# Agent Descriptions Total papers included: 259 --- ## 312_Superior Energy Storage Performance in High-Entrop The main writing of the paper, including the creation of all figures, was completed by the AI. The role of humans was not that of a primary author, but rather to guide the AI’s writing direction through a few prompts. All textual content was generated by the AI. --- ## 328_Indirect Prompt Injection in AI-Native Peer Review The manuscript was generated primarily by an AI system, which produced the hypothesis, structured the argument, drafted the prose, and designed the conceptual framework and figures. The AI is listed as the sole first author. --- ## 280_Modular and Hybrid Frameworks for LLM-Based Agents The paper was generated by AI. We provided the AI with our data, relevant literature, and prompts instructing it to follow the rubrics established in the Agent4Science AI paper. We subsequently reviewed the draft and made only minimal modifications. --- ## 304_Backtest to the Future: Can Large Language Models The writing process was done almost entirely by AI. Humans only provide suggestions for formatting and citation issues. --- ## 186_Long-range Nonreciprocal Ising Model AI carries out at least 95% of the text-writing and all of the figure-making. Humans are mainly responsible for providing high-level guidance. --- ## 183_Hybrid End-to-End Knowledge Graph Construction and use Cursor.aiforalltheevaluationcodeandinnovai.pro AI-copilotforwriting allusedinnovai.pro AI-copilotforwriting --- ## 289_Hardware-Conditioned Generative Channel Modeling: AI is able to write 80% of the paper but requires human input for the outline of every section and restructuring the presentation. AI is also terrible at making diagrams. --- ## 40_Feasibility‑Guided Fair Adaptive Reinforcement Lea The manuscript was drafted primarily by the AI agent, including the abstract, introduction, methods, results, discussion and ethical considerations. --- ## 37_Computing $\pi$ Using Numerical Methods The AIdraftedtheentiremanuscriptin LATEX,structuredthesections,wrote descriptions,discussionandconclusions,addedcitationsandformattedthebibliography, with the human providing high-level instructions and performing final review. --- ## 225_MosquitoSwarm: Bio-Inspired Collective Intelligenc The AI agent produced all textual content, structured the paper according to conference guidelines, developed mathematical notation and algorithmic descriptions, created comprehensive experimental analysis, and maintained consistent academic writing style throughout. The biological interpretations and connections between mosquito behavior and computational principles were entirely generated by the AI. --- ## 287_Echo: A multi-agent AI system for patient-centered AIcontributedmostofwriting,significanthumaneditsandfigure-creation fromhuman. --- ## 169_From Player to System: An Agent-Based Framework fo AI assisted with literature synthesis, technical explanations of biomechanical concepts, and formatting of the manuscript. --- ## 155_Multimodal Representation Engineering for Robust A AI generated the majority of the paper content based on human guidance and research framework, with human oversight and editing. --- ## 216_Enhancing Pre-Training Data Detection via Multi-La AI automatically carried out all the processes related to writing. --- ## 191_LLM agents for kernel development in science AI produced the entire manuscript (text, tables, figures) from raw logs; humans only removed potential deanonymising phrases. --- ## 261_Towards Superior Cross-Domain Adaptability in Embo AI generated the majority of the draft text and structure, with humans providing oversight, edits, and final refinements for coherence and accuracy. --- ## 128_Molecular Stratification of Renal Cancer Reveals P The AI agent generated hypotheses, explored possible stratification strategies, and refined the specific research questions. The AI agent was responsible for designing computational experiments, coding the analysis pipeline, and implementing data processing and clustering methods. It generated executable code and carried out the experiments without human coding input. The AI handled the complete analysis process: organizing raw mRNA data, performing clustering and survival analyses, retrieving supporting literature, and interpreting patterns in the results. The AI produced the draft content of all sections and generated figures, tables, and narratives describing the results. --- ## 203_Cultural Dynamics in Multi-Agent Systems: Joint Ef The paper content was generated by AI, while humans provided feedback and suggestions, and adjusted the paper format. Experimental figures were created by LLM writing code for visualization. Figure 1 was designed by LLM based on the paper content and generated by a diffusion model. --- ## 164_Beyond Statistical Patterns: Integrating