Pith. sign in

REVIEW 3 major objections 65 references

Large language models can clean noisy historical OCR and turn rigid keyword archives into natural-language searchable systems with measurable gains in ranking and answer quality.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

LLM post-OCR cleanup plus dense retrieval and cross-encoder reranking raises NDCG@10 by 31.9% over BM25 on 500k Swiss newspaper segments while cutting CER/WER substantially.

T0 review reviewed 2026-07-12 challenge →

load-bearing objection Solid end-to-end digital-library systems paper; the big NDCG lift is mostly hybrid+rerank, not OCR cleaning, and the OCR numbers rest on 94 snippets. the 3 major comments →

arxiv 2607.03440 v1 pith:XCPUAQY5 submitted 2026-07-03 cs.IR

Improving Access to Historical Archives with Real-time RAG-based Systems

Stergios Konstantinidis , Hayman Lotfy , Alexis Erne , Faruk Zahiragic , Min-Yen Kan , Michalis Vlachos This is my paper
classification cs.IR
keywords large language modelsdigital librariesretrieval-augmented generationOCRhistorical archivessemantic retrievalcross-encoder rerankingnewspaper digitization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Digitized historical newspapers are full of OCR errors and only support brittle keyword search, so much of the material stays hard to find. This paper shows that an end-to-end pipeline can change that: instruction-tuned language models first repair the OCR text, then dense retrieval plus cross-encoder reranking feed a retrieval-augmented generation system that answers ordinary questions in natural language. On half a million Swiss newspaper segments spanning 1762–2001 and 384 test queries, the best models cut character and word error rates by roughly 45% and 61% on average, while the full retrieval stack raises NDCG@10 from about 66% to 87% against a keyword baseline and improves answer correctness and context relevance with statistical significance. The authors also keep response times interactive by using approximate nearest-neighbor search, lighter rerankers, and streaming generation. A sympathetic reader cares because the work turns static cultural-heritage repositories into systems people can actually explore with ordinary questions.

Core claim

Integrating large language models for post-OCR refinement with a semantic retrieval, cross-encoder reranking, and RAG generation pipeline measurably reduces transcription error and improves ranking and answer quality on a real multi-century newspaper archive, elevating digital libraries from keyword-only repositories to interactive, semantically searchable systems.

What carries the argument

The end-to-end archival pipeline: LLM post-OCR sanitization of article text, followed by hybrid dense-plus-BM25 retrieval over approximate nearest-neighbor indexes, cross-encoder reranking of the candidate pool, and grounded answer generation. The reranker and streaming generation are the levers that trade quality against interactive latency.

Load-bearing premise

That automatic LLM-as-judge scores on a few hundred author-written queries, plus error rates measured on only about a hundred manually transcribed snippets, reliably stand in for real human relevance and for OCR noise across the full multi-century collection.

What would settle it

Re-run the same 384 queries (or a larger independent human-annotated set) with human raters scoring answer correctness and context relevance, and recompute CER/WER on a much larger random sample of the 500,000 segments; if the NDCG@10 lift and the error reductions disappear or reverse under human judgment, the central claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Digital libraries can replace or augment pure keyword search with natural-language question answering grounded in the original pages.
  • OCR quality for historical material need not be perfect if a strong reranker and grounded generator are present; cleaning still helps factual correctness.
  • Latency-aware design (lighter cross-encoders, approximate indexes, streaming tokens) makes real-time exploration of large noisy archives practical.
  • Institutions can surface previously undiscoverable content whose wording no longer matches modern query language.
  • Immersive interfaces that show supporting evidence and map places become feasible once answers are both fast and source-grounded.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same pipeline could be applied to other noisy heritage collections (handwritten letters, maps, multilingual colonial newspapers) without waiting for perfect human transcription.
  • Selective correction—running the expensive LLM only on high-error or high-value segments—could cut cost while preserving most of the retrieval gain.
  • Hybrid lexical-plus-dense indexes plus metadata filters (date, place, entities) may become the default access layer for any multi-century text archive.
  • Because smaller models often degrade OCR, production systems will likely keep frontier models for sanitization even if lighter models handle generation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. The paper presents an end-to-end pipeline for historical newspaper archives that (i) applies instruction-tuned LLMs as post-OCR refinement operators and (ii) supports natural-language QA via dense retrieval, hybrid BM25+dense retrieval, cross-encoder reranking, and RAG generation. On a 500k-segment Swiss newspaper corpus (1762–2001) and 384 author-written queries, LLM refinement reduces CER/WER by up to 44.52%/60.95% on a 94-snippet ground-truth sample (Tables 1–2). The full hybrid+reranker pipeline raises NDCG@10 from a BM25 baseline of 65.99% to 87.05% and yields statistically significant gains in RAGAS answer correctness and context relevance (Table 3). A latency-aware design (lightweight MiniLM reranker + streaming generation) and an Apple Vision Pro prototype are also demonstrated.

