PaperlayerView exact revision on arXiv

Source-linked AI summary

Large language models for automated scholarly paper review: A survey

Zhenzhen Zhuang, Jiandong Chen, Hongfeng Xu, Yuwen Jiang, Jialiang Lin

arXiv:2501.10326v2cs.AIcs.CLcs.DL

TL;DR

This survey asks how LLMs advance automated scholarly paper review while addressing related challenges. It comprehensively reviews ASPR in the LLM era, including model use, technological progress, performance, and limitations.

  • Problem

    The survey examines how LLMs promote the advancement of automated scholarly paper review and related developments.

  • Method

    The paper undertakes a comprehensive survey of ASPR in the LLM era and systematically reviews associated developments.

  • Results

    LLMs enhance ASPR efficiency, while proprietary models generally outperform open-source counterparts in identifying defective segments.

  • Takeaways & Limitations

    The survey provides a comprehensive reference for understanding LLM-driven ASPR and its progress toward implementation.

  • Takeaways & Limitations

    Existing long-text methods remain insufficient because they cannot consider full context across the entire text.

Abstract

from arXiv · show

Large language models (LLMs) have significantly impacted human society, influencing various domains. Among them, academia is not simply a domain affected by LLMs, but it is also the pivotal force in the development of LLMs. In academic publication, this phenomenon is represented during the incorporation of LLMs into the peer review mechanism for reviewing manuscripts. LLMs hold transformative potential for the full-scale implementation of automated scholarly paper review (ASPR), but they also pose new issues and challenges that need to be addressed. In this survey paper, we aim to provide a holistic view of ASPR in the era of LLMs. We begin with a survey to find out which LLMs are used to conduct ASPR. Then, we review what ASPR-related technological bottlenecks have been solved with the incorporation of LLM technology. After that, we move on to explore new methods, new datasets, new source code, and new online systems that come with LLMs for ASPR. Furthermore, we summarize the performance and issues of LLMs in ASPR, and investigate the attitudes and reactions of publishers and academia to ASPR. Lastly, we discuss the challenges and future directions associated with the development of LLMs for ASPR. This survey serves as an inspirational reference for the researchers and can promote the progress of ASPR for its actual implementation.

1. Introduction

LLMs are increasingly being incorporated into automated scholarly paper review, motivating a systematic survey of their role, resources, performance, reception, and future challenges. This survey uses snowballing to examine how LLMs advance ASPR in the era following ChatGPT.

  • LLMs have been employed to generate scholarly peer-review report content, expanding their influence within academic evaluation.
  • ASPR currently coexists with traditional peer review, serving as a human assistant intended to improve review quality and efficiency.
  • The survey asks how LLMs promote ASPR advancement and systematically reviews the emerging research body.
  • Using snowballing, the authors select seed papers, expand them through forward and backward citation searches, and include publicly available 2023–2024 studies using LLMs for ASPR.
  • Bias, fairness, ethics, accountability, and misuse remain relevant concerns, while their detailed discussion is deferred to earlier work to avoid redundancy.
  • The survey covers LLM usage, technological bottlenecks, methods, resources, performance, issues, publisher and academic reactions, challenges, and future directions.

2. Background of LLMs and their applications in ASPR

LLMs evolved from statistical and neural language models to Transformer-based systems with scaling-driven capabilities, while their ASPR applications reveal both comparative strengths and persistent limitations. The background reviews model development, efficiency techniques, architectures, and observed ASPR performance.

  • Development process of large language models: LLMs progressed from statistical and recurrent neural models to Transformer-based pre-trained systems such as BERT and GPT.
  • Development process of large language models: Scaling model parameters and training data enabled emergent capabilities including in-context learning, instruction following, and chain-of-thought reasoning.
  • Development process of large language models: Llama 3 improved zero-shot logical reasoning by 23.6% over its predecessor after incorporating 17% structured code data into pre-training.
  • Efficiency and deployment: Compression methods such as pruning, quantization, distillation, and dynamic inference reduce computational demands for resource-constrained deployment.
  • Efficiency and deployment: DistilBERT retains 97.2% of BERT-base’s GLUE performance while reducing parameter count by 40% through hierarchical distillation.
  • LLM components and architectures: Attention mechanisms, positional encoding, activation functions, and normalization strategies address architectural and long-sequence processing challenges in Transformer models.
  • Application and comparative analysis of LLMs in ASPR: In ASPR meta-reviewing, closed-source LLMs generally outperform open-source counterparts in identifying defective segments, but all models have relatively low F1 scores and recall exceeds precision.
  • Application and comparative analysis of LLMs in ASPR: GPT-4 is the most frequently referenced LLM in ASPR applications, while the “Others” category exceeds one-third of the total, indicating application diversity.

