The Era of AI Speed-Reading: Who Still Reads Literature Word by Word? — A Field Report on the Truth of AI Literature Review Tool Usage

At a research group meeting of a research-oriented university in early 2026, a second-year master’s student was found to have obvious errors in the titles of three core references in their literature review. Upon inquiry, the student admitted that the review draft was entirely generated by AI tools, with only simple formatting adjustments made, and they had never read or verified the original texts of the cited literature.

This is not an isolated case but a prevalent phenomenon in current academic research. The popularization of AI literature tools has completely revolutionized the traditional model of literature reading and review writing. Behind the dramatic improvement in efficiency, prominent problems such as content distortion, thinking inertia, and academic anomie have become increasingly severe.

Meanwhile, “Write For You”, a zero-hallucination AI literature review system developed by the Hong Kong Polytechnic University, has accumulated over 40,000 active users worldwide across dozens of countries and regions since its launch in the first quarter of 2025. Adopting cutting-edge agent technology to overcome AI hallucinations, the system enables fully traceable references and standardized logical reasoning. Users can obtain a complete literature review report within minutes simply by inputting their research topics.

The tension between efficiency improvement and academic authenticity has become a normalized proposition in academic research. To clarify the rational usage logic of AI literature tools and answer the core question of why AI empowers efficient research for some users but leads to academic failures for others, the author conducted a three-month field study. The research tracked the AI tool usage logs of graduate students from six disciplines, conducted in-depth interviews on their usage habits, practical confusions and learning experiences, and adopted 30 valid questionnaires to restore the real application status of AI literature tools. The core gap lies not in the tools themselves, but in differentiated usage strategies.

I. Field Investigation: Real Usage Scenarios in Six Laboratories

The research covered six majors including biomedicine, education, computer science, literature and history, engineering, and economics, involving master’s and doctoral students at different stages. It covered the three major disciplines of science and engineering, humanities, and social sciences, as well as both novice and senior researchers, ensuring comprehensive and representative samples. Adopting the research method of “log retention + weekly in-depth interviews”, respondents voluntarily provided AI tool usage screenshots, and the whole process recorded practical operations, problems and gains to restore real frontline application scenarios.

1.1 Efficient Users: Taking AI as an Assistant and Dominating Academic Judgment Manually

The research finds that researchers proficient in AI literature tools have established a clear human-machine division of labor, always placing manual critical thinking and academic judgment at the core.

A third-year master’s student majoring in computer science provides a representative practical model. When receiving a new research topic, instead of conducting experiments or writing manuscripts directly, they first use AI tools to conduct a comprehensive domain scan, sorting out baseline research methods, mainstream technical approaches and previous research boundaries to quickly grasp the state-of-the-art progress. With the efficient preliminary screening capability of AI, the literature research cycle is shortened from 4 weeks to 10 days. AI undertakes preliminary literature screening, domain context sorting and framework construction, while humans focus on core literature intensive reading and innovation exploration, achieving simultaneous improvement in efficiency and quality.

A first-year doctoral student majoring in education has summarized exclusive verification rules to avoid AI hallucinations from the source. When using AI to interpret literature and sort out viewpoints, they always attach a fixed instruction: “Please output the original text fragments and corresponding page numbers of the cited viewpoints”. The student emphasized in the interview: “AI generates fluent content but tends to fabricate information and misattribute viewpoints. Only tracing back to original texts can avoid misleading.” After each AI output, they conduct 2 to 3 rounds of follow-up verification, and never adopt unvalidated content.

The core commonality of efficient users is defining AI as a scientific research assistant rather than a writing substitute. Liu Zhiyuan, Associate Professor in the Department of Computer Science, Tsinghua University, pointed out that AI’s powerful capabilities in information extraction and correlation integration free researchers from massive literature, helping them abandon inefficient extensive reading and focus more on classic literature reading and in-depth academic thinking, thereby realizing effective technological empowerment for scientific research.

1.2 Inefficient Users: Taking AI as the Standard Answer and Abandoning Manual Critical Thinking

In sharp contrast, most research novices fall into the AI dependency trap, regarding AI as an all-purpose answer generator and completely abandoning manual judgment, literature verification and in-depth thinking.

A third-year doctoral student majoring in literature and history directly submitted an AI-generated review draft with only typesetting adjusted, which was eventually rejected by the supervisor for fragmented argumentation logic, insufficient dialogic correlation between literature, lack of independent research thinking, and simple stacking of literature viewpoints. A first-year engineering graduate student adopted a more extensive approach, directly copying and pasting AI-generated literature interpretation paragraphs without verifying any original texts, resulting in disjointed review content and distorted core viewpoints.