Significance. If the results hold, the work supplies a practical blueprint for modernizing digital libraries: LLM post-OCR sanitization plus hybrid retrieval and reranking can turn noisy, keyword-only heritage collections into interactive, semantically searchable systems under interactive latency budgets. Strengths include a real multi-century corpus, systematic ablations across BM25/dense/hybrid/reranked settings at multiple k with 95% CIs and paired t-tests, multi-engine OCR and multi-LLM comparisons, and an explicit quality–latency analysis. The contribution is primarily systems-and-evaluation rather than new algorithms, but the end-to-end framing and public-facing prototype are valuable for the digital-libraries and IR communities.

major comments (3)
  1. Abstract and §Impact on Information Retrieval package two results as a single causal story: LLM OCR refinement “is accompanied by” the 31.9% NDCG@10 lift (65.99%→87.05%). Table 3 shows that Semantic+BM25s+reranker on raw OCR already reaches 84.94% NDCG@10; the LLM-corrected view adds only ~2 absolute points (87.05%). The dominant IR gains therefore come from hybrid retrieval + reranking, not from OCR sanitization. The abstract and discussion should separate these effects and avoid implying that the large CER/WER reductions are the primary driver of the NDCG jump.
  2. Tables 1–2 report CER/WER on only 94 manually transcribed snippets, while the retrieval index contains 500k segments spanning three centuries of heterogeneous fonts, layouts and degradation. No evidence is given that this sample is representative of the full corpus error distribution. Because the paper’s second core claim is that OCR improvements propagate to downstream IR, either enlarge the ground-truth sample, stratify by period/periodical, or clearly qualify the OCR numbers as a limited pilot.
  3. Answer correctness, context relevance and (implicitly) graded relevance for NDCG rest on RAGAS LLM-as-judge (GPT-4o-mini) plus 384 author-written queries of which only five are unanswerable. For historical material this is a weak proxy for human relevance and factual fidelity. At minimum, a human-judged subset (or inter-annotator agreement against the LLM judge) is needed before the “statistically significant gains in answer correctness and context relevance” can be treated as load-bearing.

Circularity Check

0 steps flagged

No circularity: empirical systems paper whose CER/WER, NDCG and RAGAS results are measured against external human transcripts and held-out queries, not forced by construction from fitted inputs or self-citation.

full rationale

The paper is an end-to-end empirical evaluation of LLM post-OCR refinement plus dense+BM25+cross-encoder RAG on a 500k-segment Swiss newspaper corpus. Load-bearing claims (Tables 1-2 CER/WER reductions vs. 94 manually transcribed snippets; Table 3 NDCG@10 / answer correctness / context relevance on 384 author-written queries) rest on external ground-truth transcripts and standard IR/RAGAS metrics; none of the reported numbers is obtained by fitting a free parameter and then re-predicting a quantity that is definitionally identical to the fit. There is no uniqueness theorem, ansatz, or self-citation chain that forces the central result. Minor use of an LLM-as-judge (RAGAS with GPT-4o-mini) is a conventional evaluation practice, not a mathematical reduction of the claim to its own inputs. Consequently the derivation chain is self-contained against external benchmarks and exhibits zero circularity of the kinds enumerated.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

Systems/empirical IR paper. Load-bearing choices are engineering and evaluation assumptions rather than physical axioms: which OCR engines and LLMs are used, the size of the rerank pool, the embedding and judge models, and the claim that automatic RAGAS scores plus a small hand-transcribed OCR sample generalize. No new physical entities are postulated.