3. New technologies in the era of LLM-driven ASPR

LLMs address earlier ASPR bottlenecks by expanding context windows, incorporating multimodal inputs, supporting multi-turn review interactions, and acquiring current information through search.

  • 3.2. Multimodal modeling: Multimodal LLMs combine textual and visual data, enabling ASPR to consider figures, tables, and other non-textual paper content.This addresses the earlier focus on textual data that overlooked information relevant to scholarly evaluation.
  • 3.3. Multi-turn interaction: LLMs support multi-turn review workflows involving author responses and interactions among reviewers, extending ASPR beyond its earlier one-turn definition.Long-text conversation capabilities allow review reports to be revised across simulated peer-review phases.
  • 3.4. Instant knowledge acquisition: Search-enabled LLMs can retrieve up-to-date information when model knowledge is insufficient or outside reviewers’ expertise.The feature supports evaluating papers against recent research and assessing ideas and originality more comprehensively.

4. New methods for generating review reports by LLMs

LLM-based review generation uses prompt engineering, supervised fine-tuning, and multi-agent or staged architectures to produce more structured, specific, and actionable reports.

  • 4.1. Prompt engineering: Prompt engineering encodes review criteria, examples, checklists, and output formats to guide LLMs toward accurate, comprehensive, and compliant reports.Applications include scoring papers, identifying errors, answering checklist questions, and generating paired comparisons.
  • 4.2. Supervised fine-tuning: Supervised fine-tuning adapts pretrained models using labeled review data, including LimGen’s 4,068 papers and corresponding limitations.LimGen fine-tuning with dense passage retrieval and chain modeling generates limitations using the paper’s entire context.
  • 4.3. Multi-agent and process-based methods: Multi-turn and role-based systems model authors, reviewers, chairs, and decision-makers to simulate iterative conference review processes.AgentReview spans reviewer assessment, author discussion, reviewer-chair discussion, meta-review compilation, and chair decision-making.
  • 4.3. Multi-agent and process-based methods: Multi-agent architectures divide review generation among specialized roles and stages, producing more specific, useful, coherent, or accurate feedback.SWIF2T uses planner, investigator, reviewer, and controller components, while MARG combines leader, worker, and expert agents with refinement stages.
  • 4.3. Multi-agent and process-based methods: SEA standardizes and evaluates review comments before analysis, introducing mismatch score to assess consistency between papers and reviews.Its modules integrate GPT-4 consolidation, fine-tuning, and a consistency metric for practical and specific comments.

5. New datasets

New ASPR datasets support screening, reviewing, comment generation, quality assessment, and author response, but remain largely text-centered and weakly multimodal.

  • 5.1. Dataset categories: Recent datasets are organized by review phases including screening, main review, comment generation, quality assessment, and author response.Table 2 provides an updated review of datasets developed since earlier ASPR dataset work.
  • 5.2. Dataset contents: Most datasets contain manuscript text, metadata, decisions, comments, and reports, while some represent multi-turn review conversations.Conversation data enables models to handle iterative discussions, and expert annotations support multidimensional quality assessment.
  • 5.3. Dataset limitations: Existing ASPR datasets remain predominantly textual and often omit figures, images, and tables from manuscripts.This lack of multimodal integration restricts development and practical application of multimodal LLM-driven ASPR.

6. Source code and online systems for LLM-driven ASPR

Open source code, online systems, and increasingly capable LLM-driven approaches support ASPR reproducibility and broaden coverage across review subtasks.

  • 6.2. Reproducibility and online systems: Without source code, reproducing empirical ASPR experiments becomes nearly impossible, whereas online systems let users experience emerging review technologies.Tyser et al. provide systems for paper feedback, trend analysis, and review-quality improvement.
  • 6.1. Source code and approaches: LLM-driven ASPR approaches address screening, main review, comment generation, and author response, outperforming traditional NLP baselines in most cases.These approaches exploit LLMs’ long-text and multi-turn conversation capabilities.
  • 6.1. Source code and approaches: MAMORX was introduced as the first open-source multi-module LLM-driven ASPR system, integrating textual, graphical, citation, and external-knowledge inputs.The survey reports that this comprehensive approach has advantages over human reviewers and baseline models.
  • 6.2. Reproducibility and online systems: SEA exemplifies reproducibility by releasing its architecture, weights, source code, requirements, installation instructions, tutorials, demos, and hardware support.These materials are presented as a strong commitment to replicable ASPR research.