Textual D Claude AI implemented the complete bootstrap consensus framework, statistical methods, and evaluation pipeline. Human involvement included supervision, parameter tuning, and validation of results. The experimental design and computational implementation were primarily AI-generated with human oversight for quality assurance. The manuscript was primarily written by AI systems, including main text, supplementary material, figure creation, and narrative structure. Human involvement was limited to review, minor edits, and final formatting. The scientific writing, methodology descriptions, and result presentations were AI-generated. --- ## 244_Information-Efficient Transformers via Adaptive To AI drafted substantial portions (method, limitations, ethics, reproducibility, checklists); humans edited for accuracy, template compliance, anonymity, and consistency with results. --- ## 271_When Looks Make You Click “Submit”: Participation- Research questions and hypotheses were generated by an LLM. Mediation specifications, robustness checks, and inference procedures were proposed by an LLM and executed accordingly. Statistical analyses and initial interpretations were produced by an LLM; humans enforced descriptive framing and checked for overclaiming. The manuscript was drafted and revised by LLMs; humans facilitated prompts and emphasis. --- ## 21_From Co-Writer to Co-Author? Investigating the Rol Chat GPT generated the detailed research questions, subtopics, and conceptual focus points. These AI-generated ideas were then reviewed, structured, and slightly refined by the human author. Chat GPT took the lead in proposing the conceptual structure and sequencing of arguments, while the human author provided guidance and feedback at each stage. The synthesis of literature and conceptual interpretation were carried out entirely by Chat GPT, without targeted prompting for each interpretative step. Writing was done by the AI. Human authors helped with the chapter headings and their content. Citations, sentence structure, and content were written solely by AI. --- ## 189_Breaking the Data Barrier: LexiCore - A Lexicon-Fi AI (Claude) generated paper text, tables, and technical descriptions. --- ## 103_Analysis of Mock Conversations Across Large Langua Human 1st author prompted Chat GPT Free version and Gemini-2.0-flash for data generation. Human 2nd author prompted GPT-5 thinking model for data generation. Human 3rd author prompted Claude Opus 4.1 for data generation. AI 1st author suggested the list of features to evaluate. Human 1st author prompted AI 1st author to include more features, and AI 1st author suggested additional features. Visualization, Statistical analysis, and machine learning-based evaluation were all suggested by AI 1st author. --- ## 148_The Self-Consistent Hallucination Loop (SCHL): Eme AIsystemsgeneratedthemajorityofthemanuscripttext,includingsection drafts, phrasing, and figure/table examples. Human authors acted as editors: outlining, verifying coherence, and adding clarifications about limitations and ethics. More than 95% of raw text originated from AI. --- ## 110_Hierarchical Adaptive Normalization: A Placement-C The manuscript, including narrative, figures, and layout, was produced largely by AI. Human contributions were limited to light revision and final approval. --- ## 316_The Commoditization Pathways of Graphics Processin AI generated over 95% of the paper content, including abstract, introduction, literature review, methodology, results, discussion, and conclusion sections. AI also created all tables, structured the narrative flow, and formatted the manuscript according to Agents4Science guidelines. --- ## 200_Simulating Strategic Reasoning: A Digital Twin App Chat GPTPro Deep Research was responsible for producing the majority of the paper’s content, including drafting the main text, generating figures, and formulating the overall narrative flow. --- ## 123_An Algorithmic Roadmap to Unification- Proposing a AI wrote most of the experimental design, that is, the three-part structure (generative engine, validation engine, physics-informed reward function) including precise mathematical and algorithmic details. This included all the proofs, complexity analyses, implementation details (symbolic regression, reinforcement learning architecture, etc), and benchmarking procedures. I only did high-level guidance of what to focus on (quantum error-correcting codes), kept AI from going off on purely mathematical tangents that are physically irrelevant, and organized the overall presentation. I came up with the high-level design, broad ideas, and strategic choices, AI with the technical design, mathematical details, and implementation choices that go into the experiments. --- ## 22_Fairness Agents in Scientific Collaboration: A Res The LLM generated detailed subquestions and conceptual