 +it cited at a negligible rate in 2019 and at a tenth of its references in
846 +2025 (Table~\ref{tab:ai_flows}), but no landmark abstract reports a validated
847 +measurement of a catastrophic capability, so the disagreement is recorded, not
848 +resolved. % [E-14]
849 +The sociology of the category's citation top is the same as that of ML
850 +theory: it is the pre-LLM ``responsible AI'' literature---fairness
851 +\citep{mehrabi2021biasfairness}, explainable AI \citep{barredoarrieta2020xai,
852 +ali2023xaitrust}, explanation as the social sciences understand it
853 +\citep{miller2019explanation}---rather than the alignment literature of the
854 +window, which is young and mostly unreviewed.
855 +
856 +\subsection{Empirical harms with peer review}
857 +
858 +Where the landmark set is strongest is in peer-reviewed empirical harm. The
859 +data-poisoning threat assessment in \emph{Nature Medicine} reports, in its
860 +abstract, that replacing 0.001 per cent of the tokens of a common pretraining
861 +corpus with medical misinformation produces models more likely to propagate
862 +medical errors \citep{alber2025poisoning}. A physician-led red-teaming of four
863 +public chatbots on 222 patient-posed primary-care questions found
864 +statistically significant differences in the safety of their answers
865 +\citep{draelos2026unsafeanswers}. Experiments on current models as
866 +replacement therapists report expressions of stigma and inappropriate
867 +responses contrary to clinical guidelines \citep{moore2025stigma}. % [E-43]
868 +These three, together with the jailbreak results of Section~\ref{sec:ai-align},
869 +are what the corpus can offer as established, replicable harm; the
870 +over-reliance review \citep{zhai2024overreliance}, the healthcare bias
871 +review \citep{hasanzadeh2025biashealthcare} and the governance frameworks
872 +\citep{papagiannidis2025raigovernance} are syntheses and proposals. Energy is
873 +the one harm with a measurement standard: the inference-cost comparison in
874 +the landmark set reports, in its abstract, the first systematic comparison of
875 +task-specific and general-purpose models and finds the latter more expensive
876 +by a wide margin \citep{luccioni2024powerhungry}, in line with the
877 +electricity-footprint estimates of \citet{devries2023energy}. % [E-44]
878 +
879 +\subsection{Labour, adoption and education}
880 +
881 +The social-science records of the set carry the window's evidence on who uses
882 +the models and to what effect. Occupational exposure to language models is
883 +estimated from task decompositions in \citet{eloundou2024gpts}, published in
884 +\emph{Science}, whose landmark record, as noted, carries the wrong abstract;
885 +the preregistered field experiment with 758 consultants, whose
886 +\emph{Organization Science} version is in the set, reports that assistance
887 +improved performance on some tasks and worsened it on others of similar
888 +apparent difficulty, the ``jagged frontier'' \citep{dellacqua2023jagged}; and
889 +the adoption surveys report use by nearly half of United States adults within
890 +two years of the first product \citep{bick2026rapidadoption}. % [E-45]
891 +In education, the category where the ChatGPT commentary concentrated, the
892 +abstracts move from perception surveys \citep{chan2023students,
893 +kasneci2023chatgptgood} to policy analysis \citep{mcdonald2025hepolicies,
894 +chan2023policyframework} and to one preregistered experiment reporting harm
895 +to unassisted performance \citep{bastani2024harmlearning}, a preprint. The
896 +direction of the effect of generative-AI use on learning is thus contested
897 +between survey evidence of perceived benefit and experimental evidence of
898 +harm, and the manuscript lists it among the open problems rather than taking
899 +a side.
900 +
901 +\section{Synthesis: established, contested, speculative}
902 +\label{sec:ai-synthesis}
903 +
904 +Four reconciliations summarise what this part can say and what it cannot.
905 +
906 +\emph{Growth versus discontinuity.} The nine AI categories accelerated after
907 +2022; the acceleration is a slope change that the data-driven procedures date
908 +to the middle of 2022 or later and that steepened in 2023--2024, not a
909 +level jump in December 2022 (Tables~\ref{tab:ai_breaks} and
910 +\ref{tab:ai_altdates}). The design cannot attribute it to one product, and
911 +sentences of the form ``ChatGPT caused a surge in AI research'' are not
912 +supported by anything in this manuscript. What is supported is weaker and
913 +more useful: the trend of the LLM literature changed, the change propagated
914 +through citation to generative AI, safety and agents within three years, and
915 +it did not propagate to reinforcement learning, theory or architectures.
916 +% [E-09--E-16]
917 +
918 +\emph{Impact versus invention.} The citation-ranked landmarks of the LLM and
919 +generative-AI categories are commentary and education papers; the technical
920 +lineage is reached through arXiv records that the bibliographic linkage
921 +sometimes attached to the wrong work. A list of the most influential AI papers
922 +of the window built from this corpus without the screening of
923 +Section~\ref{sec:ai-landmarks} would be wrong in about one record in five, and
924 +would rank the first wave of ChatGPT commentary above the papers that
925 +introduced the models. % [E-02, E-06]
926 +
927 +\emph{Taxonomy versus field.} The apparent stagnation of reinforcement
928 +learning and of deep-learning architectures is at least partly a consequence
929 +of classification: the methods of the first migrated into LLM post-training,
930 +and the vocabulary of the second migrated into the LLM and foundation-model
931 +categories. % [E-37, P-08]
932 +
933 +\emph{Evidence tiers.} Twenty-eight of the 234 landmark candidates are
934 +preprints and seventy-two are dated 2026 with twelve-month citation windows;
935 +many of the window's defining technical papers---LLaMA, BLIP-2, Mamba, the
936 +agent benchmarks, the emergent-abilities debate, the catastrophic-risk
937 +agenda---have not passed peer review. The established results of this part
938 +are therefore of two kinds: corpus statistics, which are reproducible from
939 +the files named in the tables, and peer-reviewed empirical findings, of which
940 +the landmark set contains a modest number, concentrated in medicine
941 +(foundation models, poisoning, unsafe answers), science (MatterGen, the
942 +Virtual Lab, Robin, DreamerV3) and the social sciences (adoption, the jagged
943 +frontier). The contested results---emergence, sparks, pure-RL reasoning,
944 +linear versus Transformer forecasting, fusion versus text-conditioning,
945 +agentic AI as a category, the sign of the learning effect---are contested in
946 +the literature itself. The speculative ones---domain foundation models
947 +displacing pipelines, autonomous science at scale, trust frameworks for
948 +multi-agent systems---are speculative because no landmark abstract tests
949 +them.
950 +
951 +The open problems that this reading leaves are collected in
952 +\texttt{literature\_synthesis/agent\_E/open\_problems.csv} and taken up in
953 +Part~X: whether capabilities are discontinuous in scale; whether pure-RL
954 +reasoning generalises beyond verifiable rewards; whether alignment brittleness
955 +is intrinsic; how to measure hallucination without references; whether
956 +vision-language models fuse or condition; what ``agentic'' denotes; whether
957 +autonomous science agents pay; whether architecture research plateaued or was
958 +relabelled; how to evaluate on benchmarks that may be in the training data;
959 +whether catastrophic capabilities can be measured before deployment; whether
960 +domain foundation models generalise across sites; whether Transformers suit
961 +long-horizon time series; whether generative-AI use helps or harms learning;
962 +and, for this manuscript's own method, whether citation-based landmark
963 +identification can be made reliable for a field whose most-cited identifiers
964 +are the ones most often attached to the wrong record.
modified manuscript/sections/part07a_theory.tex +622 −1
@@ -1 +1,622 @@
1 −% part07a_theory — being written by Agent G
1 +% part07a_theory.tex — Part VII (theory half): algorithms and data structures, computational complexity, programming
2 +% languages, compilers, formal verification, software engineering. Agent G (Theory Specialist), 2026-11-03.
3 +% The director merges this file with part07b_security.tex into Part VII. Every corpus number is a macro of
4 +% tables/numbers.tex (prefix \Thy..., appended by tables/make_numbers.py, Agent G block) or appears in a generated table
5 +% (tables/make_tables.py, tables_part07a). Sources: results/graph/category_counts_yearly.csv, category_flows_yearly.csv,
6 +% category_cooccurrence_yearly.csv; results/econometrics/breaks.csv; literature_synthesis/landmarks/{theory,software_pl}_*;
7 +% literature_synthesis/agent_G/{claims,contributions,open_problems}.csv and review.md. Bibliography: references.bib and
8 +% references_part07a.bib (65 entries, each resolved by one live lookup; references_verification_part07a.csv).
9 +
10 +\section{Scope, evidence and reading rules for the theory half}
11 +\label{sec:thy-scope}
12 +
13 +This half of Part~VII reviews six categories of the taxonomy: algorithms and data structures, computational
14 +complexity, programming languages, compilers, formal verification and software engineering (the areas
15 +\emph{theory} and \emph{software\_pl} of the ontology, Section~\ref{sec:p2-ontology}). It has two objects. The
16 +first is the research itself: what the baseline period established in these categories, what the window added,
17 +and where the additions are contested. The second is the instrument: the corpus statistics for the six
18 +categories and, above all, the automatically selected landmark candidates, which in these areas turn out to
19 +measure the bibliographic record more than the research it records. The two objects cannot be separated,
20 +because the landmark list is the protocol's channel from corpus to synthesis, and in the theory and software
21 +areas that channel is obstructed in ways that are themselves findings.
22 +
23 +Three reading rules apply. First, no full text was read. Every statement about a work rests on its metadata
24 +(title, venue, type, year, identifiers) or on the abstract recorded in OpenAlex, and the text says which; where
25 +an abstract is the source, the phrase ``the abstract reports'' is used deliberately. Second, works available
26 +only as arXiv postings are preprints and are described as such, whatever their later fate. Third, the
27 +structural-change statistics are read as the protocol prescribes (Section~\ref{sec:m-structural}): the November
28 +2022 boundary is a reference date, the placebo distribution is the test, and nothing below attributes a change
29 +to any single event. The categories are small relative to the AI categories of Part~V (their combined primary
30 +share in 2025 is under ten per cent of the corpus), so the statistics are those of minorities inside a
31 +corpus whose growth is set elsewhere.