free parameters (5)
  • reranker candidate pool size L = 50
    Set to 50 for deployed latency/quality trade-off; directly controls NDCG vs. response time (Figures 5–6).
  • retrieval depth k = 5/10/25
    Reported at 5, 10, 25; primary headline NDCG uses k=10.
  • RAGAS correctness weights = 0.75 / 0.25
    Default wf=0.75 factual, ws=0.25 semantic similarity used for answer correctness.
  • OCR ground-truth sample size = 94 snippets
    CER/WER claims rest on 94 manually transcribed snippets, not the full 500k index.
  • query set size and composition = 384 queries
    384 author-formulated natural-language queries including 5 adversarial unanswerable items; no public multi-annotator relevance qrels.
axioms (5)
  • domain assumption Instruction-tuned LLMs can reduce OCR transcription error while preserving semantic and structural fidelity of historical text when prompted with length/format constraints.
    Core ingestion premise in Methodology (Sanitizing OCR Errors) and Tables 1–2; smaller models are shown to violate it by over-generating.
  • domain assumption Cosine similarity of dense embeddings plus cross-encoder scores is a valid proxy for semantic relevance on noisy historical newspaper text.
    Underpins the RAG formulation and NDCG evaluation throughout Methodology and Table 3.
  • domain assumption RAGAS LLM-as-judge metrics (answer correctness, context/answer relevancy) track human notions of quality for archival QA.
    Primary evaluation of answer quality in the absence of full human ground-truth pairs (Quality of answers section).
  • domain assumption Approximate nearest-neighbor graph indexing preserves sufficient recall for the subsequent reranker on this corpus.
    Stated in Methodology for scalable retrieval; not independently ablated for recall@k of the ANN stage alone.
  • standard math Standard IR ranking metrics (NDCG@k) and paired t-tests on the 384-query set support claims of statistically significant improvement.
    Used in Table 3 significance markers.

reviewed 2026-07-12 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Access to Historical Archives with Real-time RAG-based Systems." pith.science (2026). https://pith.science/paper/XCPUAQY5

@misc{pith2026260703440,
  author       = {Pith},
  title        = {Pith review of: Improving Access to Historical Archives with Real-time RAG-based Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XCPUAQY5}},
  note         = {Machine review of arXiv:2607.03440}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Digitized historical archives are large, heterogeneous cultural heritage repositories, but access methods for such archives face challenges such as noisy optical character recognition (OCR) output and rigid keyword-based retrieval, which limit retrieval quality. In this work, we present an end-to-end archival processing and retrieval framework that integrates large language models (LLMs) into the archival pipeline. Our system introduces two core components: (i) an LLM-based OCR refinement module that improves text quality, and (ii) a semantic retrieval and cross-encoder reranking pipeline supporting natural-language question answering via retrieval-augmented generation (RAG). Our evaluations are done on a historical archival dataset of 500,000 Swiss newspaper segments spanning over three centuries (1762 to 2001). Experiments are conducted across 384 natural-language test queries. Our results highlight that LLM refinements reduce OCR errors by up to 44.52% (CER) and 60.95% (WER). More importantly, this is accompanied by downstream information retrieval improvements. Compared to traditional keyword baselines, our reranking pipeline increases NDCG@10 by 31.9% (from 65.99% to 87.05%) and achieves statistically significant gains in both answer correctness and context relevance. These results demonstrate that integrating LLMs with established document processing and retrieval pipelines can elevate digital libraries from static repositories to interactive, semantically searchable archival systems.

Figures

Figures reproduced from arXiv: 2607.03440 by Alexis Erne, Faruk Zahiragic, Hayman Lotfy, Michalis Vlachos, Min-Yen Kan, Stergios Konstantinidis.

Figure 1
Figure 1. Figure 1: Top: Typical archival access pipelines rely primarily on (A) OCR and (B) keyword [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Indicative allocation of computation time across the stages of a RAG pipeline. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Examples of the datasets used. for each segment, including a synthesized title, a concise summary, and extracted keywords. Named Entity Recognition (NER): Categorized lists of entities mentioned in the text, such as Persons, Locations, Organizations, and Events. This rich metadata structure allows our system to perform hybrid search strategies—combining dense vector retrieval with hard meta￾data filtering.… view at source ↗
Figure 4
Figure 4. Figure 4: Example: Raw OCR (Tesseract) vs. raw OCR with LLM refinement (Gemini 3 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Reranking trade-off between quality (NDCG@10) and mean time (candidate doc [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Deployed system, reducing system response time by changing the reranker and [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Examples of two queries, the prototypical and the actual reply by the RAG system. [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prototypical interface using the Vision Pro augmented reality device [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