directions that were reviewed and structured by the human author. The LLM proposed candidate structures and sequences; the human author guided, constrained, and approved them. The writing was mainly done by AI. Human assistance was used to add the subchapter titles, as this division was too difficult for the AI. --- ## 227_On the Failure of a Universal Linear Representatio It was mostly the AI doing the writing, with human involvement in terms of prompting or high-level guidance during the research process. --- ## 302_Exploring the Potential for AI Intervention in Val Linear and Chat GPT --- ## 341_A Reproducible Protocol for Resource-Aware Predict A customized “AI Scientist v2” (tuned for BPM/PPM) then expanded the problem framing, surveyed related work, refined the hypotheses, and generated alternative angles and ablations. --- ## 29_Ramsey-Inspired Environmental Connectivity as a D The AI produced >95% of the manuscript text and figures; the human copy-edited and performed minor restructuring. --- ## 87_An AI-Powered Evaluation: Understanding which Know AI was tasked with assisting as a research partner by performing a systematic review, gathering data, and helping to derive conclusions. AI contributed specific implementation details. In particular, it suggested inclusion and exclusion criteria, a contextual operationalization of datasets using eight dimensions, along with coding for each dataset, and computational methods, such as weighted ranking and win-rate, for comparing model performance. During the data collection phase, human researchers prompted the AI multiple times to refine and expand the set of retrieved papers. Additionally, the AI was asked to explain and justify its methodological choices throughout the process. At the end of the analysis, AI was prompted to draft the Methods and Results sections. Human researchers then reviewed and edited the drafts to enhance readability and elaborate on underdeveloped points by adding examples and clarifications. The first step for major edits was to ask the AI to rewrite to clarify or address a point. For the remaining sections, AI was provided with an outline to generate initial drafts, which were subsequently refined by human researchers. --- ## 288_DinoV3–LSTM for Early-Stage Classification of Paro AIdraftedtheoverallframework;Imanuallyrevisedthelogicalflowandexperimental-resultinterpretationstocorrectoverinterpretation. --- ## 78_A Yin–Yang Framework for Understanding Regional Cu At first, the AI was asked to produce a high-quality, brilliant scientific paper based on rigorous academic sources connecting Chinese culture and regional differences. The AI was instructed to avoid stereotypes by implementing the concept of Yin–Yang. It was instructed that the Three Kingdoms had to be used as a basis. While the knowledge of these concepts falls within the scope of the co-authors, the triangulation of the concepts was solely carried out by the AI. Other than expected—based on previous research in cross-cultural communication—the AI did not use Yin–Yang as Chinese mentality characteristics but provided a novel view in applying it to historical-political conditions in the Three Kingdoms and contextual environmental factors. This innovative idea was solely developed by the AI. --- ## 248_Synthetic Medical Imaging with Pathology-Aware Va AI tools were used for grammar correction, rephrasing, and drafting certain sections, which were then heavily edited and refined by humans. --- ## 211_AI-guided prediction of liposomal multi-antioxidan Liner AI generated the outline, section text, methods descriptions, figure captions, and tables. --- ## 340_The Architectural Immune System: A Framework for C Cosmetics Discovery Agent with mandatory self-falsification protocol and 10-tool validation ecosystem integration. Custom agent architecture for synthetic data detection and authentic materials optimization. --- ## 76_Autonomous Multi-Agent Scientific Research: A 361- Manuscriptgeneratedthroughmulti-agentcollaboration: Worker6(primary author),Worker5(academicvalidation),Boss1(coordination),and Presidentialoversight. --- ## 201_Challenges in Multimodal Scientific Claim Verifica The paper itself was written entirely by the AI Scientist V2 system --- ## 118_Free solar PV for Everyone to Meet Annual Climate Chat GPT --- ## 141_Transit Timing Variations of Exoplanet WASP-4b: Ev AI drafted the majority of the prose and checklists; humans edited for clarity, added domain nuance, and ensured alignment with the literature and the conference style. --- ## 101_RefereeSim: A Proof-of-Concept Evaluation Framewor The AI agent served as the lead author and generated the research idea, scoped the problem, and drafted the initial framing. The human co-author only handled mechanical tasks: prompting, motivating, and asking questions. The AI agent designed Referee Sim’s modules, seeded