32 +
33 +The section proceeds from the instrument to the field. Section~\ref{sec:thy-landmarks} audits the landmark
34 +candidates; Section~\ref{sec:thy-statistics} reports the six categories' volume, structural-change and
35 +composition statistics; Sections~\ref{sec:thy-algorithms} to~\ref{sec:thy-se} review the categories; and
36 +Section~\ref{sec:thy-synthesis} draws the conclusions that Part~X and Part~XI inherit.
37 +
38 +\section{The landmark candidates: an audit}
39 +\label{sec:thy-landmarks}
40 +
41 +The protocol selects landmarks as the works with the highest field-normalised 36-month citation score
42 +(NCS36, Section~\ref{sec:m-citations}) within each primary category, with a floor of fifteen per category, plus
43 +the 200 most cited works overall; for 2025--2026 works the 12-month window replaces the 36-month one
44 +(Section~\ref{sec:data-core}). The theory area contributed \ThyLandTheoryN{} candidates and the software area
45 +\ThyLandSoftwarePlN{}, twenty-five per category, of the \SelectionLandmarks{} landmarks of the whole corpus.
46 +Before any of the \ThyAuditN{} can be called a landmark, the list has to be audited, and the audit is the first
47 +result of this section (Table~\ref{tab:theory_landmark_audit}). Each candidate received one of eight verdicts
48 +from its metadata and, where available, its abstract: an in-scope research contribution
49 +(\ThyAuditVerdictInScopeResearch{} candidates), an in-scope survey or report (\ThyAuditVerdictInScopeSurvey{}),
50 +a misclassified work outside the category and outside the theory and software scope
51 +(\ThyAuditVerdictMisclassified{}), a thesis (\ThyAuditVerdictDissertation{}), a classic re-deposited under a
52 +recent record date (\ThyAuditVerdictReprintMisdated{}), a textbook or chapter (\ThyAuditVerdictTextbook{}), a
53 +metaphor-based metaheuristic (\ThyAuditVerdictMetaheuristic{}) or a duplicate record
54 +(\ThyAuditVerdictDuplicate{}). Fewer than one candidate in five is a research contribution inside its
55 +category, and only \ThyAuditInScopeResearchPost{} of those are post-period works. Four mechanisms account
56 +for the rest.
57 +
58 +\input{tables/tab_theory_landmark_audit}
59 +
60 +\paragraph{Identifier collisions.}
61 +\ThyAuditIdCollision{} candidates carry OpenAlex identifiers that, resolved against the live database on
62 +3~November 2026, return a different work. The record titled ``Exploiting Generative AI to Scale up Intelligent
63 +Tutoring Systems'', with the DOI of an ITP~2023 paper, resolves to \emph{Attention Is All You Need}; the abstract
64 +stored for it in the corpus is that of \emph{MizAR~60 for Mizar~50}, which is what the DOI denotes
65 +\citep{jakubuv2023mizar}; its 36-month citation count in the corpus is that of the Transformer paper, which is
66 +why it ranks first in the whole software area. The ECOOP~2025 paper on functionality-specific
67 +vulnerabilities \citep{chen2025apiss} resolves to the graph-convolutional-network preprint; the SEA~2024
68 +paper on learned branching for maximum independent set \citep{langedal2024mis} to GraphSAGE; the WADS~2025
69 +Kd-tree paper \citep{pham2025kdtree} to the variational-autoencoder preprint; the SAIA~2024 robustness
70 +framework \citep{awadid2025robust} to the adversarial-training preprint of Madry and co-authors; an
71 +ESEM~2026 record \citep{naqvi2026bigger} to the Codex paper; and a DX~2024 answer-set-programming paper
72 +\citep{klampfl2024asp} to an unrelated preprint. In every case the recorded citation window matches the other
73 +work's count.\footnote{Raw responses in \texttt{literature\_synthesis/agent\_G/verification\_lookups\_raw.txt}.
74 +The mechanism (identifier merging or reuse between the snapshot of 14~October 2026 and the live database) is
75 +inferred; the consequence, that the citation windows of these records belong to other works, is observed.} The
76 +four top-ranked candidates of the software area and the top-ranked candidate of the theory area owe their rank
77 +to citations that are not theirs. The same check should be run on the other areas' lists before Part~IX uses
78 +them.
79 +
80 +\paragraph{Generic-title chapters.}
81 +Chapters titled ``Introduction to Algorithms'' (2020, 2021, 2023), ``Particle Swarm Optimization'' (2019),
82 +``Introduction to Software Testing'' (2021), ``Introduction to Graph Theory'' (2023), and Royce's 1970 paper on
83 +managing large software developments reprinted in a 2021 anthology, carry hundreds to thousands of citations in
84 +their windows, and live lookups confirm the counts. The plausible reading, offered as an interpretation, is
85 +that citations intended for the canonical work are attributed to a same-titled chapter by the indexer's
86 +reference matching. Whatever the mechanism, a chapter of a 2023 undergraduate textbook is not a landmark of
87 +2023 research.
88 +
89 +\paragraph{Re-deposited classics.}
90 +\ThyAuditVerdictReprintMisdated{} records are works whose record date is the date of a deposit, reprint or
91 +re-posting rather than of the work. Dijkstra's 1959 note on two problems in connexion with graphs appears as a
92 +2022 ACM Books chapter; Turing's 1936 paper as a 2021 MIT Press chapter; the 1994 survey of decidability for
93 +Petri nets as a 2024 arXiv posting; Lynch and Tuttle's introduction to input/output automata as a 2026 journal
94 +article \citep{lynch2026ioautomata}; the \emph{egg} equality-saturation paper \citep{willsey2021egg} as a 2026
95 +research highlight \citep{willsey2026eggcacm}; doctoral theses of the early 2000s as 2026 repository records;
96 +three theses of the L4/seL4 verification school as 2022 records; and the 2018 conference paper on online
97 +algorithms with machine-learned predictions as a 2024 arXiv posting. Because the year field is the deposit year,
98 +these records enter the post period: \ThyLandTheoryDatedTwentysix{} of the \ThyLandTheoryN{} theory candidates
99 +and \ThyLandSoftwarePlDatedTwentysix{} of the \ThyLandSoftwarePlN{} software candidates are dated 2026, a year
100 +of which the corpus holds ten months.
101 +
102 +\paragraph{Lexical misclassification and small cells.}
103 +\ThyAuditVerdictMisclassified{} candidates are outside the theory and software scope altogether. The
104 +computational-complexity list contains \ThyAuditComputationalComplexityMisclassified{} such works of its
105 +\ThyAuditComputationalComplexityN{}: technology-acceptance models of mobile learning, PLS-SEM studies of
106 +small enterprises and of students' well-being, a Routledge volume on complexity theory in the social sciences,
107 +a review of fringe-projection profilometry, captured by the word ``complexity''. The compilers list contains
108 +\ThyAuditCompilersMisclassified{}, captured by ``just-in-time'' (adaptive interventions in mental health),
109 +``compiling'' (a guide to literature reviews) and ``polyhedral'' (power-system dispatch). The
110 +programming-languages list includes a Nature paper on Calabi--Yau manifolds in black-hole scattering. The
111 +classifier's per-category precision, estimated on the validation sample (Section~\ref{sec:m-validation}), is
112 +the right place to quantify this; what the landmark list shows is the interaction between imprecision and
113 +normalisation. NCS36 divides a work's citations by the expectation for its category, year and type
114 +(Section~\ref{sec:m-citations}); cells of theses or of non-computing articles are small and lightly cited, so a
115 +thesis with a handful of citations, or a social-science article with a hundred, obtains a normalised score that
116 +no STOC paper of the same year can match. \ThyAuditVerdictDissertation{} of the \ThyAuditN{} candidates are
117 +theses.
118 +
119 +What survives is a set of \ThyAuditVerdictInScopeResearch{} research works and \ThyAuditVerdictInScopeSurvey{}
120 +surveys or reports, listed in Table~\ref{tab:theory_landmarks_discussed}. They are reviewed below with the
121 +pre-2022 lineage of Part~II and with the window's results that the ranking does not see: the major theoretical
122 +results of 2022--2025 are absent from the candidate list and enter this review through verified bibliographic
123 +records, not through the corpus ranking. Open problem~G-P1 of Part~X states the correction: type filters,
124 +live re-resolution of identifiers and venue priors at the selection stage, evaluated against a
125 +specialist-labelled sample.
126 +
127 +\input{tables/tab_theory_landmarks_discussed}
128 +
129 +\section{Volume, structural change and composition}
130 +\label{sec:thy-statistics}
131 +
132 +\subsection{Growth in absolute terms, decline in relative terms}
133 +
134 +Table~\ref{tab:theory_counts} gives the yearly counts of works carrying each category (multi-label) and of
135 +works whose primary category it is; Table~\ref{tab:theory_summary} summarises growth and structural change.
