PaperlayerView exact revision on arXiv

Source-linked AI summary

ReviewPilot: A Multi-Agent System for Screening and Data Extraction in Systematic Reviews

Zhaoqian Xue, Shaoze Zhou, Xuexin Li, Junting Mu, Haozheng Liu, Zhongyin Wang, Yichun He, Renkai Ma, Hao Zhou, Sarah Bratt, Siyuan Ma, Chen Chen, Lingyao Li

arXiv:2610.21042v1cs.CLcs.AIcs.IR

TL;DR

Systematic reviews demand extensive screening and extraction, while many literature tools do not document exclusions or extracted-value sources. ReviewPilot coordinates a configurable, PRISMA-guided workflow; on PERG it outperforms two baselines, though its evaluation does not establish performance on unseen reviews.

  • Problem

    Systematic reviews require labor-intensive screening and extraction, while deep research tools do not record which studies were excluded or where extracted values come from.

    Source: §1 · p. 1

  • Method

    ReviewPilot coordinates a PRISMA-guided workflow from search through extraction, with dialogue-configured criteria and schemas and project memory for configurations and corrections.

    Source: §2 · p. 3

  • Results

    ReviewPilot’s screening macro F1 exceeds AgentSLR’s by 5.3–10.3 points and the single-prompt LLM’s by 7.0–19.0 points on PERG.

    Source: §4.2 · p. 6

  • Takeaways & Limitations

    ReviewPilot’s workflow records screening reasons and links extracted values to source passages for inspection.

    Source: §1 · pp. 1–2

    PERG pools informed system development, so the reported gains do not establish performance on unseen reviews.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Systematic literature reviews follow the PRISMA process of screening scholarly works and extracting evidence from included studies. Deep research tools based on large language models (LLMs) can search and summarize the literature but do not record which studies are excluded or where extracted values come from. We develop a conversational multi-agent system called ReviewPilot that automates the PRISMA workflow from paper screening to information extraction. Researchers can tailor criteria and schemas through dialogue, and a project memory keeps their configurations and revisions. Each screening decision comes with its reason and each extracted value with its source. On epidemiological expert reviews, ReviewPilot outperforms a single-prompt LLM and AgentSLR across three LLMs, with screening macro F1 up to 0.776 and extraction F1 up to 0.788. Case studies in biomedicine and human-computer interaction show close agreement with human annotators, suggesting that ReviewPilot can transfer beyond epidemiology. Code is available at https://github.com/ZhaoqianXue/ReviewPilot.

1 Introduction

Systematic reviews require labor-intensive screening and extraction while preserving inspectable, PRISMA-aligned decisions. ReviewPilot addresses gaps in existing research tools with a configurable workflow that records decisions and their sources.

1 Introduction

  • Reviewers may read thousands of titles and abstracts to find eligible studies, and a review can take over a year from registration to publication.Automation can reduce workload, but researchers still need to inspect decisions and document the process based on PRISMA 2020.Source: §1 · p. 1
  • Deep research tools can search and summarize literature but do not document which papers were screened or excluded, why, or where extracted values come from.Other tools often cover only screening, extraction, or a few review stages.Source: §1 · p. 1
  • Review workflows need operational instructions for criteria and field definitions, plus retained configurations and corrections so downstream outputs can be identified for revision.ReviewPilot records counts and inclusion-exclusion criteria for a PRISMA-guided workflow.Source: §1 · p. 1
  • ReviewPilot lets researchers define criteria and schemas through dialogue, then records screening reasons and links extracted values to source passages.Its contributions include a PRISMA-guided workflow, project memory for configurations and corrections, and evaluation against a single-prompt LLM and AgentSLR.Source: §1 · pp. 1–2

2 Related Work

Prior work includes literature-synthesis tools, stage-specific review tools, and integrated workflows. ReviewPilot is presented as a PRISMA-guided system that coordinates review stages and supports review-specific configuration and revision.