65 extracted references · 16 canonical work pages

  1. [1]

    Raff, Edward and Nicholas, Charles , year =

  2. [2]

    2023 , organization =

    Huang, Qiang and Tung, Anthony KH , booktitle =. 2023 , organization =

    2023

  3. [3]

    2023 , organization =

    Elkin, Yury and Kurlin, Vitaliy , booktitle =. 2023 , organization =

    2023

  4. [4]

    2022 , publisher =

    Gu, Yuanqi and Chen, Qun and others , booktitle =. 2022 , publisher =

    2022

  5. [5]

    2022 , publisher =

    Nockels, Joe and Gooding, Paul and Ames, Sarah and Terras, Melissa , journal =. 2022 , publisher =

    2022

  6. [6]

    2023 , organization =

    2023

  7. [7]

    2011 , publisher =

    Novak, David and Batko, Michal and Zezula, Pavel , journal =. 2011 , publisher =

    2011

  8. [8]

    , title =

    Yianilos, Peter N. , title =. Proceedings of the Fourth Annual ACM-SIAM Symposium on Discrete Algorithms , pages =. 1993 , isbn =

    1993

  9. [9]

    The research on nearest neighbor search algorithm based on vantage point tree , year=

    Jiang, Daguang and Sun, Hejuan and Yi, Junkai and Zhao, Xianghui , booktitle=. The research on nearest neighbor search algorithm based on vantage point tree , year=

  10. [10]

    , title =

    Lee, Benjamin Charles Germain and Mears, Jaime and Jakeway, Eileen and Ferriter, Meghan and Adams, Chris and Yarasavage, Nathan and Thomas, Deborah and Zwaard, Kate and Weld, Daniel S. , title =. Proceedings of the 29th ACM International Conference on Information & Knowledge Management , pages =. 2020 , isbn =. doi:10.1145/3340531.3412767 , abstract =

    work page doi:10.1145/3340531.3412767 2020

  11. [11]

    2021 , volume =

    Johnson, Jeff and Douze, Matthijs and Jégou, Hervé , journal =. 2021 , volume =

    2021

  12. [12]

    IEEE International Conference on Big Data , year=

    Making History Readable , author=. IEEE International Conference on Big Data , year=

  13. [13]

    Digital Scholarship in the Humanities , volume =

    Jaillant, Lise and Rees, Arran , title =. Digital Scholarship in the Humanities , volume =. 2023 , month =

    2023

  14. [14]

    2025 , url =

    Zhang, Hao , title =. 2025 , url =. doi:10.1007/978-981-95-4861-3_40 , booktitle =

    work page doi:10.1007/978-981-95-4861-3_40 2025

  15. [15]

    Reflections on Tools, Methods and Epistemology , author =

    Digitised Historical Newspapers: A Changing Research Landscape , booktitle =. Reflections on Tools, Methods and Epistemology , author =. 2023 , lastchecked =. doi:doi:10.1515/9783110729214-001 , isbn =

    work page doi:10.1515/9783110729214-001 2023

  16. [16]

    Dagstuhl Reports , ISSN =

    Ehrmann, Maud and D\". Dagstuhl Reports , ISSN =. 2023 , volume =

    2023

  17. [17]

    Post-Correction of Historical Text Transcripts with Large Language Models: An Exploratory Study

    Boros, Emanuela and Ehrmann, Maud and Romanello, Matteo and Najem-Meyer, Sven and Kaplan, Fr \'e d \'e ric. Post-Correction of Historical Text Transcripts with Large Language Models: An Exploratory Study. Proceedings of the 8th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature. 2024. doi:1...

    work page doi:10.18653/v1/2024.latechclfl-1.14 2024

  18. [18]

    Docling Technical Report , url =

    Deep Search Team , month =. Docling Technical Report , url =. 2024 , institution =. 2408.09869 , doi =

    Pith/arXiv arXiv 2024

  19. [19]

    ICDAR 2023 Competition on Robust Layout Segmentation in Corporate Documents

    Auer, Christoph and Nassar, Ahmed and Lysak, Maksym and Dolfi, Michele and Livathinos, Nikolaos and Staar, Peter. ICDAR 2023 Competition on Robust Layout Segmentation in Corporate Documents. Document Analysis and Recognition - ICDAR 2023. 2023

    2023

  20. [20]