A graduate student who experienced review revision summarized the core misunderstandings of inefficient AI usage: lacking overall framework and logical planning, relying entirely on AI for free generation leads to structural chaos; focusing only on literature narration rather than evaluation results in simple viewpoint stacking without independent sorting, comparison and critical thinking, completely losing the core value of academic reviews. The subsequent effective improvement method is to return to academic essence, sort out domain context based on highly cited classic literature, and optimize details with AI tools.

Zheng Lei, Assistant Professor at the School of Education, Peking University, accurately summarized the differentiation logic of AI usage: “AI presents mirror feedback — it empowers competent users and exacerbates the deficiencies of incompetent ones. Excessive AI dependency among beginners leads to one-sided and even wrong literary cognition, widening the gap in scientific research capabilities. In contrast, researchers with professional judgment revise AI output through continuous questioning and verification, stimulating innovative academic thinking in human-machine interaction.”

In conclusion, the gap in AI literature tool application does not lie in the tool itself, but in whether users adhere to the three core bottom lines of manual judgment, original text verification and in-depth critical thinking. Among them, literature screening capability, widely ignored by most users, serves as the key starting point for differentiating review quality.

II. The Overlooked Bottleneck: Literature Screening Determines Review Quality

A common confusion frequently appears in the research: “Reading 50 pieces of literature with AI still fails to produce a high-quality review”. In-depth analysis reveals that the core problem lies not in literature interpretation capability, but in invalid preliminary screening — more than half of the 50 read literatures have low correlation with the core research theme, and invalid literature consumes massive time and energy, resulting in loose review content and ambiguous key points.

2.1 Breaking Traditional Misunderstandings: Reconstructing the Literature Research Process

Traditional literature reading generally follows a wrong process: retrieval and download → full-text intensive reading → value judgment → retention or abandonment. This “reading before screening” mode wastes plenty of time on invalid literature with extremely low efficiency.

The scientific logic of literature research is screening before reading and hierarchical reading: literature retrieval → preliminary screening based on titles, abstracts and keywords → secondary screening via rapid full-text browsing → in-depth intensive reading only for core literature.

Zhang Xiaodan, librarian of Huazhong University of Science and Technology, systematically sorted out the standardized five-step literature research process in academic lectures: topic analysis, literature collection, screening and analysis, reading and understanding, review writing. She emphasized that literature screening is the foundation of scientific research. Accurate screening helps researchers quickly identify core domain achievements, research hotspots and controversial issues, clarify research directions, and effectively avoid repetitive research.

Chen Ben, a graduate student at the School of Education, Peking University, vividly interpreted the screening value of AI: the literature outline, abstract and keywords generated by AI are like a scientific research “navigation map”. Even with minor deviations, they provide clear research entry points, helping novices get rid of confusion in the face of massive literature and quickly establish domain cognitive frameworks.

The core empowering value of AI lies in optimizing preliminary screening and secondary screening links, greatly reducing ineffective time consumption, and enabling researchers to focus on core literature intensive reading, viewpoint refinement and critical innovation, so as to achieve accurate efficiency improvement in scientific research.

2.2 Capability Priority: AI Screening Is Far More Reliable Than AI Summarization

Among various functions of AI literature tools, correlation ranking, topic clustering and literature screening are far more reliable than viewpoint summarization and content induction. The core reason is that literature screening only relies on pattern matching for objective classification without complex value judgment; while content summarization, viewpoint refinement and cross-literature comparison require logical reasoning, difference discrimination and controversy sorting, which are prone to AI hallucinations, viewpoint misattribution and logical confusion.

Multiple authoritative studies have confirmed that AI achieves the most significant efficiency improvement in the literature screening stage. Relevant evidence synthesis studies in 2025 show that AI tools reduce workload by 55%-64% in title and abstract screening, greatly lowering manual screening pressure. In the whole process of systematic review, AI presents prominent advantages in preliminary screening, while its empowerment effect is limited in literature quality evaluation and refined data extraction, with effective efficiency improvement only in a few scenarios.

A November 2024 study published inNature clarified the inherent shortcomings of general large language models: when sorting out academic topics, general AI tools such as ChatGPT tend to mix and splice information from authoritative academic literature, informal blogs and unknown sources, resulting in mixed and distorted content. Meanwhile, general models can only retrieve open-access abstracts and text fragments instead of full texts, leading to incomplete information input and inherent interpretation deviations.