2 Related Work

  • Deep research systems iteratively search and analyze sources to produce reports, but do not document paper filtering based on PRISMA 2020.The passage names ChatGPT (OpenAI, 2025), Gemini (Citron, 2024), Perplexity (Perplexity, 2025), OpenScholar, and AutoSurvey (Wang et al., 2024b).Source: §2 · p. 2
  • Stage-specific tools include ASReview and LGAR for screening and RobotReviewer for extraction and risk-of-bias assessment.Their described functions include active-learning prioritization, graded criterion-based ranking, and source-linked sentences.Source: §2 · p. 2
  • Integrated workflows include AgentSLR, which uses predefined eligibility criteria and domain-specific schemas, and systems describing cross-review, correction, or editable-field workflows.The passage also describes Elicit's editable criteria, evidence quotes, exclusion reasons, and PRISMA flow diagrams.Source: §2 · pp. 2–3
  • ReviewPilot coordinates literature search, screening, full-text retrieval, and extraction, with dialogue-based criteria and schemas and project memory for configurations and corrections.The paper evaluates the complete system on PERG and examines workflows and screening agreement with human annotators in two case studies.Source: §2 · p. 3

3 ReviewPilot Development

ReviewPilot coordinates five systematic-review stages while allowing researchers to configure, inspect, and correct work. Its agents generate review-specific guidance, search and assess records, retrieve papers, extract values, and retain project state.

3 ReviewPilot Development

  • ReviewPilot covers search, screening, full-text retrieval, information extraction, and semantic categorization, with researcher configuration, inspection, and overrides.Researchers can configure through dialogue or direct editing and correct decisions or values.Source: §3 · p. 3

A Lead Agent

  • The Lead Agent coordinates specialized agents and tools, starts stages only when prerequisites and dependent outputs are current, and checks required artifacts before reporting completion.It categorizes extracted values itself.Source: A Lead Agent · p. 3

3.2 Review-Specific Prompt Design

  • The Prompt Agent drafts screening guidance and extraction schemas and coding rules from review objectives, criteria, requested fields, and selected paper text.Researchers can edit and must confirm the guidance, schema, and rules; reviewer decisions and reference values are excluded from generation inputs.Source: §3.2 · p. 3

Search.

  • The Search Condition Agent builds Boolean queries by joining alternatives with OR and eligibility-required concept groups with AND, while omitting non-eligibility dimensions.Researchers review sources, date range, and limits; the search tool queries titles and abstracts in PubMed, arXiv, and OpenAlex.Source: Search · p. 3
  • Screening assesses each dated, deduplicated record against confirmed criteria and guidance, returning a rationale and one of five likelihood labels.Only probably-exclude and definitely-exclude labels exclude a record; failed model calls and other labels are retained for full-text review.Source: Search · pp. 3–4

3.4 Paper Retrieval & Information Extraction

  • The retrieval tool tries publisher, open-access, and preprint sources; extraction applies a confirmed schema and coding rules to retrieved PDFs.Returned excerpts are checked for textual location, but location does not establish semantic support for a value or coverage of every value in a multi-value field.Source: §3.4 · pp. 4–5
  • When PDFs are unavailable, researchers can enable abstract assessment and supported web evidence, with source labels and located excerpts required for accepted values.Fields without evidence remain unavailable, and papers with unavailable eligibility are not categorized.Source: §3.4 · p. 5

3.5 Semantic Categorization

  • For retained papers, researchers review and edit proposed field categories before the Lead Agent assigns single- or multiple-label values; missing values are marked unavailable.The plan includes an Other category for values that do not fit.Source: §3.5 · p. 5

4 Experiments & Results

ReviewPilot is evaluated against expert annotations on PERG screening and extraction, then its workflows and screening agreement are examined in two case studies. Across the PERG models, it leads both screening and extraction comparisons.

4.1 Experiment Setup

  • The PERG evaluation pools 5,614 screening records with 854 expert inclusions and 178 extraction papers scored across five fields.Extraction is evaluated on the same expert-included papers for all methods, independently of screening predictions.Source: §4.1 · p. 5, §4 · p. 5
  • The PERG pools also informed ReviewPilot’s screening procedure and prompt-design skill, so they were not held out from method development.Expert decisions and reference extraction values were excluded from the Prompt Agent’s inputs.Source: §4.1 · p. 6

4.2 Screening Results

  • 5.3–10.3 points over AgentSLR and 7.0–19.0 points over the single-prompt LLM: ReviewPilot has the highest screening macro F1 on each model.It also has the highest macro precision and recall on each model.Source: §4.2 · p. 6, Table 1 · p. 6
  • ReviewPilot retains 26–29% of records, compared with 33–42% for AgentSLR and 35–54% for the single-prompt LLM; its screening API cost is below AgentSLR’s but above the single-prompt LLM’s.The lower review share does not establish fewer missed inclusions or reduced human review time.Source: §4.2 · pp. 6–7