    ICDAR 2019 Competition on Post-OCR Text Correction , year=

    Rigaud, Christophe and Doucet, Antoine and Coustaty, Mickaël and Moreux, Jean-Philippe , booktitle=. ICDAR 2019 Competition on Post-OCR Text Correction , year=

    2019

  21. [21]

    EACL 2024 - 18th Conference of the European Chapter of the Association for Computational Linguistics, Proceedings of System Demonstrations , pages =

    Shahul Es and Jithin James and Luis Espinosa-Anke and Steven Schockaert , doi =. EACL 2024 - 18th Conference of the European Chapter of the Association for Computational Linguistics, Proceedings of System Demonstrations , pages =

    2024

  22. [22]

    2026 , eprint=

    A Survey of Large Language Models , author=. 2026 , eprint=

    2026

  23. [23]

    Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =

    Fan, Wenqi and Ding, Yujuan and Ning, Liangbo and Wang, Shijie and Li, Hengyun and Yin, Dawei and Chua, Tat-Seng and Li, Qing , title =. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , pages =. 2024 , isbn =. doi:10.1145/3637528.3671470 , abstract =

    doi:10.1145/3637528.3671470 2024

  24. [24]

    Combining Visual and Textual Features for Semantic Segmentation of Historical Newspapers , volume =

    Barman, Raphaël and Ehrmann, Maud and Clematide, Simon and Oliveira, Sofia and Kaplan, Frederic , year =. Combining Visual and Textual Features for Semantic Segmentation of Historical Newspapers , volume =. Journal of Data Mining & Digital Humanities , doi =

  25. [25]

    Chowdhery, Aakanksha and Narang, Sharan and Devlin, Jacob and Bosma, Maarten and Mishra, Gaurav and Roberts, Adam and Barham, Paul and Chung, Hyung Won and Sutton, Charles and Gehrmann, Sebastian and Schuh, Parker and Shi, Kensen and Tsvyashchenko, Sashank and Maynez, Joshua and Rao, Abhishek and Barnes, Parker and Tay, Yi and Shazeer, Noam and Prabhakara...

  26. [26]

    2023 , eprint=

    Artificial Intelligence in archival and historical scholarship workflow: HTS and ChatGPT , author=. 2023 , eprint=

    2023

  27. [27]

    Leveraging LLM s for Post- OCR Correction of Historical Newspapers

    Thomas, Alan and Gaizauskas, Robert and Lu, Haiping. Leveraging LLM s for Post- OCR Correction of Historical Newspapers. Proceedings of the Third Workshop on Language Technologies for Historical and Ancient Languages (LT4HALA)

  28. [29]

    2022 , month=

    Digitised Newspapers – A New Eldorado for Historians? Reflections on Tools, Methods and Epistemology , url=. 2022 , month=. doi:10.1515/9783110729214 , abstractNote=

    work page doi:10.1515/9783110729214 2022

  29. [30]

    Davis, Corey , publisher =

  30. [31]

    Historical Ink: 19th Century L atin A merican S panish Newspaper Corpus with LLM OCR Correction

    Manrique-Gomez, Laura and Montes, Tony and Rodriguez Herrera, Arturo and Manrique, Ruben. Historical Ink: 19th Century L atin A merican S panish Newspaper Corpus with LLM OCR Correction. Proceedings of the 4th International Conference on Natural Language Processing for Digital Humanities. 2024. doi:10.18653/v1/2024.nlp4dh-1.13

    work page doi:10.18653/v1/2024.nlp4dh-1.13 2024

  31. [32]

    Scancella, John , year =

  32. [33]

    Unifying Large Language Models and Knowledge Graphs: A Roadmap , year=

    Pan, Shirui and Luo, Linhao and Wang, Yufei and Chen, Chen and Wang, Jiapu and Wu, Xindong , journal=. Unifying Large Language Models and Knowledge Graphs: A Roadmap , year=

  33. [34]

    and Koerich, Alessandro and Badue, Claudine and De Souza, Alberto and Oliveira-Santos, Thiago , year =

    Paixão, Thiago and Berriel, Rodrigo and Boeres, M. and Koerich, Alessandro and Badue, Claudine and De Souza, Alberto and Oliveira-Santos, Thiago , year =. Fast(er) Reconstruction of Shredded Text Documents via Self-Supervised Deep Asymmetric Metric Learning , doi =

  34. [35]