Respondent researchers summarized a practical ironclad rule: prioritize AI for literature correlation judgment, and cautiously adopt AI for independent literature viewpoint summarization. Entrust screening to AI and reserve intensive reading and summarization for manual work — this is the most efficient and reliable application mode.

III. The Truth of AI Hallucinations: Defining the Credibility Boundary of AI Literature Output

The biggest hidden danger brought by the popularization of AI tools is researchers’ thinking inertia and lack of critical thinking. A survey of university academic status conducted by Science and Technology Daily found that “being lazy with AI” and “losing independent critical thinking ability” are common problems among students. Yang Yuxi, an undergraduate student at the Department of Chinese Language and Literature, Peking University, admitted that under the heavy burden of literature reading tasks, directly adopting AI-generated content has become a highly tempting shortcut, and long-term dependency leads to continuous degradation of autonomous reading, sorting and critical thinking capabilities.

A December 2025 academic review published in The American Journal of Surgery systematically defined the core risks of AI-assisted scientific research, clarifying that content hallucinations and cognitive overreliance are two major hidden dangers, specifically manifested as fabricated references, logical reasoning errors, failure to identify literature controversies, and confusion of differentiated research conclusions, which are the main causes of current academic review quality failures.

3.1 Three-Tier Credibility Classification: Accurately Judging the Value of AI Output

Combined with field research data and authoritative literature conclusions, AI literature output can be divided into three credibility levels with corresponding application specifications:

Category 1: Basic information extraction of single literature (Credibility: ★★★☆☆). It covers explicit information such as literature titles, research methods, basic conclusions, authors and journals. AI achieves high accuracy in extracting such objective information but tends to omit key content such as research constraints, experimental details and conclusion applicability boundaries, requiring manual verification and supplementation of original texts before adoption.

Category 2: Core viewpoint positioning of single literature (Credibility: ★★★★☆). Under the constrained instruction of “output original text fragments and mark page numbers”, AI can quickly locate core viewpoints, key experimental data and core argument paragraphs of literature. Output credibility is greatly improved in this mode, which can serve as an auxiliary intensive reading tool, though manual verification of original texts is still required.

Category 3: Comprehensive comparison and thematic induction of multiple literatures (Credibility: ★★☆☆☆). This is the most error-prone and least credible scenario for AI. The algorithm logic of AI tends to integrate common contents and eliminate research differences, while the core value of academic reviews lies in sorting out research divergences, viewpoint controversies, domain evolution and research gaps. Many respondents reported that AI frequently misattributes viewpoints, confuses conclusions and covers up controversies when inducing multiple literatures with similar themes, which is highly misleading.

3.2 Core Misunderstanding: Structural Mismatch Between AI Comprehensive Induction and Academic Review Requirements

Nature once made a straightforward comment: relying entirely on large AI models to write academic reviews from scratch is an immature scientific research behavior. The core defect of AI cross-literature induction stems from the essential mismatch between technical logic and academic needs.

Technically, the core goal of AI content generation is fluent sentences, coherent logic and unified content, tending to merge similar viewpoints and avoid contradictory controversies; while high-quality academic reviews require accurate identification of methodological differences, conclusion divergences, research limitations and domain controversial focuses across different studies.

Practically, complex academic deduction and refined case analysis are high-frequency error scenarios for AI. Tan Ziwei, an undergraduate student majoring in philosophy at Peking University, shared his learning experience: initial reliance on AI for professional theory interpretation and logical deduction analysis presented complete and fluent content, but in-depth reading revealed multiple key errors, and blind adoption would easily lead to wrong academic cognition.

The review published in The American Journal of Surgery clarified the core positioning of AI: AI is an augmentative tool for scientific research rather than a substitute for human professional expertise, which can only serve as an auxiliary means, and core academic judgment and innovative critical thinking must be completed independently by researchers.

3.3 Rational View on Zero-Hallucination AI Systems

The “Write For You” system developed by the team led by Professor Liu Yan from Hong Kong Polytechnic University adopts cutting-edge agent technology to solve the core pain point of traditional AI fabricated references, realizing 100% traceable references, standardized logical reasoning and cross-lingual academic resource integration, achieving literal zero hallucination in reference information and gaining wide recognition from global researchers since its launch.