4.3 Extraction Results

  • 3.7–4.7 points over AgentSLR and 1.5–1.7 points over the single-prompt LLM: ReviewPilot has the highest pooled extraction F1 on each model.Its pooled precision and recall are also highest on each model.Source: §4.3 · p. 7, Table 2 · p. 7
  • 0.703–0.714 verified-correct rate for ReviewPilot, versus 0.405–0.449 for AgentSLR and 0.528–0.561 for the single-prompt LLM.This rate requires a complete normalized match and a located excerpt; semantic support still needs verification.Source: §4.3 · p. 7, Table 2 · p. 7

5 Case Studies

ReviewPilot was applied to biomedical and HCI reviews, with complete workflows and separate screening audits. Human annotators’ agreement and source inspections provide evidence about its use in these two case studies.

5.1 Complete Review Workflows

  • The HCI review yielded 752 categorized studies from 2,241 records; the researcher reinstated three exclusions and regrouped the proposed application categories.Full-text assessment excluded 75 of 827 reports, and 153 reports were not retrieved.Source: §5.1 · p. 7
  • The biomedical review yielded 1,341 categorized studies from 3,000 records; the researcher reinstated four exam evaluations and kept the five proposed task categories.Full-text assessment excluded 21 of 1,362 reports; 336 reports were not retrieved and seven were not extracted.Source: §5.1 · p. 7

5.2 Human Agreement and Inspection

  • In separate audit pools of 87 biomedical and 90 HCI records, ReviewPilot’s screening κ exceeded annotator agreement in biomedicine and fell on either side of it in HCI.Two expert annotators independently labeled titles and abstracts without seeing ReviewPilot’s output.Source: §5.2 · p. 7
  • Source checks corrected three field values per topic, including restored participant counts in HCI and removal of unsupported biomedical labels.The inspections also added missing excerpts and distinguished human experiments from agent assessments in HCI.Source: §5.2 · pp. 7–8

6 Discussion and Conclusion

ReviewPilot combines review-specific configuration and project memory with screening and extraction whose decisions and values can be inspected. On PERG, it outperforms two baselines across three reported outcome measures, while the evaluation and case studies leave important questions for future work.

6 Discussion and Conclusion

  • ReviewPilot combines review-specific guidance and project memory to support systematic reviews from search through semantic categorization.Researchers confirm criteria and schemas; the Prompt Agent writes guidance and coding rules, and project memory retains configurations and corrections.Source: §6 · p. 8
  • ReviewPilot achieves higher screening macro F1, pooled extraction F1, and verified-correct rates on PERG than a single-prompt LLM and AgentSLR.The comparison is reported across three LLMs.Source: §6 · p. 8
  • PERG pools informed ReviewPilot’s development, so the reported gains do not establish performance on unseen reviews.Whether a located excerpt semantically supports a decision or value still requires human judgment.Source: §6 · p. 8
  • Case-study checks of flagged and sampled outputs led researchers to reverse some screening exclusions and correct some field values, with each change retained with its reason.ReviewPilot pairs automation with inspectable evidence and revisable decisions.Source: §6 · p. 8
  • Future work should measure effects on review effort and decision quality and evaluate reviews with other evidence types.Source: §6 · p. 9

Limitations

The evaluation leaves broad generalizability and several aspects of measurement unresolved. The authors identify limits in review coverage, retrieval, auditing, and model stability, and call for more comprehensive evaluation.

Limitations

  • PERG pools informed method development, and the evaluation also lacks repeated runs, confidence intervals, and independent audits of full-text eligibility.Macro-averaged screening metrics do not isolate include-class sensitivity; extraction covers five fields from epidemiological transmission models, while case-study audits use small, non-random pools.Source: Limitations · p. 9
  • Search is limited to PubMed, arXiv, and OpenAlex and depends on queries, per-source result limits, and successful PDF retrieval.The reported results do not quantify retrieval outcomes by access status; subscription barriers’ contribution to unretrieved reports remains unquantified.Source: Limitations · p. 9
  • Future evaluations should use held-out reviews, additional evidence types and schemas, and independent audits of eligibility and semantic support.They should also report repeated-run uncertainty and include-class recall.Source: Limitations · p. 9
  • Prospective studies with review teams are needed to assess reviewer effort and decision quality, while retrieval evaluations should report outcomes by access status.The authors also call for assessing how manual PDF submission and fallback assessment affect evidence coverage.Source: Limitations · p. 9
  • Hosted-model behavior may change across versions, and ReviewPilot’s outputs require human verification.The system is intended to support independent review procedures.Source: Limitations · p. 9

A ReviewPilot Implementation Details

The appendix presents implementation details in the order of the main text.