136 +The multi-label counts grew in every category after 2022. Between 2023 and 2025 the compound annual growth was
137 +\ThyCagrMultilabelSoftwareEngineeringPostPct\% for software engineering,
138 +\ThyCagrMultilabelComputationalComplexityPostPct\% for computational complexity,
139 +\ThyCagrMultilabelCompilersPostPct\% for compilers, \ThyCagrMultilabelAlgorithmsPostPct\% for algorithms,
140 +\ThyCagrMultilabelFormalVerificationPostPct\% for formal verification and
141 +\ThyCagrMultilabelProgrammingLanguagesPostPct\% for programming languages, against
142 +\ThyCagrMultilabelSoftwareEngineeringBaselinePct\%, \ThyCagrMultilabelComputationalComplexityBaselinePct\%,
143 +\ThyCagrMultilabelCompilersBaselinePct\%, \ThyCagrMultilabelAlgorithmsBaselinePct\%,
144 +\ThyCagrMultilabelFormalVerificationBaselinePct\% and \ThyCagrMultilabelProgrammingLanguagesBaselinePct\%
145 +between 2018 and 2022. The acceleration is real but not specific: the corpus-wide multi-label growth rose from
146 +\ThyCagrMultilabelTotalBaselinePct\% to \ThyCagrMultilabelTotalPostPct\% over the same periods, so the six
147 +categories accelerated with the corpus rather than within it. Measured by primary-label share, every category
148 +lost ground between 2018 and 2025: algorithms from \ThySharePrimaryAlgorithmsTwoZeroOneEightPct\% to
149 +\ThySharePrimaryAlgorithmsTwoZeroTwoFivePct\%, software engineering from
150 +\ThySharePrimarySoftwareEngineeringTwoZeroOneEightPct\% to \ThySharePrimarySoftwareEngineeringTwoZeroTwoFivePct\%,
151 +formal verification from \ThySharePrimaryFormalVerificationTwoZeroOneEightPct\% to
152 +\ThySharePrimaryFormalVerificationTwoZeroTwoFivePct\%, programming languages from
153 +\ThySharePrimaryProgrammingLanguagesTwoZeroOneEightPct\% to \ThySharePrimaryProgrammingLanguagesTwoZeroTwoFivePct\%,
154 +computational complexity from \ThySharePrimaryComputationalComplexityTwoZeroOneEightPct\% to
155 +\ThySharePrimaryComputationalComplexityTwoZeroTwoFivePct\% and compilers from
156 +\ThySharePrimaryCompilersTwoZeroOneEightPct\% to \ThySharePrimaryCompilersTwoZeroTwoFivePct\%. The theory and
157 +software categories are growing and shrinking at once: more works, a smaller fraction of a corpus whose growth
158 +is set by the AI categories of Part~V.
159 +
160 +\input{tables/tab_theory_counts}
161 +
162 +Two cautions attach to these numbers. The post-period growth rate spans two years and is exposed to the
163 +indexing lag of the most recent year (Section~\ref{sec:m-truncation}); the primary-label counts for 2025 are
164 +the last complete ones. And the file \texttt{results/bibliometrics/cagr\_by\_category.csv}, from which
165 +Table~\ref{tab:cagr} is generated, reports primary counts for these categories about three times smaller than
166 +those of the graph counts file (algorithms in 2018: \ThyCagrfileAlgorithmsNTwoZeroOneEight{} against
167 +\ThyCountAlgorithmsTwoZeroOneEightPrimary{}), so it predates the current classification; its growth rates are
168 +not used here, and the discrepancy is recorded for the reproducibility audit.\footnote{The generator
169 +\texttt{bibliometrics/04\_diversity\_cagr.py} reads \texttt{category\_counts\_yearly.csv}; re-running it on the
170 +current counts resolves the discrepancy. Until then, Table~\ref{tab:cagr} and Table~\ref{tab:theory_summary}
171 +disagree for these six rows.}
172 +
173 +\subsection{Structural change at the reference date}
174 +
175 +The interrupted-time-series slope change at December 2022 is positive for all six primary-label series and
176 +significant with HAC standard errors (Section~\ref{sec:m-its}): \ThyBreakAlgorithmsChatgptSlopePct\% per month
177 +for algorithms, \ThyBreakComputationalComplexityChatgptSlopePct\% for computational complexity,
178 +\ThyBreakProgrammingLanguagesChatgptSlopePct\% for programming languages,
179 +\ThyBreakCompilersChatgptSlopePct\% for compilers, \ThyBreakFormalVerificationChatgptSlopePct\% for formal
180 +verification and \ThyBreakSoftwareEngineeringChatgptSlopePct\% for software engineering. The placebo test
181 +(Section~\ref{sec:m-placebo}) does not confirm any of them at the five per cent level: the share of
182 +baseline-period pseudo-events whose slope statistic exceeds the observed one is
183 +\ThyBreakAlgorithmsPlaceboSlopeP{} for algorithms, \ThyBreakComputationalComplexityPlaceboSlopeP{} for
184 +computational complexity, \ThyBreakSoftwareEngineeringPlaceboSlopeP{} for software engineering,
185 +\ThyBreakCompilersPlaceboSlopeP{} for compilers, \ThyBreakFormalVerificationPlaceboSlopeP{} for formal
186 +verification and \ThyBreakProgrammingLanguagesPlaceboSlopeP{} for programming languages. The two theory
187 +categories are closest to the threshold; none crosses it. A slope change that pseudo-events of the baseline
188 +reproduce at these rates is a feature of series that were already accelerating, not of the reference date.
189 +
190 +\input{tables/tab_theory_summary}
191 +
192 +The level changes point the other way in five categories. At December 2022 the level fell by
193 +\ThyBreakFormalVerificationChatgptLevelPct\% for formal verification ($p = \ThyBreakFormalVerificationChatgptLevelP$),
194 +\ThyBreakComputationalComplexityChatgptLevelPct\% for computational complexity
195 +($p = \ThyBreakComputationalComplexityChatgptLevelP$), \ThyBreakAlgorithmsChatgptLevelPct\% for algorithms
196 +($p = \ThyBreakAlgorithmsChatgptLevelP$), \ThyBreakProgrammingLanguagesChatgptLevelPct\% for programming
197 +languages and \ThyBreakCompilersChatgptLevelPct\% for compilers (both not significant); only software
198 +engineering shows a positive level change (\ThyBreakSoftwareEngineeringChatgptLevelPct\%,
199 +$p = \ThyBreakSoftwareEngineeringChatgptLevelP$), and it is also the only category with a significant positive
200 +level change at the GPT-4 date (\ThyBreakSoftwareEngineeringGptFourLevel{},
201 +$p = \ThyBreakSoftwareEngineeringGptFourLevelP$; Section~\ref{sec:m-altdates}). The negative level changes
202 +are at least partly the seasonality of proceedings dates in categories dominated by conferences, and the
203 +alternative-date estimates for the o1 date (\ThyBreakComputationalComplexityOOneLevel{} for computational
204 +complexity, $p = \ThyBreakComputationalComplexityOOneLevelP$) show how sensitive a level estimate is to where
205 +the window is cut. The Bai--Perron procedure (Section~\ref{sec:m-bp}) dates a single break at
206 +\ThyBreakAlgorithmsBpBreaksBic{} for algorithms, computational complexity and compilers, two breaks for
207 +programming languages (\ThyBreakProgrammingLanguagesBpBreaksBic{}) and formal verification
208 +(\ThyBreakFormalVerificationBpBreaksBic{}), and one at \ThyBreakSoftwareEngineeringBpBreaksBic{} for software
209 +engineering: the data-driven dates fall in spring 2023 for the three categories with a single break, which is
210 +after the reference date and before the GPT-4 release had been indexed, and they say nothing about cause.
211 +
212 +\subsection{Vocabulary moves, citations do not}
213 +
214 +Composition is where the window is visible. Table~\ref{tab:theory_llm_cooccurrence} reports, for each
215 +category and year, the number of works that the multi-label classifier also assigns to the LLM category and
216 +their share of the category. Between 2022 and 2025 that share rose from
217 +\ThyCoocSoftwareEngineeringLlmTwoZeroTwoTwoPct\% to \ThyCoocSoftwareEngineeringLlmTwoZeroTwoFivePct\% in
218 +software engineering, from \ThyCoocProgrammingLanguagesLlmTwoZeroTwoTwoPct\% to
219 +\ThyCoocProgrammingLanguagesLlmTwoZeroTwoFivePct\% in programming languages, from
220 +\ThyCoocCompilersLlmTwoZeroTwoTwoPct\% to \ThyCoocCompilersLlmTwoZeroTwoFivePct\% in compilers, from
221 +\ThyCoocFormalVerificationLlmTwoZeroTwoTwoPct\% to \ThyCoocFormalVerificationLlmTwoZeroTwoFivePct\% in formal
222 +verification, and from \ThyCoocAlgorithmsLlmTwoZeroTwoTwoPct\% to \ThyCoocAlgorithmsLlmTwoZeroTwoFivePct\% in
223 +both algorithms and computational complexity. The gradient runs from the code-facing categories to the
224 +mathematical ones, and it is a gradient of vocabulary: co-classification is a lexical and topical signal
225 +(Section~\ref{sec:m-classification}), not a measure of method adoption.
226 +
227 +\input{tables/tab_theory_llm_cooccurrence}
228 +
229 +Citation flows are more conservative. In the post period, software-engineering works direct
230 +\ThyFlowSoftwareEngineeringToLlmPostPct\% of their categorised references to works of the LLM category,
231 +computational complexity \ThyFlowComputationalComplexityToLlmPostPct\%, compilers
232 +\ThyFlowCompilersToLlmPostPct\%, programming languages \ThyFlowProgrammingLanguagesToLlmPostPct\%, formal
233 +verification \ThyFlowFormalVerificationToLlmPostPct\% and algorithms \ThyFlowAlgorithmsToLlmPostPct\%; the
234 +within-category shares barely moved (algorithms \ThyFlowAlgorithmsToAlgorithmsBaselinePct\% to
235 +\ThyFlowAlgorithmsToAlgorithmsPostPct\%, software engineering
236 +\ThyFlowSoftwareEngineeringToSoftwareEngineeringBaselinePct\% to
237 +\ThyFlowSoftwareEngineeringToSoftwareEngineeringPostPct\%, formal verification
238 +\ThyFlowFormalVerificationToFormalVerificationBaselinePct\% to
239 +\ThyFlowFormalVerificationToFormalVerificationPostPct\%).\footnote{Flows are aggregated from
240 +\texttt{results/graph/category\_flows\_yearly.csv} over citing years 2018--2022 (baseline) and 2023--2026
241 +(post) and are shares of the references that fall inside the categorised corpus (Section~\ref{sec:m-citgraph}).