However, the technical boundary needs to be rationally distinguished: the zero-hallucination feature of the system only targets the authenticity of reference sources and basic literary information, rather than the complete reliability of AI content interpretation, viewpoint induction and comprehensive judgment. As stated in Nature research, high-quality systematic reviews with transparency and reproducibility still rely on in-depth manual verification and academic critical thinking, and the goal of fully automatic AI review writing is still far from realization.

In short: traceability does not equal verification. Tools can build a credible framework for literature retrieval and sorting, but the core links of academic truth-seeking, content verification and innovative thinking must be completed by researchers themselves.

IV. Efficiency Accounting: The Truth and Hidden Costs of AI Time Saving

Based on statistical data from 30 valid research questionnaires, this paper sorts out the time cost differences among traditional reading, primary AI usage and advanced AI usage, presenting the real efficiency logic of AI empowerment (data refers to trend-based reference reported by respondents).

4.1 Time Cost Comparison of Three Research Modes

Process Mode	Literature Screening	Preliminary Reading	Content Verification	Writing & Intensive Reading	Total Time Cost
Traditional Process	40h	120h	—	40h	Approx. 200h
Primary AI Usage (Blind Adoption)	10h	30h	40h (Post-rework Verification)	80h	Approx. 160h
Advanced AI Usage (Standardized Verification)	5h	15h	25h (Pre-positive Accurate Verification)	40h	Approx. 105h

Core research conclusion: AI does not absolutely save time; efficiency depends entirely on usage strategies. Primary users blindly rely on AI output and concentrate verification and error correction on the later writing stage, resulting in high rework costs and limited overall efficiency improvement. Advanced users advance verification work to avoid content errors and logical loopholes in advance, greatly reducing rework costs and nearly halving total time consumption.

Yang Ziyue, a graduate student at the School of Education, Peking University, provided a valuable practical method: using AI to quickly extract core experimental content and key information from literature to avoid inefficient full-text reading, and conducting targeted intensive reading and detail verification for AI-extracted content, realizing a benign division of labor between efficient extraction and accurate intensive reading.

4.2 Overlooked Hidden Costs of AI Usage

Most users only focus on the explicit time saved by AI but ignore three major hidden time-consuming factors, which are the core reasons why AI sometimes reduces efficiency instead:

First, prompt debugging cost. Accurate literature sorting and viewpoint extraction require standardized instructions, and 3-5 rounds of debugging are needed on average to obtain academically compliant output. As emphasized in The American Journal of Surgery, prompt engineering is the key technology to optimize AI academic output quality and avoid potential risks.

Second, content verification cost. Verifying details, tracing viewpoints and checking logic for AI interpretation of single literature takes 15-20 minutes on average, and verification time accumulates continuously in batch reading scenarios.

Third, error correction cost. AI problems such as viewpoint misattribution, logical deviation and content fabrication require repeated questioning, manual revision and secondary sorting after discovery, bringing additional invalid workload.

4.3 Optimal Paradigm of Efficient Human-Machine Collaboration

Research published on ACL 2025 proposed a mature paradigm for AI scientific research empowerment: InsightAgent, an interactive AI agent system, realizes in-depth human-machine collaboration through multi-agent architecture, semantic literature clustering and real-time visual feedback mechanisms. Verified by 9 medical researchers, the system improves systematic review quality by 27.2%, reaching 79.7% of manual writing quality, increases user satisfaction by 34.4%, and enables clinical researchers to complete high-quality systematic review drafts in only 1.5 hours, compared with months required by traditional methods.

The core advantage of the system is not one-click generation, butreal-time human-machine interaction, dynamic feedback and accurate modification, retaining manual intervention, verification and optimization authority throughout the process, which represents the future development direction of AI-assisted scientific research.

Thus, the core formula for AI scientific research efficiency improvement is concluded: Real Efficiency Gain = Time Saved by AI Rough Processing − Incremental Time for Manual Verification and Debugging. Optimizing usage strategies and reducing hidden costs are essential to realize effective technological empowerment. Zhang Xiaodan from Huazhong University of Science and Technology repeatedly emphasized that AI assistance never equals AI substitution, and only adherence to manual critical thinking can avoid technical risks and ensure academic quality.

V. Human-Machine Division Boundaries: Clarifying Tasks for AI and Humans

Based on research data, authoritative literature conclusions and frontline practical experience, this paper sorts out a standardized task division table for AI literature tools, clarifying human-machine authority boundaries and balancing efficiency and academic rigor.