    Reconstruction of shredded document based on image feature matching , journal =

    Huei-Yung Lin and Wen-Cheng Fan-Chiang , keywords =. Reconstruction of shredded document based on image feature matching , journal =. 2012 , issn =. doi:10.1016/j.eswa.2011.09.019 , url =

    work page doi:10.1016/j.eswa.2011.09.019 2012

  35. [36]

    Word Form Matters: LLM s' Semantic Reconstruction under Typoglycemia

    Wang, Chenxi and Gu, Tianle and Wei, Zhongyu and Gao, Lang and Song, Zirui and Chen, Xiuying. Word Form Matters: LLM s' Semantic Reconstruction under Typoglycemia. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.866

    work page doi:10.18653/v1/2025.findings-acl.866 2025

  36. [37]

    Sandhaus, Jane Cotler and Evan , year =

  37. [38]

    LayoutParser: A Unified Toolkit for Deep Learning Based Document Image Analysis

    Shen, Zejiang and Zhang, Ruochen and Dell, Melissa and Lee, Benjamin Charles Germain and Carlson, Jacob and Li, Weining. LayoutParser: A Unified Toolkit for Deep Learning Based Document Image Analysis. Document Analysis and Recognition -- ICDAR. 2021

    2021

  38. [39]

    Do, Thao and Tran, Dinh Phu and Vo, An and Kim, Daeyoung , title =. Proceedings of the Thirty-Ninth AAAI Conference on Artificial Intelligence and Thirty-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence , articleno =. 2025 , isbn =. doi:10.1609/aaai.v39i27.3...

    work page doi:10.1609/aaai.v39i27.35012 2025

  39. [40]

    and Staar, Peter , title =

    Pfitzmann, Birgit and Auer, Christoph and Dolfi, Michele and Nassar, Ahmed S. and Staar, Peter , title =. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2022 , isbn =. doi:10.1145/3534678.3539043 , abstract =

    doi:10.1145/3534678.3539043 2022

  40. [41]

    , title =

    Smith, R. , title =. Proceedings of the Ninth International Conference on Document Analysis and Recognition , pages =. 2007 , isbn =

    2007

  41. [42]

    2024 , doi =

    Kaufmann, Timo and Weng, Paul and Bengs, Viktor and H. 2024 , doi =

    2024

  42. [43]

    A Comprehensive Survey of LLM Alignment Techniques: RLHF, RLAIF, PPO, DPO and More , doi =

    Wang, Zhichao and Bi, Bin and Pentyala, Shiva and Ramnath, Kiran and Chaudhuri, Sougata and Mehrotra, Shubham and Zhu, Zixu and Mao, Xiang-Bo and Asur, Sitaram and Cheng , year =. A Comprehensive Survey of LLM Alignment Techniques: RLHF, RLAIF, PPO, DPO and More , doi =

  43. [44]

    2025 , eprint=

    A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications , author=. 2025 , eprint=

    2025

  44. [45]

    Gemma 3 Technical Report , year =

  45. [46]

    Gemini 2.5 Pro Model Card , year =

  46. [47]

    Gemini 3 Pro Model Card , year =

  47. [48]

    Gemini 3 Flash Model Card , year =

  48. [49]

    2025 , url =

    Zhang, Junyuan and Zhang, Qintong and Wang, Bin and Ouyang, Linke and Wen, Zichen and Li, Ying and Chow, Ka-Ho and He, Conghui and Zhang, Wentao , booktitle =. 2025 , url =

    2025

  49. [50]

    2026 , publisher =

    Song, Zhao Xun and Cheung, Kwok Wai and Jia, Zi Yun , journal =. 2026 , publisher =

    2026

  50. [51]

    2025 , doi =

    Levchenko, Maria A , booktitle =. 2025 , doi =

    2025

  51. [52]

    Vesalainen, Ari and Mäkelä, Eetu and Ruotsalainen, Laura and Tolonen, Mikko , year =. Error. doi:10.48550/arXiv.2602.14524 , abstract =

    doi:10.48550/arxiv.2602.14524

  52. [53]

    Third International Conference on Big Data, Computational Intelligence, and Applications , publisher =

    Ying Cao , title =. Third International Conference on Big Data, Computational Intelligence, and Applications , publisher =. 2026 , doi =

    2026

  53. [54]

    International Conference on Advances in Computer Vision Research and Applications (ACVRA) , abstract =