242 +The LLM category is young, which depresses its share of cited works mechanically; the comparison with the
243 +co-classification shares is therefore indicative, not a test.} Two flow changes deserve notice. The
244 +self-citation share of computational complexity fell from
245 +\ThyFlowComputationalComplexityToComputationalComplexityBaselinePct\% to
246 +\ThyFlowComputationalComplexityToComputationalComplexityPostPct\% while its references to human--computer
247 +interaction rose from \ThyFlowComputationalComplexityToHciBaselinePct\% to
248 +\ThyFlowComputationalComplexityToHciPostPct\%; given the candidate list of
249 +Section~\ref{sec:thy-landmarks}, this is more plausibly the category's contamination by social-science works
250 +than a change in theoretical research, and it is recorded as a contested reading. Inbound citations to formal
251 +verification from computer vision doubled, from \ThyInflowFormalVerificationFromComputerVisionBaselinePct\% to
252 +\ThyInflowFormalVerificationFromComputerVisionPostPct\% of the category's inbound references, which is
253 +consistent with the neural-network verification literature reviewed in Section~\ref{sec:thy-fv} being cited by
254 +robustness work; no paper-level breakdown tests this.
255 +
256 +\section{Algorithms and data structures}
257 +\label{sec:thy-algorithms}
258 +
259 +\subsection{What the window established}
260 +
261 +The category's largest results are not in the candidate list, and the first task of the review is to say what
262 +they are, from verified records. Maximum flow and minimum-cost flow were brought to almost-linear time
263 +\citep{chen2022maxflow}, closing a line of work that had run through interior-point methods and dynamic graph
264 +data structures for a decade. Single-source shortest paths with negative weights were brought to near-linear
265 +time by a combinatorial algorithm \citep{bernstein2022negative}, a result republished in the \emph{Journal of
266 +the ACM} and \emph{Communications of the ACM} in 2025. The sorting barrier for directed single-source shortest
267 +paths, the $\Omega(m + n \log n)$ cost that Dijkstra's algorithm with a priority queue cannot avoid, was
268 +broken by a deterministic $O(m \log^{2/3} n)$ algorithm \citep{duan2025sorting}, the best paper of STOC~2025.
269 +In data structures, the open-addressing bounds conjectured by Yao in 1985 were shown to be beatable: without
270 +reordering, a greedy scheme achieves expected amortised probe complexity $O(\log^2 \delta^{-1})$ at load
271 +$1 - \delta$, against the conjectured $\Theta(\delta^{-1})$ \citep{farachcolton2024openaddressing}. These
272 +four results, from 2022 to 2025, are what a specialist would list first for the window; none has the citation
273 +profile that the ranking rewards, because theoretical results are cited by a small community over long
274 +horizons, and the 36-month window and the field normalisation are both calibrated on the behaviour of
275 +applied categories.
276 +
277 +What the corpus does record is the programme of \emph{algorithms with predictions}, which the baseline period
278 +established and the window continued. The candidate list carries it through a 2024 arXiv re-posting of the 2018
279 +conference paper on ski rental and non-clairvoyant scheduling with machine-learned predictions, whose
280 +abstract describes algorithms that are ``oblivious to the performance of the predictor, improve with better
281 +predictions, but do not degrade much if the predictions are poor''; the programme's canonical references are
282 +the journal version of competitive caching with learned advice \citep{lykouris2021caching} and the overview in
283 +\emph{Communications of the ACM} \citep{mitzenmacher2022predictions}. The formal object is a pair of
284 +guarantees, consistency (the competitive ratio when the prediction is exact) and robustness (the ratio under
285 +arbitrary error), with a trade-off between them; it is the point at which online algorithms and machine
286 +learning meet without either giving up its guarantees, and it is the one place in the category where the
287 +``learning'' vocabulary of the window names a theory rather than a tool. The fine-grained complexity programme
288 +that organised algorithms research at the baseline (Section~\ref{sec:p2-theory}) is represented in the
289 +bibliography by the ICM survey \citep{vassilevskawilliams2019finegrained} and by the parameterized-algorithms
290 +monograph \citep{cygan2015parameterized}.
291 +
292 +\subsection{What the candidate list contains}
293 +
294 +Of the \ThyAuditAlgorithmsN{} algorithms candidates, \ThyAuditAlgorithmsInScopeResearch{} are research
295 +contributions and none is misclassified; the rest are textbook chapters, re-deposited classics, theses and
296 +metaheuristics. The two post-period research candidates are modest and both carry citation counts that belong
297 +to other works: the WADS~2025 paper extends Kd-trees to Bregman divergences, shows that the triangle inequality
298 +is not needed for correct pruning, and reports two orders of magnitude over linear search for
299 +Kullback--Leibler nearest-neighbour queries in up to a hundred dimensions \citep{pham2025kdtree}; the SEA~2024
300 +paper replaces the highest-degree branching rule of branch-and-reduce for maximum independent set with a graph
301 +neural network \citep{langedal2024mis}. Both abstracts describe engineering contributions with experimental
302 +evaluation, which is what algorithm engineering is; neither is the kind of result that defines the window.
303 +
304 +The other research genre in the list is the metaphor-based metaheuristic: the Arithmetic Optimization Algorithm
305 +\citep{abualigah2021aoa} and the Aquila Optimizer \citep{abualigah2021aquila} in 2021, and the artificial
306 +lemming, glider snake, gray langur and starfish optimisers in 2025--2026, published in engineering and
307 +AI-review journals and evaluated on CEC benchmark functions and constrained design problems. The abstracts share
308 +a template, a natural metaphor mathematically modelled into exploration and exploitation phases and
309 +``outperforming or matching'' a dozen earlier metaheuristics. The operational-research literature has argued
310 +since \citet{sorensen2013metaphor} that the genre re-describes existing methods under new names and that its
311 +benchmarks do not discriminate; the position was restated in a ``call for action'' \citep{aranha2021metaphor}
312 +and substantiated for six widely cited algorithms, shown to be equivalent to or special cases of earlier
313 +methods \citep{camachovillalon2022exposing}. The corpus does not adjudicate this dispute; what it shows is
314 +the genre's weight in the highly cited stratum of the algorithms category, where six of twenty-five candidates
315 +belong to it and where the post-2022 research articles with the highest normalised scores are of this kind.
316 +Whether the genre grows faster than the rest of the category and absorbs a growing share of its citations is an
317 +open problem of Part~X (G-P7), answerable with a keyword-defined subseries.
318 +
319 +\subsection{Algorithms discovered by learning systems}
320 +
321 +Three Nature papers and one preprint report algorithms found by reinforcement learning or program search:
322 +faster matrix-multiplication tensor decompositions for small sizes \citep{fawzi2022alphatensor}, shorter
323 +sorting networks for short sequences, merged into a standard library \citep{mankowitz2023alphadev}, new
324 +constructions for the cap-set problem and bin packing from evolved programs \citep{romeraparedes2024funsearch},
325 +and, as a 2025 preprint, a coding agent that reports improvements on a collection of open mathematical and
326 +algorithmic problems \citep{novikov2025alphaevolve}. All four are reports by the developers; the improvements
327 +are on specific instances; and no independent reproduction or extension by the theory community appears in the
328 +verified bibliography. The question these works pose to the category, whether search over programs changes how
329 +algorithms are designed or only how constants are tuned, is one the corpus cannot answer at this granularity;
330 +it is stated as open problem G-P9. The honest position is that the window's algorithmic breakthroughs were
331 +obtained by people, and that the learned-discovery results are promising, contested in scope, and awaiting
332 +replication.
333 +
334 +\section{Computational complexity}
335 +\label{sec:thy-complexity}
336 +
337 +The category's statistics are the least trustworthy of the six: \ThyAuditComputationalComplexityMisclassified{}
338 +of its \ThyAuditComputationalComplexityN{} candidates are outside computing, and the flow and co-occurrence
339 +changes of Section~\ref{sec:thy-statistics} are to that extent artefacts. The two in-scope research candidates
340 +are a STOC~2019 paper and a 2026 \emph{Theoretical Computer Science} article on SAT in spiking neural P systems
341 +(no abstract). The STOC paper is a genuine landmark of the baseline: its abstract describes the reachability
342 +problem for Petri nets as open since the 1960s, with Lipton's 1976 exponential-space lower bound unimproved, and
343 +establishes a non-elementary lower bound \citep{czerwinski2019petri}. The story was completed within the
344 +baseline period, when reachability in vector addition systems was shown Ackermann-complete
345 +\citep{czerwinski2021ackermann,leroux2021notprimitive}, matching the upper bound of Leroux and Schmitz; the
346 +window inherits a closed problem whose closure the corpus ranks below a technology-acceptance survey.
347 +
348 +The window's own contribution to complexity theory that a specialist would name first is the simulation of
349 +time $t$ in space $O(\sqrt{t \log t})$ \citep{williams2025sqrtspace}, which improves the Hopcroft--Paul--Valiant
350 +simulation of 1975 after fifty years and implies, through a known argument, that there are problems solvable
351 +in linear space that need nearly quadratic time on multitape machines; the paper is a STOC~2025 result
352 +republished in the \emph{Journal of the ACM} in 2026. In additive combinatorics, the strong bounds for
353 +three-term arithmetic progressions \citep{kelley2023threeap} brought the Roth-type exponent to a form conjectured
354 +since Behrend. Neither appears in the candidate list; both enter through verified records, and the general
355 +point is the one made for algorithms: results of this kind are cited slowly and by few, and a 36-month
356 +field-normalised window does not see them.