7. ASPR enhancement with LLMs

LLMs have expanded ASPR beyond early format validation, proofreading, plagiarism detection, and templated reports by improving review efficiency and supporting screening, summarization, verification, error detection, and feedback generation.

  • Overall enhancement: LLM-driven ASPR significantly improves review performance and functionality compared with earlier automation based on validation, proofreading, plagiarism detection, and templates.The surveyed advances include new generation methods, datasets, and open-source code.
  • Efficiency advancement: LLMs accelerate peer review by rapidly assessing manuscript suitability, identifying potential issues and ethical concerns, and streamlining publication workflows.They also substantially alleviate human reviewers’ workload.
  • Efficiency advancement: Authors rated LLM-generated review comments as equally helpful as human comments, with an average score of 3 out of 5.This result indicates perceived value of LLM assistance during the coexistence of ASPR and human review.
  • Summarization: LLMs generate abstractive manuscript summaries and quality-evaluation rationales, with reported strengths in fluency, factuality, flexibility, and REF-criteria alignment.Their summarization captures essential points without directly copying the reviewed manuscript.
  • Screening empowerment: LLMs support screening by evaluating topical relevance and manuscript quality, achieving up to 90% overall accuracy in abstract screening tasks.They can serve as first-pass filters for journal fit, quality, plagiarism, and preprint screening.

8. Main existing issues

Despite advances in ASPR, LLMs remain limited in deep, domain-specific manuscript evaluation and can produce generic, biased, unstable, or factually incorrect reviews. Confidentiality risks also arise when unpublished manuscripts are processed by external models.

  • Insufficient knowledge and comprehension: LLMs still cannot fully replace human reviewers because ASPR requires deep knowledge, experience, critical evaluation, and contextual judgment.The survey identifies limitations in assessing originality, soundness, scientific contribution, and acceptance-worthiness.
  • Insufficient knowledge and comprehension: LLMs struggle with specialized terminology, recent advances, domain-specific expertise, and the broader field context needed to assess complex research contributions.Their comments are often brief, generic, or out of scope rather than tailored to manuscript content.
  • Biased and unstable review: LLM evaluations can be inconsistent, overly favorable, and biased: one study found no reviewed paper received the lowest score, while repeated evaluations fluctuate.Training-data biases may affect responses to demographic groups, institutions, countries, minority perspectives, and unfamiliar content.
  • Reliability concerns: LLMs may hallucinate fabricated information and references, misjudge weak manuscripts, and accept authors’ claims without thorough verification.These outputs can appear convincing while remaining factually incorrect.
  • Critical-evaluation limitations: LLMs often miss deeper theoretical flaws, missing metrics, overstated conclusions, writing-quality problems, and overall correctness, while producing unstable evaluations across repeated reviews.Their assessments may reduce manuscript evaluation to superficial summarization rather than critical review.
  • Confidentiality: Processing unpublished manuscripts with LLMs can create confidentiality risks because manuscript content may be stored, processed, or incorporated into future training data.The survey identifies these risks as a primary concern in LLM-enabled ASPR.

9. Publisher policies on AI-generated content tools in peer review

Publisher policies generally restrict reviewers’ use of external AIGC tools because of confidentiality, accountability, bias, and review-quality concerns, while some permit limited assistance under safeguards. Several publishers instead provide in-house systems for editorial screening and integrity checks.

  • General policies: Most publishers prohibit reviewers from using AIGC tools to generate or assist in writing review reports.The stated reasons include risks of unauthorized manuscript exposure, compromised confidentiality, reduced reviewer accountability, bias, and harms to review quality and fairness.
  • Permitted uses: Some publishers allow AIGC tools for improving review quality, readability, concept understanding, or review comments, with confidentiality requirements.ACM and ACL permit specified uses without requiring disclosure, whereas JAMA Network requires reviewers to identify the tool and its application.
  • Confidentiality: Editors and reviewers must not upload manuscripts to platforms that retain data or fail to comply with confidentiality requirements.This restriction is shared across publishers despite differences in whether AIGC use is otherwise permitted.
  • In-house systems: Frontiers’ AIRA can generate up to 20 recommendations in seconds, including language, figure-integrity, plagiarism, and conflict-of-interest checks.AIRA is presented as an in-house AI tool for enhancing editorial workflow efficiency.
  • In-house systems: Wiley’s Research Exchange supports reviewer matching and research-integrity screening, including AI-content detection, identity verification, reference checks, and suspicious-language detection.The platform integrates these capabilities into the publishing workflow, while prohibiting external AIGC tools for evaluations or editorial decisions.
  • Policy gaps: For publishers without publicly stated reviewer-AIGC policies, COPE guidance addresses authors’ AI use but not reviewers’ use, leaving permission unclear.These publishers are classified as “Unknown” in Table 4 based on the information available when the survey was written.

