Scope at a glance
How to read this publication
- Scope
- Scope: this overview examines scientific and technical literature workflows, especially evidence synthesis, provenance, verification, and replication. It is the first technical strand within the broader AI for Knowledge Creation research direction, not a complete account of every form of knowledge creation.
- Stance
- Publish evidence with an inspectable provenance trail. Revise, hold, or escalate when predefined verification and uncertainty criteria are not met.
- Decision
- Publish / revise / hold
Evidence taxonomy
Different claims carry different weight
Literature finding
Conclusion or result reported by an external source.
Proposed method
Architecture or workflow inAi recommends or intends to evaluate.
Prototype
Implemented but experimental.
Current practice
Actually used by inAi today.
Research target
Desired threshold or future outcome, not a demonstrated result.
Open question
Unresolved.
Verification-first thesis
Abstract and scope boundary
Open scholarly indexes now expose research at a scale no individual can inspect manually. Without provenance and verification, LLM-assisted reading can amplify useful evidence alongside low-quality, fabricated, or machine-generated material. This overview proposes a verification-first workflow that combines provenance graphs, citation-context checks, scientific error benchmarks, replication packages, and calibrated publish, revise, or hold decisions. Any statistical guarantee applies only to the defined loss, calibration data, scoring function, and assumptions; it is not a certificate that a long-form synthesis is true.
Literature findings
State of the field—with proxy limits visible
Provenance and citation integrity
- FineCite improves fine-grained citation-context extraction; separate citation-function classifiers such as SciCite can then label whether a citation provides background, uses a method, supports a result, or contrasts with prior work. These signals should assist review rather than determine reference validity by themselves.
- Supporting, contrasting, and mentioning classifications are contextual signals rather than truth labels; they should be inspected at the citation-statement level and combined with source quality and claim-level evidence.
- Citation integrity debt: unresolved problems in a claim’s source chain, such as missing identifiers, metadata mismatches, inaccessible evidence, incorrect citation context, unsupported inference, or transformations that break traceability.
Novelty estimation
- SchNovel provides a useful evaluation proxy, but its labels assume that the later paper in each pair is more novel. Publication recency should therefore not be treated as ground-truth novelty.
- GraphMind — graph- and retrieval-assisted novelty classification; 3,063 ICLR/NeurIPS papers; accuracy 0.52–0.74 depending on model; peer-review originality scores used as proxy labels; rationale comparisons used GPT-4o as judge; EMNLP 2025 demo paper.
- Post-vintage update — 2026 work strengthens the case for caution. NovBench and RINoBench report substantial gaps between model novelty judgments and expert evaluation, while axiomatic testing shows that no single existing novelty metric behaves reliably across all tested criteria. Novelty scores should therefore be treated as evidence-organising signals, not final verdicts.
Scientific verification
- Across repeated runs, models rarely rediscovered the same errors, so verification performance is not only low but unstable.
- Evaluate candidate verification methods against SPOT-style paper-level error-detection tasks.
- PRISMM-Bench v3 overall model range 27.8%–53.9%
Three structured tasks across 384 reviewer-flagged inconsistencies from 353 papers; accepted at ICLR 2026. 4. PRISMM-Bench evaluates structured identification, remedy, and matching tasks. Its scores should not be interpreted as end-to-end peer-review reliability. 5. CliniFact contributes 1,970 clinical claim–evidence instances for domain-specific factual and logical claim verification against linked publication abstracts. It should be treated as a clinical benchmark, not a general paper-contradiction dataset.
Replication workflows
- Use the reporting guideline appropriate to the review type—for example PRISMA 2020 for systematic reviews and PRISMA-ScR for scoping reviews—and pair it with separate provenance and artifact-packaging standards.
- Include a datasheet when a dataset is released, a model card when a trained model is released, and a general artifact manifest for every package.
- Keep the manifest and runnable example lightweight. Archive larger data, models, and outputs through persistent repositories with identifiers, hashes, licenses, and versioned links. A package should document code, environment, scripts, parameters, seeds where meaningful, expected outputs, hardware and runtime requirements, data-access conditions, and provenance.
- Full reproduction / Partial reproduction / Method inspection only / Access-controlled reproduction
- EuroSys artifact evaluation from 2021–2025 reports 161 Artifacts Available, 136 Functional, 75 Results Reproduced, and approximately 58% participation among accepted papers. These figures describe venue-specific artifact-evaluation outcomes, not a universal reproducibility rate.
Human-time savings
- Abstract screening — 119,691 records; best ensemble at perfect sensitivity reduced workload by 41.81%; performance varied substantially by review and requires domain-specific validation.