5.1 AI Literature Task Division Table

Task Type	AI Credibility	Practical Operation Suggestions	Core Basis
Literature Retrieval and Correlation Screening	★★★★☆	Fully entrusted to AI with secondary manual review	Evidence-based studies confirm AI achieves the most significant efficiency improvement and highest accuracy in literature screening
Basic Literature Information Extraction	★★★★★	Fully automated without repetitive manual work	Objective information such as titles, authors and journals has almost no AI error risk
Single Literature Structural Sorting	★★★☆☆	AI generates preliminary framework with manual preview and optimization	Serves as intensive reading navigation but cannot replace in-depth manual reading
Single Literature Core Viewpoint Positioning	★★★★☆	AI-assisted positioning + manual sentence-by-sentence original text verification	High credibility under precise instructions; traceability verification completely avoids hallucination risks
Multiple Literature Comparative Summary	★★☆☆☆	Only used for draft reference with full manual viewpoint verification	AI cannot effectively identify literature controversies and research differences, prone to viewpoint confusion
Research Gap and Innovation Point Extraction	★☆☆☆☆	AI usage is not recommended; fully completed manually	Requires domain academic accumulation, research intuition and critical thinking irreplaceable by AI
Reference Format Sorting	★★★★★	Fully automated typesetting and proofreading by AI	Standardized rules without subjective deviation; AI efficiency and accuracy surpass manual work

5.2 Four Universal Safe AI Usage Principles

Principle 1: Use AI for subtraction rather than addition. The core value of AI is to screen and filter invalid information to help researchers focus on core literature, rather than generate new viewpoints and conclusions out of nothing. Researchers should always be alert to AI biases, hallucinations and opacity, and maintain continuous critical supervision.

Principle 2: Reject un traceable content. Any viewpoints, data and conclusions generated by AI that cannot be accurately positioned to specific original text paragraphs shall be regarded as questionable and prohibited from direct adoption in reviews, avoiding implicit misleading caused by AI forced correlation and fabricated viewpoints.

Principle 3: Never rely on AI abstracts for core literature. The 5-10 key literatures directly related to research themes and core arguments must be fully read and independently sorted out to consolidate the academic foundation of reviews and avoid core content distortion.

Principle 4: Reserve special time for verification. In scientific research time budgeting, no less than 50% of AI usage time should be reserved for content verification, error correction and optimization, advancing risk control to avoid large-scale later rework.

VI. Capability Restructuring: New Academic Rules and Competencies in the AI Era

The research reveals a prominent capability differentiation phenomenon: junior graduate students and research novices have the highest AI dependency but the weakest ability to judge and correct AI output errors; senior researchers can rationally utilize AI empowerment while adhering to academic bottom lines. This capability inversion is reshaping the core competency structure of modern scientific researchers.

Liu Zhiyuan pointed out that AI-assisted scientific research does not lower research thresholds or weaken thinking abilities, but puts forward higher requirements for researchers. In the context of technological iteration, researchers must master the ability to identify AI errors, integrate effective information and conduct independent innovative thinking, and never abandon subjective academic judgment. The American Journal of Surgery also warned that unrestricted AI dependency will continuously weaken academic autonomy and become a hidden cause of scientific research capability degradation.

6.1 Dialectical View on the Elimination and Reconstruction of Literary Intuition

There is a widespread academic concern: traditional researchers accumulate academic intuition and domain cognition by reading hundreds of literatures; will researchers lose academic sensitivity and research judgment if only a small number of core literatures are read after AI screening?

This concern requires dialectical analysis. Zheng Lei from Peking University believes that academic laziness and opportunistic learning behavior induced by AI are essentially extensions of traditional academic irregularities, and AI only amplifies existing problems rather than creating new ones. The core value of AI is to help new researchers quickly complete domain panoramic scanning, shortening the half-year academic entry cycle to 1-2 months, freeing researchers from inefficient extensive reading and enabling earlier participation in core links such as experimental verification and innovative research.

True academic intuition lies not in extensive reading, but in intensive reading and differentiated discrimination. The mode of AI screening for high-quality literature plus manual in-depth reading helps researchers accurately accumulate high-quality academic cognition and reshape efficient and precise modern research intuition.

6.2 Standardized Implementation: Institutional Boundaries of AI Scientific Research Usage

With the in-depth integration of AI into academic research, successive standardized guidelines have been issued, marking that AI-assisted scientific research has officially entered the era of standardized governance from free exploration.