A ReviewPilot Implementation Details

A.1 System and Orchestration

The system’s orchestration uses task-specific action contracts and checks for required stage artifacts before marking a stage complete.

A.1 System and Orchestration

  • A Lead Agent checks for required artifacts before reporting a stage complete, while fixed skill mappings append task-specific skills to system prompts.The implementation uses a Starlette backend and JavaScript interface, with a project-level model setting for search setup, conversation, screening, and extraction.Source: §A.1 · p. 12

A.2 Search

The Search Condition Agent forms a Boolean query from required eligibility groups and their alternative terms. The search tool applies source-specific syntax while preserving Boolean grouping and records each source’s query settings, counts, and errors.

  • The Search Condition Agent forms a Boolean query for a research description.Source: §A.2 · p. 12
  • Required eligibility groups define the groups of alternatives used in the query.Source: §A.2 · p. 12
  • The search tool adapts the query syntax for PubMed, arXiv, and OpenAlex while preserving Boolean grouping and recording per-source queries, settings, counts, and errors.Source: §A.2 · p. 12

A.3 Screening Preprocessing

Screening preprocessing retains uncertain-date records, removes exact and near duplicates, and records a reason for each removal.

  • Records with missing dates or date intervals overlapping the search range are retained.Source: §A.3 · p. 12
  • Exact duplicates are detected after normalizing title case, punctuation, and whitespace.Source: §A.3 · p. 12
  • Near duplicates are detected by title-and-author TF-IDF cosine similarity at a default threshold of 0.70, while distinct nonempty DOIs or PubMed identifiers keep records separate.Source: §A.3 · p. 12
  • Each record removal is saved with its reason.Source: §A.3 · p. 12

A.4 Full-Text Retrieval

ReviewPilot retrieves full texts, validates and records extracted values, supports field categorization, and lets researchers inspect and revise project configurations.

A.4 Full-Text Retrieval

  • The PDF Download Tool tries sources based on record identifiers and metadata, and checks the opening-page text against the expected title.Retrieval routes include direct links, publisher pages, DOI resolution, repositories, and preprint servers; requests use throttling and retries, with an optional browser fallback.Source: §A.4 · pp. 12–13

A.5 Extraction

  • Extraction checks field presence, declared types, and value constraints, then matches proposed excerpts to normalized PDF text and records their page.The PDF text is read with page markers; schema fields specify a name, JSON-compatible type, and description.Source: §A.5 · p. 13

A.6 Categorization

  • For each selected field, the Lead Agent proposes up to five categories from as many as 150 distinct extracted values, with outputs saved alongside assignments.The plan supports single- or multiple-label assignment and an Other category when needed; papers without a value are marked unavailable.Source: §A.6 · p. 13

A.7 Interface, Memory, and Revision

  • The interface separates project navigation, conversation, and stage-specific editing; sample runs of one to five papers remain separate from full-run outputs.Project memory supports importing configurations and rejects edits submitted against an older revision.Source: §A.7 · p. 13

B PERG Data, Criteria, and Extraction Fields

The PERG data and screening rules define the material and criteria used for evaluation; the appendix also identifies five evaluated extraction fields.

C Baselines

The section specifies the single-prompt and AgentSLR comparison setups, including their prompts, and points to documentation of models, prices, and request settings.