    Zhouzhen Shi and Yi Chen , title =. International Conference on Advances in Computer Vision Research and Applications (ACVRA) , abstract =. 2025 , doi =

    2025

  54. [55]

    Post- OCR Correction of Historical G erman Periodicals using LLM s

    Danilova, Vera and Aangenendt, Gijs. Post- OCR Correction of Historical G erman Periodicals using LLM s. Proceedings of the Third Workshop on Resources and Representations for Under-Resourced Languages and Domains (RESOURCEFUL-2025). 2025

    2025

  55. [56]

    2023 , isbn =

    Ward, Austin and Avula, Sandeep and Cheng, Hao-Fei and Sarwar, Sheikh Muhammad and Murdock, Vanessa and Agichtein, Eugene , title =. 2023 , isbn =. doi:10.1145/3539618.3592057 , booktitle =

    work page doi:10.1145/3539618.3592057 2023

  56. [57]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Bruch, Sebastian and Nardini, Franco Maria and Rulli, Cosimo and Venturini, Rossano , title =. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2024 , isbn =. doi:10.1145/3626772.3657769 , abstract =

    doi:10.1145/3626772.3657769 2024

  57. [58]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Li, Xiaoxi and Dou, Zhicheng and Zhou, Yujia and Liu, Fangchao , title =. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2024 , isbn =. doi:10.1145/3626772.3657778 , abstract =

    doi:10.1145/3626772.3657778 2024

  58. [59]

    Insights Into the Usability of a Digital Art Archive in Mixed Reality , year =

    Martins, Tiago and Rom\. Insights Into the Usability of a Digital Art Archive in Mixed Reality , year =. Proceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems , articleno =. doi:10.1145/3706599.3720092 , abstract =

    work page doi:10.1145/3706599.3720092

  59. [60]

    Information Processing & Management , volume =

    Zahra Elmi , keywords =. Information Processing & Management , volume =. 2026 , issn =. doi:10.1016/j.ipm.2026.104861 , url =

    work page doi:10.1016/j.ipm.2026.104861 2026

  60. [61]

    A hierarchical RL model for Chinese medical NER with radical semantics and dynamic decision , journal =

    Guangce Ruan and Lei Xia , keywords =. A hierarchical RL model for Chinese medical NER with radical semantics and dynamic decision , journal =. 2026 , issn =. doi:10.1016/j.ipm.2026.104784 , url =

    work page doi:10.1016/j.ipm.2026.104784 2026

  61. [62]

    Journal of the Association for Information Science and Technology , month =

    Struggling with digitized historical newspapers: Contextual barriers to information interaction in history research activities , author =. Journal of the Association for Information Science and Technology , month =. doi:10.1002/ASI.24608 , issn =

    work page doi:10.1002/asi.24608

  62. [63]

    Journal of the Association for Information Science and Technology , volume =

    Integrated interdisciplinary workflows for research on historical newspapers: Perspectives from humanities scholars, computer scientists, and librarians , author =. Journal of the Association for Information Science and Technology , volume =. 2022 , doi =

    2022

  63. [64]

    A gated multi-scale attention framework for automated short answer grading in Turkish , journal =

    Mustafa Kaya and Mahir Kavun , keywords =. A gated multi-scale attention framework for automated short answer grading in Turkish , journal =. 2026 , issn =. doi:10.1016/j.ipm.2026.104791 , url =

    work page doi:10.1016/j.ipm.2026.104791 2026

  64. [65]

    Veritas: Structuring and verifying LLM knowledge for logically consistent behavior tree generation in LLM-based agents , journal =

    Hao Tang and Feng Zhang and Kejia Wan and Songyi Lu and Xi Zhang and Yiping Yao and Xinhai Xu , keywords =. Veritas: Structuring and verifying LLM knowledge for logically consistent behavior tree generation in LLM-based agents , journal =. 2026 , issn =. doi:10.1016/j.ipm.2026.104830 , url =

    work page doi:10.1016/j.ipm.2026.104830 2026

  65. [66]

    Automatic Short-Answer Grading via BERT-Based Deep Neural Networks , year=

    Zhu, Xinhua and Wu, Han and Zhang, Lanfang , journal=. Automatic Short-Answer Grading via BERT-Based Deep Neural Networks , year=

This paper was first reviewed by grok-4.5 on July 12, 2026.