- Clinical evidence synthesis — TrialReviewBench: 100 systematic reviews and 2,220 studies; in a human-AI pilot, recall improved by 71.4% relative to manual work and screening time fell by 44.2%; clinical-domain result.
- Time savings are workflow-, domain-, model-, threshold-, and reviewer-dependent. Report the operating point and any sensitivity or recall tradeoff beside the time-saving figure.
Conformal risk control
- Conformal risk control can calibrate a decision rule for a specified measurable loss on held-out data under stated assumptions. A deployment must define the loss, scoring function, calibration set, sample size, target risk, and monitoring process. The guarantee applies to that formal setup; it does not certify the truth of a long-form synthesis, eliminate unknown errors, or remain valid automatically under domain shift.
- An output may proceed only when required source and metadata checks pass, no unresolved detected contradictions remain, and any calibrated decision rule meets its stated operating conditions. Otherwise the system should revise, hold, or escalate the output, recording the evidence, model and tool versions, threshold configuration, reviewer decision, and rationale.
Research integrity and paper-mill pressure
Low-quality, fabricated, machine-generated, or systematically manipulated research material raises the burden on evidence screening. Provenance and integrity signals can support triage, but suspected manipulation requires human investigation and should not be inferred from one automated detector.
Illustrative schema
Claim–evidence provenance
Claim–evidence provenance model
Illustrative claim–evidence provenance model. Citation, support, contradiction, derivation, and review are represented as different relations; transformed outputs retain links to their source versions and processing steps.
Illustrative schema — identifiers and values are placeholders, not results from an inAi study.
Evidence index
Scientific-verification benchmarks remain far from dependable
Scientific-verification benchmarks — version-pinned results
Different tasks and metrics · version pinned · no pooled leaderboard · confidence intervals not reported here
Across repeated runs, models rarely rediscovered the same errors, so verification performance is not only low but unstable.
PRISMM-Bench v3 overall model range 27.8%–53.9%
Three structured tasks across 384 reviewer-flagged inconsistencies from 353 papers; accepted at ICLR 2026.
| Source | Status | Domain / task | Corpus | Reported result | Required caveat |
|---|---|---|---|---|---|
| [SPOT](https://arxiv.org/abs/2505.11855) | Preprint | Paper-level scientific error detection | 83 papers; 91 confirmed errors | Best 6.1% precision; 21.1% recall | Errors were author/human validated; repeated-run rediscovery was poor; benchmark, not production detector |
| [PRISMM-Bench v1](https://arxiv.org/abs/2510.16505v1) | Preprint at the Oct 2025 cutoff | Structured multimodal inconsistency identification, remedy, and pair matching | 262 inconsistencies; 242 papers | 26.1%–54.2% overall model range | Pin v1; not open-ended peer review |
| [PRISMM-Bench v3](https://arxiv.org/abs/2510.16505) | Accepted ICLR 2026 | Updated structured multimodal benchmark | 384 inconsistencies; 353 papers | 27.8%–53.9% | Use only in post-vintage update or update the whole page |
| [GraphMind](https://doi.org/10.18653/v1/2025.emnlp-demos.21) | Peer-reviewed EMNLP 2025 demo | Graph- and retrieval-assisted novelty classification | 3,063 ICLR/NeurIPS papers | GraphMind accuracy 0.52–0.74 depending on model | Median reviewer originality scores are proxy labels; rationales judged with GPT-4o |
| [SchNovel](https://doi.org/10.18653/v1/2025.aisd-main.5) | Peer-reviewed workshop paper | Pairwise scholarly novelty proxy | 15,000 paper pairs; six fields | RAG-Novelty improves over reported baselines | Later paper is assumed more novel; recency is a noisy proxy |
| [FineCite](https://doi.org/10.18653/v1/2025.findings-acl.1259) | Peer-reviewed Findings ACL 2025 | Citation-context extraction | 1,056 manually annotated contexts plus public benchmarks | Up to 25% improvement on citation-context analysis | Not itself a general citation-integrity or truth model |
| [CliniFact](https://doi.org/10.1038/s41597-025-04417-x) | Peer-reviewed Scientific Data 2025 | Clinical claim verification | 1,970 claim–evidence instances | Benchmark dataset | Clinical and abstract-level evidence; not generic contradiction mining |
| [JAMIA screening](https://doi.org/10.1093/jamia/ocaf050) | Peer-reviewed 2025 | Systematic-review abstract screening | 119,691 records | 41.81% workload reduction at perfect sensitivity for best ensemble | Performance varied across reviews; domain-specific validation needed |
| [TrialMind](https://doi.org/10.1038/s41746-025-01840-7) | Peer-reviewed npj Digital Medicine 2025 | Clinical evidence synthesis | 100 reviews; 2,220 studies | Human-AI pilot: +71.4% recall; −44.2% screening time | Pilot result in clinical evidence synthesis, not universal effect |
| [EuroSys AE](https://doi.org/10.1145/3736731.3746152) | Peer-reviewed 2025 | Artifact evaluation | 2021–2025 venue records | 161 Available; 136 Functional; 75 Results Reproduced; ~58% participation | Counts and participation are venue-specific; not a universal reproducibility percentage |
The records use different domains, task formulations, corpora, and metrics. They form an evidence index, not one comparable leaderboard.