357 +
358 +The boundary with language models is where the category's co-classification share grew, yet the relevant
359 +works are mostly classified elsewhere. The proof that transformers with logarithmic precision lie in uniform
360 +$\mathrm{TC}^0$ \citep{merrill2023parallelism}, and the survey of what formal languages transformers can
361 +express \citep{strobl2024survey}, bound what a single forward pass can compute and make the case that
362 +chain-of-thought computation is where additional expressive power must come from; they are, in the verified
363 +bibliography, the window's main contribution of complexity theory to the study of the models that dominate
364 +Part~V. Whether such works are classified under computational complexity or under natural language processing
365 +is an empirical question about the taxonomy (open problem G-P8). The remainder of the candidate list is
366 +teaching material, including the fourth edition of the Kolmogorov-complexity monograph \citep{li2019kolmogorov}
367 +and undergraduate chapters on complexity and on parameterized complexity, and re-deposited classics, including
368 +Turing's 1936 paper and Wolfram's cellular-automata chapters.
369 +
370 +\section{Programming languages}
371 +\label{sec:thy-pl}
372 +
373 +The baseline lineage was set out in Section~\ref{sec:p2-pl}: Rust and its semantic foundation
374 +\citep{matsakis2014rust,jung2017rustbelt,jung2021saferust}, Lean~4 as a prover that is also a programming
375 +language \citep{demoura2021lean4} with its mathematical library \citep{mathlib2020}, and the verified
376 +artefacts CompCert \citep{leroy2009compcert} and seL4 \citep{klein2009sel4} that showed production software
377 +can be proved correct. The candidate list records this lineage only obliquely, through three theses of the
378 +L4/seL4 school re-deposited in 2022 (separation logic for programs in virtual memory, formal memory models for
379 +C systems code, a verified multiprocessor kernel design), which the ranking places in the post period. In the
380 +window the lineage continued with Verus \citep{lattuada2024verus}, a verification framework for Rust programs
381 +presented at SOSP~2024, which treats the ownership discipline of the language as the foundation for
382 +specification and proof rather than as an obstacle to them; the content of Verus is known here from its verified
383 +title and venue only.
384 +
385 +Of the \ThyAuditProgrammingLanguagesN{} candidates, \ThyAuditProgrammingLanguagesInScopeResearch{} are
386 +research contributions. Two of them mark the entry of language models into the community's top venue in
387 +2024. LLift \citep{li2024llift} combines static analysis with an LLM to find use-before-initialisation bugs in
388 +the Linux kernel; its abstract reports post-constraint guidance of the path analysis, procedures for handling
389 +``the unpredictable nature of LLMs'', and four previously unknown bugs acknowledged by the kernel community.
390 +WhiteFox \citep{yang2024whitefox} is described by its abstract as the first white-box compiler fuzzer that uses
391 +an LLM reading the source code of optimisations to generate test programs that exercise them, against
392 +black- and grey-box fuzzing that lacks that information and symbolic execution that does not scale to a
393 +compiler. Both papers report bugs found; neither reports a guarantee, and the abstracts do not claim one. The
394 +natural comparison is with the two instruments that the compiler community built for correctness before the
395 +window, Csmith's random differential testing, which found hundreds of bugs in production C compilers
396 +\citep{yang2011csmith}, and Alive2's bounded translation validation of LLVM optimisations
397 +\citep{lopes2021alive2}; whether LLM-generated tests can be combined with validation of that kind, so that the
398 +share of reported bugs that are confirmed can be measured per optimisation pass, is open problem G-P4. The
399 +third research candidate, ZEUS \citep{kalra2018zeus}, is the baseline representative of program analysis for a
400 +new target: abstract interpretation and constrained Horn clauses applied to smart contracts, with an abstract
401 +reporting that about 95\% of 22,400 contracts analysed were vulnerable. The universal fuzzer of
402 +\citet{xia2024fuzz4all} and the LLM-based bug-reproduction study of \citet{kang2023fewshottesters} are the
403 +software-engineering counterparts of the same move.
404 +
405 +The category's vocabulary moved more than any except software engineering's: co-classification with the LLM
406 +category reached \ThyCoocProgrammingLanguagesLlmTwoZeroTwoFivePct\% in 2025. Its references to LLM works are
407 +\ThyFlowProgrammingLanguagesToLlmPostPct\%, and its structural ties are unchanged, with
408 +\ThyFlowProgrammingLanguagesToFormalVerificationPostPct\% of its references going to formal verification and
409 +\ThyFlowProgrammingLanguagesToSoftwareEngineeringPostPct\% to software engineering in the post period, against
410 +\ThyFlowProgrammingLanguagesToFormalVerificationBaselinePct\% and
411 +\ThyFlowProgrammingLanguagesToSoftwareEngineeringBaselinePct\% at the baseline. Programming languages is, in
412 +the citation graph, the hinge between verification and engineering, and the window did not move the hinge.
413 +
414 +\section{Compilers}
415 +\label{sec:thy-compilers}
416 +
417 +Compilers is the smallest of the six categories, \ThySharePrimaryCompilersTwoZeroTwoFivePct\% of primary labels
418 +in 2025, and the most lexically contaminated after complexity. Its in-scope candidates nevertheless trace a
419 +coherent line. At the baseline, MLIR \citep{lattner2021mlir} is the infrastructure whose abstract promises to
420 +reduce ``the cost of building domain specific compilers'' and to connect existing ones; Tiramisu
421 +\citep{baghdadi2019tiramisu} is a polyhedral compiler whose scheduling language separates algorithm, loop
422 +transformation, data layout and communication so that one algorithm can target multicores, GPUs and
423 +distributed machines; ExaStencils is a four-layer domain-specific language for multigrid solvers. Two surveys
424 +establish that learned compiler heuristics are a programme older than the window: the \emph{Proceedings of the
425 +IEEE} survey of machine learning in compiler optimisation \citep{wang2018mlcompiler} and the \emph{Computing
426 +Surveys} review of autotuning, optimisation selection and phase ordering \citep{ashouri2018autotuning}, which
427 +dates the first attempts to the mid-1990s. Three MIT and Delaware theses on learned optimisation selection and
428 +on sparse tensor-algebra compilation complete the baseline picture.
429 +
430 +The post-period candidates show the programme's change of scale. LLM4VV \citep{munley2024llm4vv} uses language
431 +models to generate a validation test suite for directive-based parallel programming compilers (metadata only;
432 +no abstract). LLM Compiler \citep{cummins2025llmcompiler} is a suite of foundation models built on Code Llama
433 +and trained, per its abstract, on 546 billion tokens of LLVM intermediate representation and assembly, then
434 +instruction-tuned to predict and emulate compiler behaviour; fine-tuned versions reach 77\% of the optimising
435 +potential of an autotuning search for code size and disassemble x86-64 and ARM assembly back to LLVM-IR. The
436 +preprint that preceded it \citep{cummins2023llmcompiler} framed the task as predicting optimisation-pass
437 +sequences. The shift is from learning a heuristic inside a compiler, with features chosen by compiler writers,
438 +to pre-training a model on the compiler's own artefacts; what the shift leaves open is the question that
439 +compiler research has always had to answer last, correctness, which the field handled through differential
440 +testing \citep{yang2011csmith} and translation validation \citep{lopes2021alive2} rather than through model
441 +evaluation. Co-classification with the LLM category reached \ThyCoocCompilersLlmTwoZeroTwoFivePct\% in 2025;
442 +references to LLM works are \ThyFlowCompilersToLlmPostPct\%, and the category's citations continue to go
443 +first to computer architecture (\ThyFlowCompilersToComputerArchitecturePostPct\% of references in the post
444 +period) and to itself. The deep-learning compiler stack recorded in the main bibliography, TVM, Triton and
445 +PyTorch~2 \citep{chen2018tvm,tillet2019triton,ansel2024pytorch2}, and the image-processing DSL that preceded
446 +it \citep{ragankelley2013halide}, belong to Part~VI and are only noted here as the other post-2018 development
447 +of the category.
448 +
449 +\section{Formal verification}
450 +\label{sec:thy-fv}
451 +
452 +Formal verification has the most in-scope research candidates of the six
453 +(\ThyAuditFormalVerificationInScopeResearch{} of \ThyAuditFormalVerificationN{}) and the clearest internal
454 +structure, which the review follows: solvers and competitions, neural-network verification, verification of new
455 +targets, learned theorem proving, and the theory of verification.
456 +
457 +\paragraph{Solvers and competitions.}
458 +cvc5 \citep{barbosa2022cvc5} is the system description of the SMT solver that succeeded CVC4, evaluated on all
459 +SMT-LIB benchmarks against CVC4 and Z3; the Handbook of Satisfiability in its second edition
460 +\citep{biere2021handbook} is the reference work, present twice in the candidate list under two records. The
461 +SV-COMP~2024 report \citep{beyer2024svcomp} describes the largest edition of the software-verification
462 +competition to that date, 76 tools from 34 teams on 30,300 verification tasks for C programs, with a second
463 +track for witness validation and a new witness format. Competitions of this kind are the category's benchmark
464 +culture, and they are a point of contrast with the benchmark culture of Part~V: the tasks are fixed programs
465 +with fixed specifications, a tool's result is a proof or a counterexample that a validator checks, and the
466 +score is not self-reported.
467 +
468 +\paragraph{Neural-network verification.}
469 +The baseline built a sub-field: Reluplex \citep{katz2017reluplex} extended the simplex method to ReLU
470 +constraints; the 2018 symbolic-interval preprint in the candidate list argued, per its abstract, that existing
471 +analyses ``struggle to scale beyond very small networks'' and proposed interval propagation that does;
472 +DeepPoly \citep{singh2019deeppoly} introduced an abstract domain of floating-point polyhedra with intervals and
473 +transformers for ReLU, sigmoid, tanh and maxpool, and reported the first proofs of robustness to rotations
474 +with interpolation; Marabou \citep{katz2019marabou} packaged SMT-based verification as a framework accepting
475 +TensorFlow models. The VNN-COMP report \citep{brix2023vnncomp} summarises the first three years of the
476 +competition that organised the sub-field. The doubling of formal verification's inbound citations from
477 +computer vision in the post period (Section~\ref{sec:thy-statistics}) is consistent with this literature being
478 +cited by robustness work, and the sub-field is the main channel through which the category meets the
479 +machine-learning categories; whether it has guarantees that survive the scale of the models of Part~V is not a
480 +question these abstracts address.
