PaperlayerView exact revision on arXiv

Source-linked AI summary

Paper Pilot: A Human-in-the-Loop Expert System for Evidence-Traceable Scientific Manuscript Generation in Applied Sciences

Nidhi Jha, Siddharth Chaudhary, Ajinkya Kulkarni

arXiv:2608.28596v1cs.AI

TL;DR

Paper Pilot addresses the governance gap in LLM-assisted scientific workflows, where ideas, methods, results, and claims can propagate without mandatory human approval or artifact-level traceability. It adapts CARE into an approval-gated, evidence-locked manuscript workflow with human authority over scientific decisions. In a first citation-grounding benchmark, the rules eliminated fabricated citations and surfaced evidence gaps, while broader result-grounding, revision, and robustness evaluations remain preliminary.

  • Problem

    Existing scientific-agent workflows lack a standard mandatory end-to-end set of human approval gates and artifact-level traceability for transitions from ideas to manuscript claims.

  • Method

    Paper Pilot adapts CARE into a human-in-the-loop workflow with eight manuscript-owner approval gates, no-pass criteria, claim classification, audit logs, advisory LLM review, and evidence-locked revision control.

  • Results

    The citation-grounding benchmark showed that Paper Pilot supported section-wise drafting, claim control, evidence-locked revision, and human approval while ungated drafting fabricated up to a quarter of citations and never flagged an evidence gap.

  • Takeaways & Limitations

    Paper Pilot positions LLM-assisted scientific writing as a controlled human-AI decision-support process rather than a fully autonomous authorship pipeline.

  • Takeaways & Limitations

    Empirical validation is limited: citation grounding used real arXiv papers, while result-grounding and revision benchmarks were preliminary synthetic evaluations without a full applied-science user study.

Abstract

from arXiv · show

Large language model (LLM) agents are increasingly embedded in scientific workflows for literature analysis, drafting, and review. Existing systems advance autonomous discovery and manuscript generation, but do not resolve the governance problem that arises when ideas, methods, results, and claims propagate through AI-assisted workflows without mandatory human approval or artifact-level traceability. This paper proposes Paper Pilot, a human-in-the-loop expert system for evidence-traceable scientific manuscript generation in applied sciences. It adapts the Collaborative Agent Reasoning Engineering (CARE) methodology to manuscript development through manuscript-owner approval gates, explicit no-pass criteria, claim classification, audit logging, advisory LLM review, and evidence-locked revision control. The framework defines eight approval gates across the idea-to-claim pipeline and distinguishes literature-grounded from artifact-grounded claims, requiring reported numbers and interpretations to remain traceable to approved evidence; its system prompt is openly released for deployment in ChatGPT, Gemini, Claude, or institutional LLM environments. As a first empirical validation, we evaluate the citation-grounding layer with a controlled, mechanically scored benchmark (two commercial LLMs, real arXiv papers, no LLM judge): under coverage pressure ungated drafters fabricated up to 25% of their citations and never flagged an evidence gap, whereas the same models under Paper Pilot's evidence-locked rules produced zero fabricated citations and surfaced the planted gaps as explicit placeholders. Preliminary results for result grounding, revision, and adversarial robustness point the same way; full evaluation is left to future work. Paper Pilot positions LLM-assisted writing as a controlled human-AI decision-support process rather than a fully autonomous authorship pipeline.

1 Introduction

Paper Pilot addresses governance and traceability gaps in increasingly integrated LLM-assisted scientific workflows by placing manuscript development under mandatory human approval and evidence controls.