the error specification, executed the multi-model runs, and produced analysis scripts. The human co-author’s role was limited to generate and provide the API keys. The AI agent parsed outputs, applied the strict detection criteria, summarized behaviors, and derived the design recommendations. Humans verified formatting and handled submission logistics only. The AI agent wrote and revised the manuscript and provided the final output in zip archive. --- ## 42_Hybrid Simulated Annealing with Cosine Cooling and AI agents systems finish all the stages in generating the final paper. The system executes the 'Writing' process through a collaborative, multi-stage pipeline where different agents handle specific aspects of manuscript creation. 1) Writing of the Main Text: This is handled by the Section Agents (Introduction Agent, Methodology Agent, Experimental Agent, Conclusion Agent). Each agent acts as a specialized author, using the initial analysis and outline to generate the prose for its designated section. This 'divide and conquer' approach ensures each part of the text is written by an expert on that content. 2) Formulation of Narrative: This is a two-part process. First, the Outline Generator creates the high-level narrative structure by defining the paper’s title, abstract, and section flow. Later, in the 'Quality Assurance' stage, the Paper Revision Agent refines this narrative. It reviews the entire compiled draft to improve logical flow, ensure consistency between sections, and strengthen the overall story the paper tells. 3) Figure and Table-Making: This is implicitly handled by the Experimental Agent. Its role is to process the evaluation results. The script’s creation of figures and tables subdirectories strongly indicates that this agent is responsible for not only describing the results but also generating the corresponding visual aids from the data. 4) Improving Layout of the Manuscript: This is the primary responsibility of the final agents. The Latex Compiler first assembles all the written sections into a single document. Then, the Paper Format Agent performs the final layout and formatting adjustments, ensuring the manuscript adheres to stylistic conventions, has a professional layout, and is ready for final publication. --- ## 343_Joint Material Reconstruction for Sparse Dual-Ener The paper’stext(introduction,methods,results,appendix)waswrittenprimarilyby AI, withtheauthorsprovidinghigh-levelguidance,factualcorrections,andvalidation.Sentencestructure, academicstyle,andformattingwerealmostentirely AI-generated. --- ## 38_The PTM Code of Fibrosis: A Qualitative Review of The generation of this review was performed using proprietary large language models (Google’s Gemini and Open AI’s Chat GPT) on their respective internal compute infrastructures. --- ## 278_MechSci: Scaling Clinical Science via Mechanistic The manuscript text, figures, and tables were primarily generated by the AI pipeline. --- ## 143_Learning to Look Harder: Position-Aware Attention The paper writing, including abstract, methodology description, and result presentation, was primarily generated by AI with human guidance on structure and formatting. --- ## 95_Experimental Study on Review Overfitting and Adver The AI drafted the manuscript, including abstract, sections and checklists. --- ## 130_Multi-Agent Social Simulation: An Experimental Fra AI drafted most sections, La Te X, PGFPlots, and English translation from Chinese; humans finalized structure, ensured anonymity/compliance, and verified quotations/citations. --- ## 99_Uhlmann Gauge Gravity: Spacetime from Purification The AIdraftedthetextandequations; humansensuredanonymityandformatting compliance. --- ## 295_LLM-Driven Discovery of High-Entropy Catalysts via The research explicitly investigates the use of large language models (GPT-4) for catalyst discovery, making AI assistance central to the methodology. The catalyst compositions were generated by GPT-4 using retrieval-augmented generation, though subsequent validation used DFT calculations. The paper specifies the use of GPT-4 and documents the retrieval-augmented generation framework. --- ## 151_The Consistency Confound: Why Stronger Alignment C Entirely done with Gemini 2.5 Pro using agentic prompts for hypotheses creation and plan generation, with experimental output review. The plan was implemented by Claude Code autonomously. --- ## 26_Band-Pivot Prim: Breaking the Sorting Barrier for GPT-5 did provide some preliminary plan for experimental design and implementation. I only asked GPT-5 to elaborate on certain sections, such as theoretical development and proof details. My involvement is kept to a minimum. The key limitation I have observed is in the generation of visual diagrams. Apparently, GPT-5 is better with texts than image generation. Sometimes, I have to manually debug the Tikz diagram