481 +
482 +\paragraph{New targets.}
483 +VerX \citep{permenev2020verx} is described by its abstract as the first automated verifier of functional
484 +properties of Ethereum smart contracts, combining reduction of temporal properties to reachability, a symbolic
485 +execution engine for the Ethereum virtual machine, and delayed predicate abstraction, evaluated on 83
486 +properties of 12 projects. The answer-set-programming monitor for automated driving \citep{klampfl2024asp}
487 +and the theses on parametric and multi-objective Markov models represent runtime and probabilistic
488 +verification; the Warwick thesis on model checking quantum protocols, re-deposited with a 2026 date, is a
489 +baseline work of the previous decade.
490 +
491 +\paragraph{Learned theorem proving.}
492 +The abstract attached in the corpus to the ITP~2023 paper, under a wrong title (Section~\ref{sec:thy-landmarks}),
493 +is the clearest statement in the candidate set of what machine learning has done for automated reasoning:
494 +MizAR \citep{jakubuv2023mizar} proves about 60\% of the theorems of the Mizar Mathematical Library in the hammer
495 +setting and 75\% when given the premises of the human proofs, through the E and Vampire provers with the ENIGMA
496 +and Deepire learned guidance and an incremental loop that grows a corpus of millions of proofs while training
497 +stronger systems on it. This is the saturation-prover line. The language-model line is represented by Baldur
498 +\citep{first2023baldur}, which generates and repairs whole Isabelle proofs; by LeanDojo
499 +\citep{yang2023leandojo} and DeepSeek-Prover \citep{xin2024deepseekprover}, both preprints, which report
500 +retrieval-augmented and synthetic-data-trained provers on Lean benchmarks; and by the position paper on formal
501 +mathematical reasoning as a frontier \citep{yang2024formalmath}, also a preprint. The figures these works
502 +report are on different benchmarks, under different compute budgets, with the possibility of overlap between
503 +training corpora and test theorems that the abstracts do not exclude; a benchmark of theorems posted after the
504 +training cut-off, with proof checking rather than self-reported success, is the direction stated as open
505 +problem G-P5. The machine-learned mathematics of Part~V, olympiad geometry \citep{trinh2024alphageometry} and
506 +program search \citep{romeraparedes2024funsearch}, is adjacent but distinct: it produces objects that a human or
507 +a checker then verifies, where the works above produce the proofs.
508 +
509 +\paragraph{Theory of verification.}
510 +Equality saturation as implemented in \emph{egg} \citep{willsey2021egg}, whose abstract describes e-graph
511 +rewriting that ``only adds information'' and extraction by a user-supplied cost function, is a baseline result
512 +republished as a research highlight in 2026; the quantitative interface automata of the CSL~2026 paper on reward
513 +interfaces, whose abstract introduces multi-valued correctness requirements and best-effort satisfaction into
514 +interface theories, is the one post-period research candidate of this kind.
515 +
516 +The category's primary share halved between 2018 and 2025 while its multi-label count grew by eleven per cent
517 +(Table~\ref{tab:theory_counts}). The arithmetic of the argmax assignment (Section~\ref{sec:m-score}) suggests
518 +why: a work that carries a verification label and an AI or security label loses the primary label to the
519 +larger category when the scores are close, so the primary series understates a category whose work is
520 +increasingly done in combination. Measuring the share of works whose primary label switched, by co-assigned
521 +category and year, is open problem G-P10.
522 +
523 +\section{Software engineering}
524 +\label{sec:thy-se}
525 +
526 +Software engineering is where the window's change is largest, best documented, and most contested. Its
527 +baseline was not AI-free (Section~\ref{sec:p2-pl}): the Codex models \citep{chen2021codex}, the
528 +code-generation benchmarks \citep{austin2021mbpp}, the systematic evaluation of code models
529 +\citep{xu2022codelms} and the first productivity assessment of neural code completion
530 +\citep{ziegler2022copilot} precede the reference date, and the survey of machine-learning testing
531 +\citep{zhang2022mltesting}, 144 papers on testing properties, components and workflows of learned systems,
532 +shows the category already organised around learned software as an object of test. The window's surveys
533 +reverse the direction of the relation. The survey of software testing with large language models
534 +\citep{wang2024llmtesting} analyses 102 studies and reports, per its abstract, that test-case preparation and
535 +program repair are the most common uses; the two systematic reviews of language models for software
536 +engineering \citep{fan2023llm4se,hou2024llm4se} and the code-generation survey \citep{jiang2024codegensurvey}
537 +map the same territory. Method papers from the main bibliography and the Part~VII-a bibliography give the
538 +texture: program repair in the era of pre-trained models \citep{xia2023apr}, bug reproduction from reports
539 +\citep{kang2023fewshottesters}, universal fuzzing \citep{xia2024fuzz4all}, and the agentic setting defined by
540 +SWE-bench \citep{jimenez2024swebench} and SWE-agent \citep{yang2024sweagent}, both preprints, in which a model
541 +is asked to resolve a GitHub issue end to end. Among the candidates, APISS \citep{chen2025apiss} retrieves
542 +functionality-equivalent APIs through LLM embeddings to migrate proof-of-concept exploits, and a 2026 paper
543 +proposes a multi-scale temporal model for just-in-time defect prediction; both are method contributions judged
544 +from their abstracts.
545 +
546 +\paragraph{Productivity: two findings that need reconciling.}
547 +The most cited post-period empirical candidate reports three randomised field experiments at Microsoft,
548 +Accenture and an unnamed Fortune~100 company, run by the firms in the course of business, in which a random
549 +subset of 4,867 developers received an AI coding assistant; the pooled estimate is a 26.08\% increase in
550 +completed tasks with a standard error of 10.3\%, each experiment being ``noisy'' on its own, and less
551 +experienced developers adopting more and gaining more \citep{cui2026genaiwork}. The earlier controlled
552 +experiment on a single task reported a larger speed-up \citep{peng2023copilot}, and the field studies of
553 +generative AI at work outside software \citep{brynjolfsson2023genaiwork,noy2023productivity,dellacqua2023jagged}
554 +report the same pattern of larger gains for less experienced workers. Against these stands a 2025 randomised
555 +trial preprint in which experienced open-source developers working on repositories they maintain took longer
556 +to complete issues with AI tools than without \citep{becker2025metr}. The two findings are not contradictory
557 +once population and outcome are distinguished: enterprise developers completing assigned tasks versus expert
558 +maintainers resolving issues in mature codebases they know well, and tasks completed versus time per issue.
559 +But the reconciliation is a reading of two abstracts, the second of a preprint, and the design that would
560 +settle it, a preregistered multi-site trial stratified by experience and repository maturity with time-to-merge
561 +and defect rate as joint outcomes, has not been run (open problem G-P3). The user studies of how programmers
562 +interact with code-generating models \citep{barke2023groundedcopilot,vaithilingam2022expectation,liang2024usability}
563 +supply the mechanism, acceleration mode against exploration mode, that any such trial would need to measure.
564 +
565 +\paragraph{Security of generated code.}
566 +The security evaluations agree with one another. Copilot's suggestions were found vulnerable at a substantial
567 +rate in 2022 \citep{pearce2022asleep}; the CCS~2023 user study found that participants with an AI assistant
568 +wrote less secure code and believed it more secure \citep{perry2023insecure}; and the ICPEC~2025 study in the
569 +candidate list, evaluating three open models on a filtered Big-Vul subset with eight CWE categories, reports
570 +``a sharp contrast between high detection rates and markedly poor classification accuracy''
571 +\citep{gasparlopes2025vuln}. High detection or completion paired with poor fine-grained security reasoning is
572 +the consistent pattern across three designs, and it is the category's clearest negative result about the
573 +models.
574 +
575 +\paragraph{The statistics.}
576 +Software engineering has the highest co-classification with the LLM category of the six
577 +(\ThyCoocSoftwareEngineeringLlmTwoZeroTwoFivePct\% of its works in 2025, from
578 +\ThyCoocSoftwareEngineeringLlmTwoZeroTwoTwoPct\% in 2022), the largest post-period growth
579 +(\ThyCagrMultilabelSoftwareEngineeringPostPct\% per year), the only positive level change at the reference
580 +date and the only significant positive level change at the GPT-4 date (Section~\ref{sec:thy-statistics}).
581 +Its references to LLM works are \ThyFlowSoftwareEngineeringToLlmPostPct\% in the post period, from
582 +\ThyFlowSoftwareEngineeringToLlmBaselinePct\%, and inbound citations from the LLM category are
583 +\ThyInflowSoftwareEngineeringFromLlmPostPct\% of its inbound references. Whether a vocabulary shift of this
584 +size is a method shift, that is, whether the works co-classified with LLM cite LLM method papers at a higher
585 +rate than the category's baseline, can be tested on the knowledge graph of Part~III once method nodes are
586 +extracted for the high-priority set; it is open problem G-P2. The research-software case of the Astropy
587 +project \citep[in the candidate list;][]{collberg2016repeatability} and the reproducibility literature of
588 +Part~IX are the other side of the same category, and the only part of it where the window changed nothing
589 +visible.
590 +
591 +\section{Synthesis}
592 +\label{sec:thy-synthesis}
593 +
594 +Three conclusions follow, in decreasing order of evidential strength.
595 +
596 +The first is about the instrument. In the theory and software areas the automatically selected landmark
597 +candidates do not identify landmarks: \ThyAuditVerdictInScopeResearch{} of \ThyAuditN{} are in-scope research
598 +works, \ThyAuditIdCollision{} carry citation windows that belong to other works, and the rest are chapters with
599 +generic titles, re-deposited classics, theses in small normalisation cells and lexical misclassifications. The
600 +defects are identifiable and partly correctable at the selection stage, and the audit table
601 +(Table~\ref{tab:theory_landmark_audit}) should accompany any use of the list; the major theoretical results of
602 +the window, almost-linear flow, near-linear negative-weight shortest paths, the directed sorting barrier,
603 +optimal open addressing, square-root-space simulation and the 3-progression bounds, are invisible to a
604 +36-month field-normalised ranking and were brought into this review through verified records. The same
605 +blindness will affect any bibliometric answer to RQ3 and RQ7 for these categories.