  • Motivation: Existing scientific-agent systems increasingly integrate ideation, coding, experimentation, visualization, drafting, and review across closed-loop workflows.Related systems demonstrate growing automation from bounded assistance toward coordinated scientific workflows.
  • Motivation: Unsupported decisions can propagate from weak ideas or invalid methods into unreliable results and insufficiently supported manuscript claims.The paper identifies this propagation as a critical control problem in autonomous or semi-autonomous workflows.
  • Proposed framework: The framework extends citation-grounded writing to artifact-grounded claims, linking reported computational and analytical content to inspectable evidence.Relevant artifacts include code, data, seeds, environments, logs, configurations, outputs, and generated figures.
  • Proposed framework: Paper Pilot adapts CARE into a human-in-the-loop expert system with eight manuscript-owner approval gates across the idea-to-claim pipeline.The gates govern scope, literature, methods, evidence, claims, reviewer override, sections, and final release.
  • Authority boundary: LLM agents provide drafting and review support, while the manuscript owner retains authority over scientific validity, methodological adequacy, evidence sufficiency, and final claims.This authority boundary distinguishes Paper Pilot from fully autonomous paper-generation systems.

2 Related Work

Related systems improve autonomous scientific discovery, manuscript generation, scholarly synthesis, and automated review, but Paper Pilot focuses on governability, evidence traceability, and human authority.

  • Autonomous scientific workflows: Recent multi-agent systems connect hypothesis generation, experimentation, verification, manuscript preparation, and review across increasingly autonomous scientific workflows.Examples include InternAgent, OmniScientist, co-scientist models, and shared agent research infrastructure.
  • Writing and synthesis systems: PaperOrchestra, ARISE, and AutoSurvey2 demonstrate multi-agent manuscript writing and literature synthesis at scale.These systems target submission-ready writing, citation-guided survey generation, or higher-level literature reviews.
  • Evidence governance: Citation grounding and writing quality do not ensure that computational claims are reproducibly linked to code, data, seeds, environments, logs, and outputs.Citation-rich manuscripts can still contain metrics, plots, comparisons, ablations, or conclusions lacking artifact-level support.
  • Reviewer authority: Automated reviewers can provide scalable feedback but may exhibit drift, systematic bias, calibration failure, or overconfidence.Treating reviewer output as final authority could promote unsupported content because it appears persuasive to another model.
  • Paper Pilot positioning: Paper Pilot complements autonomous and writing-agent systems by adding mandatory approval gates, evidence-locked revision, and human authorization of scientific decisions.Its CARE-informed approach specifies reusable artifacts, verification procedures, and stage-gated controls rather than relying on ad hoc prompt refinement.

3 System Overview and Methodology

Paper Pilot converts structured user inputs into evidence-bounded manuscript sections through CARE-informed, section-wise drafting, claim classification, approval, and evidence-locked revision.

  • System overview: Paper Pilot adapts CARE to controlled, auditable, evidence-traceable manuscript production rather than autonomous paper generation.LLM agents assist with drafting, checking, revision, gap identification, and traceability management while the manuscript owner remains authoritative.
  • System purpose: The workflow accepts structured inputs and generates sections only after auditing evidence, identifying missing information, declaring assumptions, and blocking or marking unsupported claims.This process is implemented through general-purpose LLM environments and publicly released materials support inspection and reproducibility.
  • System purpose: Manuscript text must remain coupled to evidence, with unsupported data, methods, citations, and claims blocked, clarified, placeholder-marked, or escalated to the manuscript owner.Paper Pilot therefore functions as both a writing assistant and a governance layer.
  • System actors: The manuscript owner approves framing, literature interpretation, methods, results, revisions, unresolved placeholders, and submission readiness.The LLM agent layer organizes inputs, drafts prose, checks support, classifies claims, proposes revisions, and maintains change logs without final authority.
  • Workflow: The eight-stage workflow comprises manuscript setup, input audit, section planning, evidence-grounded drafting, claim classification, human approval, revision control, and release.Setup defines journal, manuscript type, scope, system name, and article structure; later stages preserve audit records and unresolved decisions.
  • Claim control and revision: Claim classification labels material statements as grounded_fact, inference, assumption, or tbd_placeholder, distinguishing approval-ready text from text requiring verification or revision.Evidence-locked revision extends citation-grounded revision to computational and user-provided evidence by keeping manuscript changes linked to their basis.

