cs.CLSep 5, 2026
SaveCheckpoints Are Not Enough: Trust Calibration in CoSLR, a Human-AI System for Systematic Literature Reviews
Organizations: Tampere University, Tampere, Finland
Abstract
Systematic Literature Reviews (SLRs) are essential for evidence-based research but remain time-consuming, requiring researchers to manage large volumes of publications across planning, screening, analysis, and reporting. Large language models (LLMs) can now produce fluent, well-structured review text, which makes it difficult to distinguish synthesis that was verified by a researcher from synthesis that merely appears authoritative. This raises the risk that unverified AI-generated synthesis enters the scholarly record carrying the credibility of a systematic review. We present CoSLR, a Human-AI collaborative multi-agent system that supports the SLR workflow through a modular three-phase pipeline using large language models and Retrieval-Augmented Generation (RAG), and that places explicit, mandatory human checkpoints on the path between generated output and its acceptance. In a survey-based study with 63 participants, the system was received positively: 27 of 63 participants (42.9 percent) rated its usability highly, indicating that the mandatory checkpoints did not come at the cost of a workable interface. However, a checkpoint safeguards the review only if researchers use it to verify: 22 of 63 participants (34.9 percent) reported that they would trust AI-generated summaries and reports without additional human checking after only a short interaction with the system. These findings indicate that Human-AI collaboration can support literature review work, but that the effectiveness of human oversight depends on whether users are willing to exercise it. This is a calibration problem that interface design must address directly, not assume.
Figures & tables
| Study / Tool | Automation Focus | Techniques Used | Main Limitation |
|---|---|---|---|
| Abstrackr ( Wallace et al., 2010 ) | Title and abstract screening | Supervised machine learning with user feedback | Requires substantial manual labeling and offers limited explainability. |
| ASReview ( van de Schoot et al., 2021 ) | Screening prioritization | Active learning with text-based ranking | Automates mainly one phase of the review workflow. |
| Rayyan ( Ouzzani et al., 2016 ) | Collaborative screening support | Workflow management and relevance filtering | Primarily a decision-support platform rather than full automation. |
| DistillerSR ( Hamel et al., 2020 ) | Review workflow management | Machine learning-assisted prioritization | Depends on human review for final decisions. |
| RobotReviewer ( Marshall et al., 2016 ) | Risk-of-bias assessment and extraction | NLP and machine learning | Strongly domain-specific to biomedical evidence synthesis. |
| EPPI-Reviewer ( Thomas et al., 2023 ) | Screening and review management | Automation-supported review workflow tools | Limited end-to-end intelligence across the full review lifecycle. |
Table 1. Comparison of representative works on systematic literature review automation
Figure 1. The SLR process and the points of Human–AI collaboration. Following the Kitchenham and Charters lifecycle ( Kitchenham and Charters, 2007 ) , the review proceeds top to bottom through the stages shown on the left. For each stage, the centre column names the AI agent that produces a draft output and the right column states the researcher’s role. Amber boxes denote mandatory human checkpoints, where a researcher decision is required before the workflow proceeds; dashed boxes denote lighter human oversight. Double-headed arrows indicate the propose–validate cycle that keeps the researcher in control throughout. A workflow diagram of the Systematic Literature Review process. The diagram shows SLR stages arranged vertically, corresponding AI agents that support each stage, and researcher responsibilities. Mandatory human checkpoints and lighter oversight points are marked to show where researchers validate AI-generated outputs.
Figure 2. High-level architecture of the proposed Human–AI collaborative system for automating Systematic Literature Reviews (SLRs). The system is organized into three phases: (1) planning and retrieval, where LLM agents generate research objectives, research questions, and search queries; (2) ingestion and indexing, where retrieved and uploaded papers are processed, chunked, embedded, and stored in a vector index; and (3) grounded interaction and reporting, where retrieval-augmented generation (RAG) enables evidence-grounded analysis and automated report generation. Human checkpoints and trace logs ensure transparency, validation, and reproducibility throughout the workflow. A high-level architecture diagram of the proposed Human–AI collaborative SLR system. The diagram shows three main phases: planning and retrieval, ingestion and indexing, and grounded interaction and reporting. It also shows LLM agents, document processing, vector indexing, retrieval-augmented generation, human checkpoints, and trace logs that support transparency and validation.