10. Suggestions from academia

Academia is more receptive than publishers to LLM use in review, provided reviewers protect data, disclose involvement, retain responsibility, receive training, and avoid exclusive reliance on models. Suggested practices also include private hosting, model selection by task, repeated prompting, and alignment with human values.

  • Academic stance: Academia generally takes an open-minded approach to LLM-assisted review while maintaining strong concern about data security.Academic recommendations emphasize responsible use and further development rather than broad prohibition.
  • Data security: Sensitive or protected materials should not be entered into LLMs unless appropriate data-security measures are in place.Privately hosted LLMs can keep confidential data on the user’s own server instead of sending it to external providers.
  • Data security: Private hosting improves data security but requires technical expertise, substantial hardware, and may involve lower performance from open-source models.These constraints can make private deployment difficult despite its privacy benefits.
  • Training: Training should cover LLM ethics, benefits, limitations, and biases, with regular updates as the technology evolves.The proposed training is intended to foster responsible use and awareness of model weaknesses.
  • Transparency and responsibility: Reviewers and editors are encouraged to disclose LLM involvement, retain responsibility for reports, and share experiences to improve transparency and collective learning.Suggestions include self-declaration or machine-generated watermarks and clear reporting of tools and functions used.
  • Operational practice: Researchers should not rely exclusively on LLMs because technical breakdowns remain possible.Additional recommendations include matching model complexity to task demands, dynamically switching models, using current versions, averaging multiple evaluation rounds, and splitting manuscripts into smaller sections.
  • Alignment and oversight: LLM value judgments should align with human researchers’ perspectives and values rather than reflecting model-inherent viewpoints or biases.Models should also be monitored for compliance with journal guidelines, review forms, and editorial or professional rules.

11. Challenges and future directions

The survey synthesizes LLM technologies, resources, policies, responses, and suggestions for ASPR, then identifies open challenges spanning reliability, multimodal data, reasoning, security, and deployment. Future directions emphasize more trustworthy, auditable, secure, and resource-efficient systems.

  • Correcting hallucinations: Hallucinations remain a persistent ASPR challenge because probabilistic generation can favor plausible language over factual correctness.Evaluation metrics may overemphasize fluency while neglecting factual consistency, especially in complex scenarios.
  • Correcting hallucinations: Future hallucination-mitigation directions include dynamic knowledge fusion, multimodal fact-checking, and interpretability enhancement.These strategies aim to align generated outputs with rigorous academic and other critical standards.
  • Generating with multimodal inputs: ASPR research using multimodal inputs remains limited because high-quality, richly annotated, balanced datasets are scarce.Acquisition, annotation, cleaning, and balancing difficulties constrain models’ ability to learn cross-modal associations and interactions.
  • Generating with multimodal inputs: Well-curated multimodal datasets are identified as a promising direction for advancing ASPR research.The survey stresses the need for further ASPR work on such datasets.
  • Applying reasoning models: Reasoning models could move ASPR beyond statistical pattern matching by deconstructing argumentative logic and supporting decision auditing through complete thought chains.Their ASPR application remains exploratory, with risks involving hallucinations from excessive reasoning and trade-offs between timeliness and computational resources.
  • Mitigating generative attacks: Generative attacks can fabricate references, endorse incorrect claims, reconstruct unpublished manuscript content, or infer author identities.The survey proposes layered defenses spanning model robustness, system access and encryption, and application-level audit trails and early warnings.
  • Mitigating generative attacks: Integrating defenses across model, system, and application layers can establish safeguards and trust boundaries for ASPR.The proposed framework includes atypical-input detection, abnormal-behavior suppression, access controls, encryption, audit trails, and early-warning systems.
  • Deploying low-resource privately hosted models: Privately hosted models protect ASPR confidentiality but face high GPU, energy, and implementation costs.Compression, hardware acceleration, caching, dynamic scheduling, mixed precision, monitoring, and elastic allocation are proposed to enable lower-resource deployment.

Declaration of competing interest

The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.

  • The authors declare no known competing financial interests or personal relationships influencing this work.
Loading 2501.10326v2…