606 +
607 +The second is about the statistics. The six categories grew in absolute terms and declined in relative terms
608 +after 2022; their slope changes at the reference date are positive but not distinguishable from pseudo-events
609 +in the baseline; their vocabulary moved toward language models in proportion to their distance from
610 +mathematics; and their citation behaviour moved little. For RQ2 the answer is that none of the six is among
611 +the categories with structural change that survives the placebo test, and for RQ5 that the relationship with
612 +the foundation-model categories changed in what these works say far more than in what they cite.
613 +
614 +The third is about the research. The substantive developments of the window are of two kinds that the corpus
615 +treats very differently. Theoretical results were obtained by people, at the rate and in the venues of the
616 +previous decade, and are cited accordingly. The entry of language models into verification, program analysis,
617 +compiler testing and software engineering is visible in the co-classification statistics and in the abstracts
618 +of the in-scope candidates, and it is where the open problems lie: guarantees for learned program analysis and
619 +compiler testing (G-P4), comparability of learned provers (G-P5), the productivity effect on experienced
620 +developers (G-P3), and the distinction between a vocabulary and a method (G-P2). What this half of Part~VII
621 +can say with confidence is which of these questions the corpus can answer with its present instruments, and
622 +which it cannot.
added results/graph/semantic_crossing_yearly.csv +361 −0
@@ -0,0 +1,361 @@
1 +year,primary_category,n_edges,share_cross_category
2 +2018,ai_accelerators,6157,0.5181
3 +2018,ai_agents,30233,0.548
4 +2018,ai_safety,5267,0.569
5 +2018,algorithms,170250,0.3159
6 +2018,applied_cryptography,60756,0.4146
7 +2018,autonomous_scientific_discovery,495,0.9253
8 +2018,blockchain,35962,0.2207
9 +2018,cloud_computing,58539,0.4688
10 +2018,compilers,4635,0.8194
11 +2018,computational_biology,92838,0.1982
12 +2018,computational_complexity,20738,0.6304
13 +2018,computer_architecture,140938,0.3222
14 +2018,computer_graphics,63877,0.4237
15 +2018,computer_networks,446365,0.1723
16 +2018,computer_vision,463491,0.1823
17 +2018,cybersecurity,183347,0.3546
18 +2018,database_systems,102960,0.4521
19 +2018,deep_learning_architectures,74595,0.4948
20 +2018,distributed_computing,75532,0.5767
21 +2018,edge_computing,34523,0.5445
22 +2018,federated_learning,726,0.8209
23 +2018,formal_verification,33962,0.465
24 +2018,foundation_models,245,0.9061
25 +2018,generative_ai,5206,0.6498
26 +2018,hci,216031,0.2699
27 +2018,information_retrieval,71592,0.3751
28 +2018,llm,4817,0.8337
29 +2018,ml_theory,91278,0.4192
30 +2018,multimodal_ai,2851,0.7555
31 +2018,nlp,158887,0.2678
32 +2018,operating_systems,9487,0.8001
33 +2018,privacy_preserving_computation,19705,0.4446
34 +2018,programming_languages,27407,0.6148
35 +2018,quantum_computing,39415,0.1159
36 +2018,reinforcement_learning,22774,0.351
37 +2018,robotics,134764,0.242
38 +2018,scientific_computing,86496,0.4539
39 +2018,software_engineering,179454,0.2979
40 +2018,speech_audio,77812,0.1873
41 +2018,sustainable_computing,3256,0.8068
42 +2019,ai_accelerators,7263,0.531
43 +2019,ai_agents,30144,0.5601
44 +2019,ai_safety,8150,0.5583
45 +2019,algorithms,179415,0.3109
46 +2019,applied_cryptography,68116,0.4176
47 +2019,autonomous_scientific_discovery,715,0.9105
48 +2019,blockchain,52593,0.2132
49 +2019,cloud_computing,57803,0.4702
50 +2019,compilers,4164,0.8343
51 +2019,computational_biology,109708,0.2013
52 +2019,computational_complexity,22012,0.6335
53 +2019,computer_architecture,137403,0.3238
54 +2019,computer_graphics,60635,0.4342
55 +2019,computer_networks,447623,0.1791
56 +2019,computer_vision,554061,0.1766
57 +2019,cybersecurity,202702,0.3475
58 +2019,database_systems,107527,0.4559
59 +2019,deep_learning_architectures,97754,0.4835
60 +2019,distributed_computing,76991,0.5673
61 +2019,edge_computing,45246,0.5424
62 +2019,federated_learning,1161,0.8243
63 +2019,formal_verification,29780,0.4782
64 +2019,foundation_models,506,0.8972
65 +2019,generative_ai,8293,0.6549
66 +2019,hci,232616,0.266
67 +2019,information_retrieval,74631,0.3586
68 +2019,llm,6560,0.8125
69 +2019,ml_theory,118641,0.4323
70 +2019,multimodal_ai,4264,0.7207
71 +2019,nlp,179279,0.2567
72 +2019,operating_systems,10082,0.7343
73 +2019,privacy_preserving_computation,21380,0.4506
74 +2019,programming_languages,28291,0.6236
75 +2019,quantum_computing,48912,0.1143
76 +2019,reinforcement_learning,34163,0.3233
77 +2019,robotics,151045,0.2428
78 +2019,scientific_computing,96671,0.4379
79 +2019,software_engineering,181111,0.3012
80 +2019,speech_audio,85969,0.1821
81 +2019,sustainable_computing,3832,0.7985
82 +2020,ai_accelerators,8555,0.5239
83 +2020,ai_agents,29472,0.5535
84 +2020,ai_safety,12179,0.538
85 +2020,algorithms,178532,0.3103
86 +2020,applied_cryptography,66281,0.408
87 +2020,autonomous_scientific_discovery,1271,0.8607
88 +2020,blockchain,59231,0.2082
89 +2020,cloud_computing,52776,0.4858
90 +2020,compilers,4503,0.8368
91 +2020,computational_biology,135533,0.1698
92 +2020,computational_complexity,21487,0.6388
93 +2020,computer_architecture,127745,0.3222
94 +2020,computer_graphics,60112,0.4218
95 +2020,computer_networks,417940,0.1881
96 +2020,computer_vision,577782,0.1761
97 +2020,cybersecurity,210760,0.3413
98 +2020,database_systems,98900,0.4658
99 +2020,deep_learning_architectures,120600,0.4763
100 +2020,distributed_computing,72448,0.5751
101 +2020,edge_computing,47740,0.5417
102 +2020,federated_learning,1722,0.7706
103 +2020,formal_verification,29104,0.4773
104 +2020,foundation_models,822,0.9051
105 +2020,generative_ai,10311,0.6379
106 +2020,hci,229085,0.2661
107 +2020,information_retrieval,73086,0.3568
108 +2020,llm,11303,0.7639
109 +2020,ml_theory,136518,0.4254
110 +2020,multimodal_ai,5280,0.7193
111 +2020,nlp,199186,0.2716
112 +2020,operating_systems,8975,0.7552
113 +2020,privacy_preserving_computation,26886,0.4459
114 +2020,programming_languages,25835,0.6379
115 +2020,quantum_computing,54478,0.1196
116 +2020,reinforcement_learning,41970,0.3203
117 +2020,robotics,159762,0.2454
118 +2020,scientific_computing,105432,0.4301
119 +2020,software_engineering,178500,0.2989
120 +2020,speech_audio,86196,0.1777
121 +2020,sustainable_computing,3712,0.7702
122 +2021,ai_accelerators,10185,0.4986
123 +2021,ai_agents,29211,0.5379
124 +2021,ai_safety,15348,0.5426
125 +2021,algorithms,185099,0.3075
126 +2021,applied_cryptography,65691,0.4139
127 +2021,autonomous_scientific_discovery,1520,0.8138
128 +2021,blockchain,68145,0.2024
129 +2021,cloud_computing,53243,0.4808
130 +2021,compilers,4166,0.8346
131 +2021,computational_biology,145854,0.1767
132 +2021,computational_complexity,21138,0.6401
133 +2021,computer_architecture,126775,0.3227
134 +2021,computer_graphics,59871,0.432
135 +2021,computer_networks,410121,0.1944
136 +2021,computer_vision,681479,0.1697
137 +2021,cybersecurity,221529,0.3434
138 +2021,database_systems,98155,0.4755
139 +2021,deep_learning_architectures,148163,0.4828
140 +2021,distributed_computing,69103,0.5739
141 +2021,edge_computing,55656,0.5292
142 +2021,federated_learning,4389,0.7471
143 +2021,formal_verification,30190,0.4816
144 +2021,foundation_models,1619,0.9036
145 +2021,generative_ai,13848,0.6419
146 +2021,hci,230614,0.2602
147 +2021,information_retrieval,75040,0.3477
148 +2021,llm,16560,0.7556
149 +2021,ml_theory,145324,0.4186
150 +2021,multimodal_ai,6653,0.7089
151 +2021,nlp,220164,0.2669
152 +2021,operating_systems,8703,0.7807
153 +2021,privacy_preserving_computation,29902,0.4454
154 +2021,programming_languages,25656,0.6097
155 +2021,quantum_computing,64423,0.1201
156 +2021,reinforcement_learning,50682,0.3234
157 +2021,robotics,174798,0.2535
158 +2021,scientific_computing,112748,0.4206
159 +2021,software_engineering,169864,0.3095
160 +2021,speech_audio,93912,0.1818
161 +2021,sustainable_computing,3960,0.7631
162 +2022,ai_accelerators,11638,0.4852