3.6 Approval-gate model

Paper Pilot operationalizes human control through mandatory approval gates and claim-level evidence records that prevent unsupported content from entering the final manuscript.

  • Approval-gate model: Mandatory manuscript-owner gates prevent advisory agent outputs from becoming final scientific claims without human authorization.The gates cover the manuscript pipeline and are paired with no-pass criteria, audit requirements, and revision or escalation conditions.
  • Evidence record: The minimal evidence record links each material claim to its evidence, approval state, metadata, and revision history.Supported evidence may include citations, user inputs, computational artifacts, figures, tables, logs, or repository files.
  • Scope: The evidence record is a lightweight traceability specification rather than a complete provenance infrastructure.It is designed for general-purpose LLM writing environments and can later support schema validation, approval checks, and claim-level orchestration.
  • Release conditions: Claims cannot pass claim approval without an evidence identifier, taxonomy assignment, and manuscript-owner approval status.Assumptions and tbd_placeholder claims may remain in internal drafts but require explicit acceptance as unresolved limitations or removal before release.

3.8 Claim extraction, classification, and validation pipeline

Paper Pilot converts drafted prose into reviewable claim records, matches claims to evidence, and requires manuscript-owner validation while keeping LLM review advisory.

  • Pipeline: The claim pipeline extracts candidate claims, assigns types and taxonomy labels, matches evidence, and validates the result with the manuscript owner.Its purpose is to support human review rather than replace it.
  • Claim extraction: Candidate claims include assertions about facts, procedures, results, interpretations, comparisons, limitations, or contributions.Transitional and rhetorical sentences are excluded unless they contain substantive assertions.
  • Claim taxonomy: A grounded_fact has approved support, an inference is bounded interpretation, an assumption is unverified, and a tbd_placeholder marks missing information.The taxonomy distinguishes evidentiary status independently from content category.
  • Evidence matching: Claims without linked evidence receive an unsupported risk flag and fail the evidence-matching stage.Literature, method, result, and interpretation claims are matched to their corresponding approved sources or artifacts.
  • Human validation: Manuscript owners may approve, revise, relabel, request evidence, reject, or accept a clearly labeled limitation, so LLM classifications remain advisory.This validation step prevents generated classifications from becoming self-authorizing.
  • Reviewer boundary: LLM reviewers may assess coherence, completeness, structure, citations, and journal alignment, but cannot approve validity, evidence sufficiency, or release.Their outputs are treated as decision support rather than decision authority.
  • Audit logging: Audit logs record inputs, missing information, assumptions, changes, classifications, citations, risks, and manuscript-owner decisions.The records support reconstruction of how claims were accepted, revised, or rejected.
  • Demonstrated resolution: All unresolved assumptions, inferences, and placeholders in the manuscript were reviewed and adjudicated before publication.The authors resolved material claims through evidence collection, literature validation, expert review, and manuscript-owner approval.

4 Workflow Demonstration: Applying Paper Pilot to Manuscript Development

The workflow demonstration shows Paper Pilot developing a manuscript section by section, with human inputs controlling framing, claims, evidence, revisions, and release boundaries.