C Baselines

  • The single-prompt baseline received the same records, papers, eligibility rules, and extraction schema as ReviewPilot, but no review-specific guidance, coding rules, skills, or schema validation.Source: C Baselines · p. 13
  • AgentSLR uses its published screening template unchanged; implementation changes are listed in Table D.2.Source: C Baselines · p. 14
  • The screening prompt applies all inclusion criteria and no exclusion criteria, includes uncertain records, and returns an include decision with a reason.Source: C Baselines · p. 14
  • The extraction prompt requests five PERG fields across a paper’s dynamic transmission models and field-level evidence quotes, page numbers, and support statuses.The fields are evaluated in Table B.2.Source: C Baselines · p. 14, Table B.2 · p. 14

D Run Configuration

  • Table D.1 lists models and prices for cost estimates, while Table D.3 lists each method’s request settings.Source: D Run Configuration · p. 14

E Additional Results

Tables E.1 and E.2 provide additional breakdowns of the results in Tables 1 and 2 by pathogen and field, using the same metrics and existing runs.

E Additional Results

F Case-Study Details

This section documents biomedical and HCI case-study workflows and the prompts, eligibility guidance, and domain-specific coding rules used by ReviewPilot’s agents. It also describes evidence-provenance-preserving extraction and illustrates the system’s staged interface for screening, extraction, and analysis.

F Case-Study Details

  • 1,341 of 3,000 biomedical records were included after screening and full-text assessment; the workflow also shows unretrieved and retrieved-but-unextracted reports.The figure gives examples of a survey exclusion, a reversed exclusion of a medical-exam evaluation, and extraction of model names and expert evaluation with source excerpts.Source: Figure F.1 · p. 17
  • Researcher inspections checked flagged records and stratified samples, while extraction checks compared sampled values with full-text evidence and added missing excerpts.HCI sampling included participant counts, technical evaluations without users, other application areas, fine-tuning or RAG, and papers with user-facing methods but missing participant counts. Biomedical checks covered expert evaluation, clinical notes, and other tasks; these inspection counts do not estimate corpus-level error rates.Source: F Case-Study Details · pp. 14, 19, Table F.2 · p. 18

G ReviewPilot Prompts

  • The appendix reproduces prompts for the Lead, Search Condition, Prompt, Screening, and Extraction Agents and the Lead Agent’s categorization prompts; search and PDF-download tools make no LLM calls.Agent skills are appended to system prompts in a designated block, and shaded text marks slots filled for each review or record.Source: G ReviewPilot Prompts · p. 19

G.1 Lead Agent

  • The Lead Agent answers project questions and turns chat requests into draft edits to screening criteria or extraction schemas for researcher confirmation.Project answers use saved configurations and confirmed decisions as authority; chat edits preserve unaffected rules, and schema edits are made through supported commands.Source: §G.1 · pp. 19–20

G.2 Search Condition Agent

  • The Search Condition Agent creates and revises concept strategies, whose Boolean queries OR terms within eligibility groups and AND required groups.Invalid responses are retried once with the validation reason appended; the Prompt Agent uses the same retry.Source: §G.2 · p. 20

G.3 Prompt Agent

  • The Prompt Agent drafts and reviews screening guidance and extraction coding rules against a skill checklist, while generating the extraction schema.Review-specific guidance interprets confirmed criteria or schema for the records the review will meet; it cannot add requirements or fields or drop stated ones.Source: §G.3 · p. 23

G.4 Screening Agent

  • The Screening Agent evaluates what a study itself did against the review focus and criteria, uses uncertainty when title and abstract evidence leaves eligibility open, and gives a one-sentence reason.It excludes only when the study’s own work clearly falls outside the criteria or could not supply the information sought; background mentions alone do not count.Source: §G.4 · p. 29