163 +2022,ai_agents,27771,0.5497
164 +2022,ai_safety,19034,0.5405
165 +2022,algorithms,181230,0.311
166 +2022,applied_cryptography,67510,0.4226
167 +2022,autonomous_scientific_discovery,1523,0.8772
168 +2022,blockchain,80994,0.2108
169 +2022,cloud_computing,51091,0.4913
170 +2022,compilers,3954,0.8462
171 +2022,computational_biology,150600,0.1813
172 +2022,computational_complexity,19809,0.6552
173 +2022,computer_architecture,120398,0.3229
174 +2022,computer_graphics,58221,0.4419
175 +2022,computer_networks,401999,0.199
176 +2022,computer_vision,718729,0.1678
177 +2022,cybersecurity,229807,0.3474
178 +2022,database_systems,96781,0.4772
179 +2022,deep_learning_architectures,159287,0.4821
180 +2022,distributed_computing,65426,0.589
181 +2022,edge_computing,62183,0.5375
182 +2022,federated_learning,6353,0.7244
183 +2022,formal_verification,26817,0.4724
184 +2022,foundation_models,2359,0.8978
185 +2022,generative_ai,15112,0.6619
186 +2022,hci,246850,0.2561
187 +2022,information_retrieval,77390,0.3335
188 +2022,llm,21803,0.7023
189 +2022,ml_theory,147928,0.4129
190 +2022,multimodal_ai,9216,0.6969
191 +2022,nlp,224621,0.2853
192 +2022,operating_systems,7330,0.7935
193 +2022,privacy_preserving_computation,32588,0.4538
194 +2022,programming_languages,24406,0.6179
195 +2022,quantum_computing,68971,0.1223
196 +2022,reinforcement_learning,54682,0.3177
197 +2022,robotics,184921,0.2512
198 +2022,scientific_computing,113486,0.4277
199 +2022,software_engineering,171531,0.3128
200 +2022,speech_audio,97128,0.1778
201 +2022,sustainable_computing,4700,0.7532
202 +2023,ai_accelerators,13046,0.4962
203 +2023,ai_agents,31675,0.5545
204 +2023,ai_safety,26172,0.5404
205 +2023,algorithms,184947,0.3131
206 +2023,applied_cryptography,74594,0.4263
207 +2023,autonomous_scientific_discovery,2873,0.8173
208 +2023,blockchain,101857,0.2044
209 +2023,cloud_computing,56168,0.4943
210 +2023,compilers,4189,0.835
211 +2023,computational_biology,159302,0.1853
212 +2023,computational_complexity,20196,0.6614
213 +2023,computer_architecture,131865,0.3229
214 +2023,computer_graphics,67297,0.4532
215 +2023,computer_networks,433815,0.2031
216 +2023,computer_vision,861207,0.1686
217 +2023,cybersecurity,274554,0.335
218 +2023,database_systems,103963,0.4807
219 +2023,deep_learning_architectures,184269,0.4815
220 +2023,distributed_computing,69801,0.5846
221 +2023,edge_computing,68563,0.5329
222 +2023,federated_learning,9556,0.7345
223 +2023,formal_verification,30583,0.4822
224 +2023,foundation_models,4193,0.9094
225 +2023,generative_ai,25933,0.6454
226 +2023,hci,288325,0.2557
227 +2023,information_retrieval,87299,0.328
228 +2023,llm,77315,0.51
229 +2023,ml_theory,171365,0.419
230 +2023,multimodal_ai,14035,0.702
231 +2023,nlp,259107,0.3093
232 +2023,operating_systems,8169,0.7975
233 +2023,privacy_preserving_computation,38950,0.4841
234 +2023,programming_languages,25578,0.63
235 +2023,quantum_computing,82434,0.1331
236 +2023,reinforcement_learning,65006,0.3221
237 +2023,robotics,209839,0.2523
238 +2023,scientific_computing,120013,0.4333
239 +2023,software_engineering,191926,0.3164
240 +2023,speech_audio,111431,0.1851
241 +2023,sustainable_computing,6610,0.7452
242 +2024,ai_accelerators,14257,0.5063
243 +2024,ai_agents,37986,0.5616
244 +2024,ai_safety,41126,0.5402
245 +2024,algorithms,202214,0.3203
246 +2024,applied_cryptography,84262,0.4261
247 +2024,autonomous_scientific_discovery,3282,0.848
248 +2024,blockchain,111052,0.2038
249 +2024,cloud_computing,64842,0.49
250 +2024,compilers,4755,0.8307
251 +2024,computational_biology,179730,0.1943
252 +2024,computational_complexity,22786,0.6901
253 +2024,computer_architecture,141510,0.322
254 +2024,computer_graphics,75793,0.4667
255 +2024,computer_networks,455866,0.2086
256 +2024,computer_vision,990283,0.1717
257 +2024,cybersecurity,344043,0.3253
258 +2024,database_systems,108454,0.4992
259 +2024,deep_learning_architectures,214093,0.4833
260 +2024,distributed_computing,75588,0.5921
261 +2024,edge_computing,77537,0.5292
262 +2024,federated_learning,12221,0.7153
263 +2024,formal_verification,29732,0.4824
264 +2024,foundation_models,7719,0.8851
265 +2024,generative_ai,49797,0.6391
266 +2024,hci,351074,0.2598
267 +2024,information_retrieval,99364,0.3423
268 +2024,llm,191956,0.4647
269 +2024,ml_theory,200633,0.4223
270 +2024,multimodal_ai,25075,0.72
271 +2024,nlp,296010,0.3364
272 +2024,operating_systems,8522,0.7941
273 +2024,privacy_preserving_computation,51216,0.4972
274 +2024,programming_languages,25672,0.648
275 +2024,quantum_computing,99990,0.1384
276 +2024,reinforcement_learning,77161,0.3317
277 +2024,robotics,248034,0.261
278 +2024,scientific_computing,132038,0.4453
279 +2024,software_engineering,222332,0.3241
280 +2024,speech_audio,123868,0.1872
281 +2024,sustainable_computing,7770,0.739
282 +2025,ai_accelerators,15217,0.5092
283 +2025,ai_agents,56916,0.6012
284 +2025,ai_safety,69356,0.5776
285 +2025,algorithms,228420,0.3291
286 +2025,applied_cryptography,98047,0.4293
287 +2025,autonomous_scientific_discovery,5478,0.8058
288 +2025,blockchain,121711,0.2119
289 +2025,cloud_computing,79661,0.4989
290 +2025,compilers,4902,0.8413
291 +2025,computational_biology,211631,0.2004
292 +2025,computational_complexity,25331,0.6998
293 +2025,computer_architecture,163238,0.3277
294 +2025,computer_graphics,87276,0.4866
295 +2025,computer_networks,496882,0.2216
296 +2025,computer_vision,1109954,0.1822
297 +2025,cybersecurity,435890,0.3327
298 +2025,database_systems,120604,0.5151
299 +2025,deep_learning_architectures,245500,0.5031
300 +2025,distributed_computing,89968,0.6073
301 +2025,edge_computing,85425,0.5495
302 +2025,federated_learning,13772,0.7339
303 +2025,formal_verification,32650,0.4987
304 +2025,foundation_models,13678,0.8794
305 +2025,generative_ai,82172,0.6386
306 +2025,hci,402677,0.2858
307 +2025,information_retrieval,121877,0.374
308 +2025,llm,343828,0.4589
309 +2025,ml_theory,245517,0.4307
310 +2025,multimodal_ai,48170,0.7293
311 +2025,nlp,324909,0.374
312 +2025,operating_systems,9257,0.8159
313 +2025,privacy_preserving_computation,66195,0.5222
314 +2025,programming_languages,27870,0.6435
315 +2025,quantum_computing,133108,0.1459
316 +2025,reinforcement_learning,99228,0.3679
317 +2025,robotics,269974,0.2587
318 +2025,scientific_computing,143714,0.4435
319 +2025,software_engineering,267473,0.3336
320 +2025,speech_audio,137912,0.2049
321 +2025,sustainable_computing,9666,0.7883
322 +2026,ai_accelerators,10745,0.5301
323 +2026,ai_agents,74557,0.6445
324 +2026,ai_safety,75373,0.6246
325 +2026,algorithms,192827,0.3229
326 +2026,applied_cryptography,67603,0.4474
327 +2026,autonomous_scientific_discovery,6582,0.8004
328 +2026,blockchain,69117,0.2277
329 +2026,cloud_computing,46035,0.5551
330 +2026,compilers,4061,0.8527
331 +2026,computational_biology,170553,0.1955
332 +2026,computational_complexity,24972,0.6565
333 +2026,computer_architecture,100060,0.3776
334 +2026,computer_graphics,68366,0.486
335 +2026,computer_networks,261852,0.2593
336 +2026,computer_vision,730253,0.2052
337 +2026,cybersecurity,311429,0.3431
338 +2026,database_systems,90810,0.5413
339 +2026,deep_learning_architectures,172458,0.5203
340 +2026,distributed_computing,63582,0.6299
341 +2026,edge_computing,47987,0.5856
342 +2026,federated_learning,10721,0.752
343 +2026,formal_verification,32567,0.5272
344 +2026,foundation_models,19258,0.8476
345 +2026,generative_ai,84628,0.6465
346 +2026,hci,313063,0.303
347 +2026,information_retrieval,109436,0.4269
348 +2026,llm,372690,0.4594
349 +2026,ml_theory,208330,0.4467
350 +2026,multimodal_ai,55739,0.7366
351 +2026,nlp,242621,0.4251
352 +2026,operating_systems,6866,0.8504
353 +2026,privacy_preserving_computation,48458,0.5405
354 +2026,programming_languages,23414,0.635
355 +2026,quantum_computing,140037,0.129
356 +2026,reinforcement_learning,86233,0.4415
357 +2026,robotics,194894,0.2596
358 +2026,scientific_computing,125821,0.4213
359 +2026,software_engineering,230827,0.3578
360 +2026,speech_audio,109792,0.2175
361 +2026,sustainable_computing,6619,0.7892
362
Menu

Theme
↑↓ navigate ↵ open esc close / or ⌘K anywhere