  • Workflow demonstration: The demonstration illustrates section-wise drafting, evidence auditing, claim classification, approval gates, and revision logging rather than quantitative model evaluation.It uses the manuscript’s own development as a representative applied-science writing scenario.
  • Design principle: The workflow treats manuscript generation as controlled interaction among user evidence, LLM drafting, and manuscript-owner approval.This distinguishes the process from fully autonomous paper-generation systems.
  • Problem definition: The manuscript owner defined the target problem and research directions concerning approval gates, reviewer override, evidence traceability, and revision controls.These directions were later consolidated into the paper’s research questions.
  • Human direction: Specifying the target journal changed the framing from a general preprint to an applied expert-system article grounded in CARE.The manuscript owner also specified the system name, Paper Pilot.
  • Approval gates: Each major transition required manuscript-owner input or confirmation before topic, system name, or methodological foundation could be finalized.The demonstration makes the approval gates visible in the manuscript-development sequence.
  • No-pass behavior: Gate-based progression prevented premature claims about software architecture, benchmark results, operational deployment, or an unavailable repository URL.The workflow retained missing information as required input rather than inventing it.
  • Claim governance: Claim classification separated owner-provided and literature-supported grounded facts from derived inferences about Paper Pilot’s positioning and workflow.The demonstration shows how evidentiary status was assigned during manuscript development.
  • Revision control: Evidence-locked revision structurally reframed the manuscript after the journal change and integrated newly supplied system information as grounded content.Unsupported implications of full autonomy were removed rather than silently preserved.

5 Evaluation: A Mechanically Scored Benchmark Suite

Paper Pilot evaluates evidence-locked drafting across citation grounding, result grounding, revision, and adversarial robustness using mechanically scored benchmarks without an LLM judge. The validated citation benchmark shows that gating eliminates fabricated citations and surfaces unsupported topics, while other benchmarks provide preliminary evidence of similar benefits and remaining vulnerabilities.

  • Benchmark design: The suite compares gated and ungated drafting across citation grounding, result grounding, evidence-locked revision, and adversarial robustness.Two commercial models were tested under repeated, fixed scenarios, with mechanical scoring and no LLM judge.
  • Citation grounding: 25.2% of gpt-5.2’s distinct ungated citations were fabricated, while neither ungated model flagged any of the 90 unsupported trap subtopics.gpt-4o-mini instead covered all 60 of its trap subtopics with uncited prose.
  • Citation grounding: Zero fabricated citations occurred among 168 gated citations, and 97.8% of trap subtopics received explicit [TBD: ...] placeholders instead of unsupported coverage.Out-of-scope citations also fell to zero under gating.
  • Result grounding: Under preliminary result grounding, ungated fabrication affected 100% of gpt-4o-mini traps and 73% of gpt-5.2 traps, falling under gating to 27% and zero, respectively.Gated drafts surfaced the corresponding gaps as placeholders in 73% and 100% of cases.
  • Revision and robustness: Evidence-locked revision eliminated added superlatives and roughly halved drift, but residual hedge removal remained; adversarial injection caused zero breaches, while authority escalation breached 72% of weaker-model traps and 8% of stronger-model traps.These findings indicate benefits from prompt-level gating alongside vulnerabilities that motivate orchestration-layer enforcement.
  • Benchmark design: The primary validated evaluation uses real arXiv papers with externally verifiable ground truth, whereas the other three benchmarks use synthetic or illustrative settings and remain preliminary.The authors limit the empirical claim to the citation-grounding benchmark.

6 Discussion

Paper Pilot reframes LLM-assisted manuscript generation as a governance-oriented, human-authorized workflow in which evidence checks, approval gates, and traceability constrain drafting and revision.

  • Approval gates: Eight mandatory approval gates govern progression from scope and literature through method, evidence, claims, review override, section approval, and final release.
  • Approval gates: No-pass criteria stop progression when scope, citations, implementation details, numbers, plots, assumptions, or placeholders remain unresolved.
  • Human authority and reviewer boundaries: Paper Pilot assigns manuscript owners final authority over scientific validity, methodological adequacy, evidence sufficiency, and manuscript release.
  • Evidence traceability: Artifact-level traceability links reported numbers, plots, tables, and claims to literature, user evidence, or computational artifacts before approval.
  • Empirical evidence: Prompt-level enforcement halved revision drift without eliminating it, motivating stronger orchestration-layer enforcement.

7 Limitations and Future Work

Paper Pilot is currently a workflow-level specification whose prompting and human review provide partial enforcement, while several components and evaluations remain incomplete.