Proposed method
Replication and publish / revise / hold
Replication package completeness
Research claim → source and version manifest → code and environment → data or access statement → run instructions → expected outputs and hashes → reproduction scope → independent outcome
Publish / revise / hold decision
Detectors generate signals · reviewers interpret evidence · calibration governs a defined loss · no layer certifies truth alone
A publish-or-hold decision combines source integrity, claim-level verification, scope review, and a calibrated operating rule. Statistical calibration does not replace evidence review.
Workflow patterns
Method map
- No index is complete. Records should retain source-of-record, retrieval date, identifier, version, access status, and correction or retraction state; duplicate records should be resolved without erasing source-specific metadata.
- Extract candidate claims from text, tables, and figures; retain exact source spans and modality; evaluate verification methods using SPOT-, PRISMM-, and domain-specific tasks such as CliniFact; route unresolved cases to domain review.
- Draft from sources that pass documented metadata, eligibility, and quality checks. Preserve source status and unresolved concerns rather than converting source quality into a single clean/not-clean label.
- Use multiple retrieval views and include an explicit search for disconfirming or out-of-neighbourhood evidence. Neighbourhood methods organize attention; they do not establish completeness.
- Package code, environment, scripts, parameters, expected outputs, data-access conditions, identifiers, hashes, and provenance. Use an artifact manifest for every package; add a datasheet for released datasets and a model card for released trained models. Store large objects in persistent repositories and link them through versioned identifiers rather than enforcing a universal bundle-size ceiling.
- Use model disagreement as one escalation signal. Agreement alone does not clear a claim; high-stakes outputs require independent evidence-level validation and appropriate human review.
- Monitor citation-function distributions for anomalies and inspect unusual concentration. Do not enforce a universal quota across document genres or disciplines without validation.
- Transfer principle: business packets can reuse provenance and claim-trace methods, but source quality, timeliness, authority, confidentiality, and legal significance must be evaluated for the business domain. A resolving URL is not evidence that a commercial claim is true.
- Email may notify reviewers, but the authoritative record should be a versioned, access-controlled audit log containing the decision, reviewer, timestamp, source versions, evidence links, model and tool versions, threshold configuration, and rationale.
2026 update
Post-vintage research note
Post-vintage update — 2026
Research published after the original October 2025 cutoff reinforces two conclusions. First, automated novelty judgment remains difficult: NovBench and RINoBench report substantial gaps between model evaluations and expert judgment, even when model rationales appear plausible. Second, scientific error detection remains incomplete: PaperAudit-Bench and FLAWS expand evaluation beyond the original SPOT and PRISMM settings, but current systems still miss many substantive errors.
New work on conformal risk control also sharpens the boundary around statistical guarantees. Calibration can control a defined loss under stated conditions, but strict risk targets may be infeasible for difficult tasks, and distribution shift requires explicit monitoring or adaptation. These developments strengthen the case for evidence-level review, visible abstention, and human escalation rather than automatic publication.
Open questions
What remains unresolved
35.1 Expert/model divergence in novelty
How should a system respond when model rationales appear persuasive but the final novelty judgment diverges from domain experts?
Relevant sources: RINoBench, NovBench.
35.2 Proxy validity
Which novelty metrics measure novelty rather than impact, paper quality, reviewer preference, age, or field popularity?
Relevant source: Axiomatic novelty benchmark.
35.3 Natural versus synthetic errors
How should results be compared across author-confirmed errors, reviewer-flagged inconsistencies, and synthetically inserted flaws?
Relevant sources: SPOT, PRISMM, FLAWS, PaperAudit-Bench.
35.4 Process-level auditability
How can retrieval, extraction, transformation, calculation, model inference, and human editing be represented as one inspectable provenance chain?
Relevant standards: W3C PROV and Workflow Run RO-Crate.
35.5 Calibration under domain and workflow shift
When a model, prompt, corpus, field, or task changes, how quickly do calibration guarantees become unreliable, and what triggers recalibration?