Figure 3. RAG-based interactive analysis component used in the proposed system, adapted from Hasan et al. ( Hasan et al., 2025 ) . User queries are matched against indexed document embeddings to retrieve relevant evidence, which is then provided to the language model to generate grounded responses in the interactive query interface. A retrieval-augmented generation component diagram. The diagram shows a user query being converted into an embedding, matched against indexed document embeddings, and used to retrieve relevant evidence passages. The retrieved evidence is provided to a language model, which generates a grounded response for the interactive query interface.
| Section | Focus / Measured Dimension | No. of Questions | Question Type |
|---|---|---|---|
| Demographic Information | Captures participant role, document usage frequency, explored content, and usage setting. | 4 | Categorical / multiple choice |
| Usability and Interaction | Evaluates support for defining research objectives and questions, generating search strings, screening articles, report generation, and question-answering interaction. | 10 | Likert-scale / binary |
| Cognitive Load & User Experience | Measures mental demand, uncertainty during use, and whether the system reduces technical handling effort. | 3 | Likert-scale |
| Learning & Retention | Assesses perceived support for understanding material, identifying key concepts, and improving confidence in discussing or applying content. | 3 | Likert-scale |
| Productivity and Workflow | Evaluates perceived accuracy, usefulness of summaries, efficiency, time savings, retrieval breadth, risk of missing studies, and preferred level of manual intervention. | 8 | Likert-scale / categorical |
| Recommendation & Concerns | Captures adoption intention, preference over traditional approaches, and privacy, security, or ethical concerns. | 3 | Likert-scale / binary |
Table 2. Structure of the post-interaction questionnaire
Figure 4. Distribution of participant roles reported in the evaluation study. Because respondents could select multiple roles, several role combinations are present. A bar chart showing the distribution of participant roles in the evaluation study. The chart reports the number of respondents who selected each role, such as researcher, student, professional, and teacher. Since participants could select more than one role, the counts are not mutually exclusive.
| Response Excerpt | Initial Code |
|---|---|
| “Add more data sources like IEEE” | More data sources |
| “Search string need to be more precise” | Query precision |
| “The created report properly cited the papers” | Citation traceability |
| “Would like a few more and clearer checkpoints” | Manual checkpoints |
Table 3. Illustrative examples of initial coding in the qualitative analysis
| Code Group | Example Codes | Related Theme |
|---|---|---|
| Planning support | RQ generation, objective formulation, search query generation | Support for RQs and search formulation |
| Synthesis support | report generation, data extraction, structured output, summaries | Report generation and synthesis support |
| Retrieval limitations | more data sources, query precision, missing papers, direct PDF access | Retrieval quality and coverage |
| Evidence transparency | citation traceability, source visibility, seeing analyzed papers | Transparency and evidence traceability |
| User control | manual checkpoints, workflow separation, validation control, context access issues | Workflow control and manual checkpoints |
Table 4. Illustrative code groups used in the qualitative analysis
| Dimension | Representative items | Avg. Mean |
|---|---|---|
| Usability | Q&A interaction, report generation, responsiveness, objective/RQ framing, data extraction | 3.78 |
| Learning support | Understanding material, key concepts, confidence in applying content | 3.68 |
| Productivity | Answer accuracy, report summaries, comprehension support | 3.68 |
| Trust & oversight | Trust in generated outputs without human checking | 3.02 |
Table 5. Grouped descriptive summary of Likert-scale responses
| Item | Mean | Var. | SD |
|---|---|---|---|
| Ease of Q&A interaction | 3.62 | 0.79 | 0.89 |
| Report generation intuitiveness | 3.71 | 0.66 | 0.81 |
| Responsiveness / speed | 3.90 | 0.80 | 0.89 |
| Relevance / accuracy of answers | 3.57 | 0.80 | 0.89 |
| Trust without human checking | 3.02 | 1.24 | 1.11 |
Table 6. Descriptive statistics for selected Likert-scale items (N=63)
Figure 5. Distribution of perceived time savings reported by participants when using the proposed SLR automation system. A bar chart showing participants' perceived time savings when using the proposed SLR automation system. The chart compares three categories of reported time savings: less than 25 percent, 25 to 50 percent, and more than 50 percent.
| Theme | Description | Implication |
|---|---|---|
| Support for research question and search formulation | Participants valued the system’s ability to generate research questions, objectives, and search strings. | Indicates perceived usefulness in the planning stage of the SLR workflow. |
| Report generation and synthesis support | Users appreciated automated synthesis, report generation, and data extraction features. | Suggests that the system is especially valuable for analysis and structured reporting. |
| Retrieval quality and coverage | Participants requested broader database integration and more precise search strings. | Highlights retrieval as a key limitation and a priority for future improvement. |
| Transparency and evidence traceability | Users wanted clearer links between generated outputs and the papers used as evidence. | Reinforces the importance of evidence grounding and explainability in Human–AI systems. |
| Workflow control and manual checkpoints | Several responses emphasized the need for stronger workflow control and manual validation. | Supports the value of human-in-the-loop design rather than full black-box automation. |
Table 7. Themes identified from open-ended survey responses
Figure 6. Distribution of trust in AI-generated output without additional human checking (1 = very low, 5 = very high), overall and split by whether the participant selected the Researcher role. The group difference was not statistically detectable (Mann–Whitney , ); the split is shown for descriptive comparison. A diverging stacked bar chart showing the distribution of trust ratings for three groups: all participants, researchers, and non-researchers. Each bar is centred on the neutral rating and shows the percentage of participants at each point of a five-point trust scale.