G.4.1 PERG Review Focus and Domain Rules

  • For the Lassa review, records qualify through specified quantitative human-outbreak evidence, transmission or epidemic models, human seroprevalence or measured risk factors, or specified animal transmission or vector parameters.A background mention alone is insufficient; risk factors require statistical association measures, and animal-only qualifying parameters include R0, Rt, Re, r, growth rate, mutation rate, and specified vector measures.Source: §G.4.1 · pp. 29–30
  • Lassa guidance retains unclear abstracts when the work type normally produces eligible results, and retains title-only records unless the title shows an excluded type or topic.Work types named for uncertain cases include outbreak analyses, serosurveys, parameter-estimation studies, and transmission models; systematic reviews or meta-analyses receive special consideration when they quantitatively synthesize eligible results.Source: §G.4.1 · p. 30
  • Henipa guidance covers quantitative Nipah/Hendra human outbreak, transmission, risk-factor, seroprevalence, and specified animal/vector evidence; animal infection or serology alone does not qualify.When abstracts are unclear, retain work types likely to yield eligible evidence; consider reviews for full-text assessment when they may report pooled eligible estimates.Source: §G.4.1 · pp. 30–31
  • SARS-CoV rules require eligible natural-transmission evidence; SARS-CoV-2, MERS-CoV, and accidental laboratory outbreaks do not qualify on their own.Eligible evidence includes quantitative outbreak details, transmission models or estimates, risk-factor analyses, seroprevalence, and specified genetic or animal/vector parameters.Source: §G.4.1 · pp. 31–32
  • Zika eligibility requires the study’s own quantitative Zika evidence, including human outbreak, transmission, timing, severity, risk-factor, or seroprevalence results, or specified animal/vector parameters.A background mention or unmeasured proposed risk factor does not qualify; eligible animal and vector parameters are specified in the guidance.Source: §G.4.1 · pp. 32–33
  • The Zika review excludes studies lacking specified transmission or vector parameters, including narrowly ecological, clinical, diagnostic, laboratory, review, and non-original-report records, subject to stated exceptions.The rules also distinguish uncertain abstracts by the kind of work and retain records without abstracts unless the title establishes exclusion.Source: §G.4.1 · p. 33
  • When an abstract leaves eligibility unclear, reviewers retain work likely to produce a qualifying result, probably exclude work unlikely to do so, and use the title when no abstract is available.Systematic reviews and meta-analyses receive the criteria’s special consideration rather than being treated as original research by default.Source: §G.4.1 · p. 33

G.5 Extraction Agent

  • The extraction agent receives the scope, finalized schema, task and coding rules with the paper text, and records field-level evidence as exact quotes, page numbers, and support statuses.For papers without full text, it uses the title and abstract, then asks the web step to fill fields that remain empty.Source: §G.5 · pp. 33–34

G.5.1 PERG Coding Rules

  • The Prompt Agent creates pathogen-specific coding rules that classify each paper’s modeled work rather than background mentions, covering model type, stochasticity, code availability, transmission routes, and interventions.Rules are written for Ebola, SARS, Zika, and Lassa; for example, probabilistic fitting alone does not make transmission stochastic.Source: §G.5.1 · pp. 35–38
  • PERG coding distinguishes modeled transmission routes, including vector or animal-to-human, direct human-to-human, sexual, and airborne or close contact, and codes each explicitly modeled route.When a more direct route is specified, it takes precedence over the general airborne or close-contact code.Source: §G.5.1 · p. 38
  • Interventions are coded only when represented as interventions in the model; listed categories cover quarantine, treatment, animal control, vaccination, and other specified measures.An unlisted modeled intervention is coded Other; a contact rate, seasonal reservoir dynamic, or saturation term is not an intervention unless tied to an implemented intervention.Source: §G.5.1 · p. 38
  • Animal-reservoir control measures such as rodent trapping, food protection, and sanitation are coded as Vector/Animal control, while vaccine campaigns and baiting are Vaccination.Modeled interventions are distinguished from background discussion and model dynamics that are not tied to an implemented intervention.Source: §G.5.1 · p. 38

G.6 Categorization

  • ReviewPilot plans broad semantic categories from all papers’ values for a selected field, then assigns each paper to one or several supported categories.The Lead Agent performs the categorization calls, with a Categorization Skill appended to each system prompt.Source: §G.6 · p. 38

H System Interface

  • The interface records screening decisions and extraction evidence, while marking full-text retrieval and extraction as partial where papers lack PDFs or remain unextracted.The activity card reflects 1,701 included and 1,014 excluded before four records were reinstated; the final canvas and PRISMA flow show updated counts. The extraction view displays a verbatim PDF excerpt supporting GPT-4o.Source: H System Interface · p. 39, Figure H.3 · p. 41, Figure H.5 · p. 42
  • ReviewPilot’s figures show search setup with query, sources, limits and date range, finalized screening criteria, full-text retrieval, and a generated PRISMA 2020 flow reporting 1,341 included and categorized studies.The PRISMA flow starts from 3,000 identified records. PDFs were retrieved for 1,369 of 1,705 records; 336 failures are listed with reasons for retry or manual supply.Source: Figure H.1 · p. 40, Figure H.2 · p. 40, Figure H.4 · p. 42

Related papers

Loading 2610.21042v1…