In May 2026, the Chinese Society for Graduate Education issued the Guidelines for Standardizing the Application of Artificial Intelligence Tools in Graduate Theses and Practical Achievements, the first guiding document regulating AI tool usage in domestic graduate education. It clarified four core principles: first, user responsibility, with students as the primary responsible subject for academic achievements and supervisors undertaking guidance and review responsibilities; second, independent core innovation, requiring core arguments, research ideas and innovative contributions to be independently completed by researchers and prohibiting AI ghostwriting and plagiarism; third,full transparent disclosure, requiring active AI usage statements in papers to truthfully mark tool names, versions, purposes, application links and verification processes; fourth, regular AI usage inquiry in defenses, setting special sessions to verify standardized AI application in thesis defenses.

Previously, the Chinese Academy of Sciences issued scientific research integrity reminders, explicitly prohibiting direct adoption of unverified AI-generated research reports, literature reviews and topic selection suggestions. Sichuan University issued special specifications, clarifying clear boundaries: AI is allowed to assist in building review frameworks and optimizing formatting, but core viewpoints, research innovations and logical argumentation must be independently completed by students.

A unified consensus has been formed in international academia:Nature has repeatedly published comments calling for explicit AI usage disclosure mechanisms in academic journals; top universities including Harvard University and MIT have successively released AI scientific research guidelines, all adhering to the core principles of technological assistance, human dominance, human responsibility and full transparency.

Meanwhile, the publishing industry is constantly improving the supervision system. A 2025 Issue 9 study in Communication and Copyright pointed out that generative AI intervention in academic writing has triggered new academic problems such as hidden plagiarism, uncertain content authenticity and ambiguous publishing ethics, requiring multidisciplinary collaboration to improve AI content detection technology and academic standard systems, and consolidate the bottom line of publishing and scientific research integrity.

Conclusion

Returning to the opening academic failure case: the graduate student who produced distorted reviews due to AI abuse subsequently adjusted the usage logic. Adopting AI for domain scanning, preliminary literature screening and framework construction, and focusing manual work on core literature intensive reading, viewpoint tracing and innovative critical thinking with full verification of all AI output content, the final revised review was highly recognized by the supervisor.

This is the ultimate value of AI literature tools:AI does not replace researchers in reading, but helps them screen worthy literature. Professor Cao Jiannong from Hong Kong Polytechnic University stated that the team will rely on existing zero-hallucination literature technology to continuously build an AI Super Research Brain covering the whole scientific research process of all disciplines to empower academic innovation. However, no matter how technology iterates, AI can never replace researchers’ independent judgment, in-depth critical thinking and academic innovation.

Tools accelerate scientific research progress, but every step of in-depth exploration, critical thinking and innovative achievement toward academic truth must be completed by researchers themselves.

Utilize AI’s speed to liberate inefficient labor, and rely on human depth to consolidate academic foundation — this is the correct answer for researchers in the era of AI speed-reading.

References

[1] Rashid M, et al. Machine Learning Tools To (Semi-)Automate Evidence Synthesis: A Rapid Review and Evidence Map. Effective Health Care Program, 2025.

[2] Celik S U. Integrating artificial intelligence into scientific writing: a narrative review for clinical and surgical researchers. The American Journal of Surgery, 2025.

[3] Liu Y Y. Common Problems and Countermeasures of Review Articles in Sci-Tech Journals under Generative Artificial Intelligence. Communication and Copyright, 2025(9).

[4] Liu Y, Cao J N. Research and Application of Zero-Hallucination AI Literature Review System “Write For You”. Hong Kong Polytechnic University Research Achievements, 2025.

[5] Chinese Society for Graduate Education. Guidelines for Standardizing the Application of Artificial Intelligence Tools in Graduate Theses and Practical Achievements. 2026.

[6] Pearson H. Can AI review the scientific literature — and figure out what it all means?. Nature, 2024, 635(7962): 276-278.

[7] Qiu R, Chen S, Su Y, et al. Completing A Systematic Review in Hours instead of Months with Interactive AI Agents. Proceedings of the 63rd ACL, 2025.

[8] Sichuan University. Trial Specifications for the Application of Artificial Intelligence Tools in Undergraduate Education and Teaching. 2025.

[9] Chinese Academy of Sciences. Integrity Reminders for Standardizing the Application of Artificial Intelligence Technology in Scientific Research. 2025.