  • Current implementation: The current implementation lacks an executable orchestration layer that automatically validates schemas, blocks unsupported claims, or prevents unapproved revisions.
  • Evidence schema: The evidence record is lightweight and does not yet implement automated artifact verification, formal provenance reasoning, or integration with established provenance standards.
  • Claim classification: Claim extraction and classification remain semi-automated, requiring manuscript-owner validation without reported accuracy or cross-domain consistency measures.
  • Evaluation scope: Only citation grounding is empirically validated on real arXiv papers; result-grounding and revision benchmarks use synthetic illustrative data.
  • Future work: Future work should evaluate strong baselines, approval time, cognitive load, traceability, inter-rater agreement, reviewer drift, and an applied-science case study.

8 Conclusion

Paper Pilot adapts CARE into a human-authorized, evidence-traceable workflow for LLM-assisted scientific manuscript development. Its citation benchmark supports the framework’s grounding mechanism, while other benchmarks remain preliminary and expose residual enforcement limits.

  • Conclusion: Paper Pilot addresses governance risks by requiring human approval, traceable support, and explicit no-pass criteria across manuscript development.
  • Conclusion: LLM agents assist with drafting, literature positioning, claim classification, revision, and gap identification, while the manuscript owner retains approval, override, and release authority.
  • Conclusion: The framework defines approval gates across the idea-to-claim pipeline and reserves scientific validity, methodological adequacy, evidence sufficiency, and final release for the human author.
  • Conclusion: The citation-grounding benchmark on real arXiv papers found ungated drafting fabricated up to a quarter of citations, while evidence-locked rules eliminated fabricated citations and surfaced planted gaps as placeholders.
  • Conclusion: Paper Pilot provides a governance-oriented framework for producing, revising, approving, and tracing LLM-generated manuscript text to evidence.

Declarations

The declarations report author contributions, human review of Paper Pilot-assisted manuscript preparation, released benchmark resources, and no funding, conflicts, or human-subject research.

  • Author contributions: Jha and Chaudhary contributed to conceptualization, methodology, and writing, while Kulkarni contributed software, validation, testing, benchmarking, and engineering support.
  • AI-use statement: The authors used Paper Pilot for drafting and review support, with all AI-generated outputs reviewed, revised, and approved by the manuscript owner.
  • Data and materials: Benchmark scenarios, raw generated drafts, mechanical scoring records, citation-verification tooling, and scoring code are released with the paper’s companion repository.
  • Ethics: No human-subject or observational dataset was generated or analyzed, and the study involved no participants, animals, or identifiable personal data.
  • Funding and competing interests: The study received no specific grant, and the authors declared no known competing financial interests or personal relationships.

A Benchmark System Prompts

The benchmark compares two system prompts under an identical user prompt, with the gated condition condensing Paper Pilot’s evidence-locked drafting rules for single-turn drafting.

  • Both benchmark conditions received the same scenario, six required subtopics, allowed literature set, fetched metadata, and citation requirement.
  • The gated prompt operationalizes Sections 3.5 and 3.6 as condensed evidence-locked rules for single-turn drafting.

A.1 Ungated condition

The ungated condition asks for polished, well-cited scholarly prose while allowing citations from literature beyond the papers provided by the user.

  • The ungated prompt requests polished scholarly prose with inline citations containing author, year, and arXiv identifier.
  • The drafter may use any literature it knows in addition to the papers supplied by the user.

A.2 Gated condition (Paper Pilot drafting rules)

The gated condition restricts drafting to user-provided sources, prohibits fabricated citation information, and requires placeholders when evidence is unavailable; unsupported citations fail approval.

  • Paper Pilot may cite only papers explicitly provided by the user and must not invent citation details or findings.
  • Unsupported subtopics require an exact evidence-needed placeholder rather than coverage from memory.
  • A citation outside the provided set causes the draft to fail its approval gate, whereas placeholders are acceptable.
Loading 2608.28596v1…