Relevant sources: domain-shift-aware conformal work and 2026 CRC feasibility analysis.
35.6 Multimodal and numerical verification
How can systems verify values and relationships across prose, tables, charts, equations, supplementary files, and code without reducing the task to multiple choice?
35.7 Human audit burden and overreliance
Does a verification assistant reduce work, or does it produce enough plausible false alarms that expert review becomes slower or more biased? How should review interfaces expose uncertainty without creating automation bias?
Primary sources
References and correction architecture
Open primary reference list
Provenance, citation context, and citation function
- Jantsch et al. FineCite: A Novel Approach for Fine-Grained Citation Context Analysis. Findings of ACL 2025. DOI
- Cohan et al. Structural Scaffolds for Citation Intent Classification in Scientific Publications. NAACL 2019. DOI
- Nicholson et al. scite: A smart citation index that displays the context of citations and classifies their intent. Quantitative Science Studies, 2021. Article
- W3C. PROV-O: The PROV Ontology. Standard
- W3C. PROV Model Primer. Primer
Novelty
- Lin et al. Evaluating and Enhancing Large Language Models for Novelty Assessment in Scholarly Publications. AISD 2025. DOI
- da Silva et al. GraphMind: Interactive Novelty Assessment System for Accelerating Scientific Discovery. EMNLP 2025 System Demonstrations. DOI
- Wu et al. NovBench: Evaluating Large Language Models on Academic Paper Novelty Assessment. Findings of ACL 2026. DOI
- Schopf and Färber. Is this Idea Novel? An Automated Benchmark for Judgment of Research Ideas. 2026 preprint. arXiv
- Liu and Zhai. An Axiomatic Benchmark for Evaluation of Scientific Novelty Metrics. 2026 preprint. arXiv
Verification and error detection
- Son et al. When AI Co-Scientists Fail: SPOT—a Benchmark for Automated Verification of Scientific Research. 2025 preprint. arXiv
- Selch et al. PRISMM-Bench: A Benchmark of Peer-Review Grounded Multimodal Inconsistencies. Accepted ICLR 2026. Current arXiv · v1
- Zhang et al. A dataset for evaluating clinical research claims in large language models. Scientific Data, 2025. DOI
- Tu et al. PaperAudit-Bench: Benchmarking Error Detection in Research Papers for Critical Automated Peer Review. 2026 preprint. arXiv
- Xi et al. FLAWS: A Benchmark for Error Identification and Localization in Scientific Papers. 2025 preprint. arXiv
Evidence synthesis and human time
- Sanghera et al. High-performance automated abstract screening with large language model ensembles. JAMIA, 2025. DOI
- Wang et al. Accelerating clinical evidence synthesis with large language models. npj Digital Medicine, 2025. DOI
- Page et al. The PRISMA 2020 statement. BMJ, 2021. DOI
- Tricco et al. PRISMA Extension for Scoping Reviews. Annals of Internal Medicine, 2018. DOI
Risk calibration
- Angelopoulos et al. Conformal Risk Control. ICLR 2024. OpenReview
- Cherian et al. Large language model validity via enhanced conformal prediction methods. NeurIPS 2024. Paper
- Kotte. When Can Conformal Risk Control Certify LLM Outputs? 2026 preprint. arXiv
- Domain-Shift-Aware Conformal Prediction for Large Language Models. 2026 version. arXiv HTML
Reproducibility, artifacts, and software metadata
- D’Elia et al. Lessons Learned from Five Years of Artifact Evaluations at EuroSys. ACM REP 2025. DOI
- ACM. Artifact Review and Badging. Guidance
- RO-Crate. Research Object Crate. Specification portal
- Workflow Run RO-Crate. Workflow execution provenance profiles. Specification portal
- Gebru et al. Datasheets for Datasets. Communications of the ACM, 2021. DOI
- Mitchell et al. Model Cards for Model Reporting. FAT* 2019. DOI
- Citation File Format. CITATION.cff. Specification
- CodeMeta. Research software metadata vocabulary. Project
- Wilkinson et al. The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data, 2016. DOI
Research integrity
Open technical glossary
- provenance;
- citance;
- citation context;
- citation intent;
- bibliographic coupling;
- co-citation;
- novelty proxy;
- contradiction detection;
- replication / reproduction / repeatability;
- WSS@95;
- conformal risk control;
- abstention;
- exchangeability;
- domain shift;
- paper mill.
Version history
Publication record
Version 1.1 — corrected source dates and benchmark metadata; clarified proposed versus implemented methods; added references and post-vintage notes.