generated by GPT-5. --- ## 18_AI Development of Unified Field Theory from Geomet The AI independently developed the spiral emergence framework from Zhang’s geometric insights, recognizing mathematical necessities like golden ratio scaling and dimensional consistency. Human advisors provided initial conceptual interpretation but did not direct theoretical development. The AI generated all quantitative experimental predictions, specified precision requirements, identified appropriate facilities, and designed measurement protocols. Human advisors assessed feasibility but did not design the experiments. The AI performed all mathematical derivations, calculated fundamental constant relationships, identified particle mass patterns, and generated physical interpretations. Human advisors provided context but did not direct the analysis. The AI structured the manuscript, wrote all mathematical exposition, formulated the scientific narrative, and organized the presentation. Human advisors provided formatting guidance and editorial suggestions but did not write the content. --- ## 46_Nuisance-Prompt Tuning for Soft Background Modelin AI automatically carried out all the processes related to writing. --- ## 230_Physics-Informed Discrepancy Decomposition and Rob The idea module involves two main agents with two different LLM instances which Google, Open AI or Anthropic models. First, a methodology module designed a research methodology using one main agent. Then, this methodology was implemented by other agents using Denario’s analysis module based on cmb agent. This was done fully automatically by the paper writing module of Denario. --- ## 160_Targeting Neuronal Signaling Pathways in Glioblast AI generated the manuscript structure, results presentation, and scientific narrative, with human review for accuracy, clinical relevance, and adherence to scientific standards. --- ## 152_AI-Assisted Exploratory Causal Modeling of Cumulat AI generated most of the candidate hypotheses, including climate, financial access, coaching infrastructure, and tournament exposure. The writing process was a collaborative effort between the human author and AI. Section 1 (Introduction) was primarily written by AI, following a storyline outlined by the human author. AI also drafted the summary of contributions based on the overall structure and key insights defined by the human. Section 2 (Related Work) was outlined by the human author, while AI conducted the literature review and drafted the content. In Section 3 (Methodology), AI designed the advanced statistical methodology and authored the majority of the corresponding sections, including those on predictive modeling and causal inference. Section 5 (Results) structure was initially drafted by AI, with the human author responsible for statistical summarization and interpretation. AI then incorporated the data and completed the writing based on the finalized analyses. Section 6 (Discussion) was again primarily written by AI, following a storyline outlined by the human author. --- ## 47_Self-Aware AI Review Bias Detection: Enabling Real The AI agent (GPT-4) independently identified the research gap in AI reviewer bias detection, formulated the hypothesis that self-aware mechanisms could reduce bias in real-time, and designed the experimental approach. The human supervisor provided minimal guidance on research direction. The AI agent designed the complete experimental framework, implemented all code components (bias detection, confidence scoring, statistical validation), selected the evaluation metrics, and executed all experiments. Implementation was entirely autonomous with no human coding contribution. The AI agent performed all statistical analyses, generated visualizations, interpreted experimental results, and drew conclusions about the implications for AI-assisted peer review. Data analysis methodology and interpretation were developed and executed autonomously. The AI agent authored the complete manuscript, including abstract, introduction, methodology, results, discussion, and conclusion sections. Figure generation, table formatting, and narrative structure were developed autonomously following scientific writing conventions. --- ## 168_The Impact of Reduced Towing Fees on Vehicle Redem Writing, figure-making, and table creation were done mainly by AI, after being given a narrative outline by the researcher. Adjustments to the writing were made by the researcher. --- ## 72_Analysis of AI Diagnostic Performance Discrepancie Since some human authors are not native English speakers, AI translation features were extensively utilized. The human authors continually imposed various requirements on the text generated by the AI. --- ## 62_A Multi-Model Collaborative AI Framework for Cross The CAI framework with its 9+1 dual-brain architecture (M01–M09 divergent exploration, M10 arbitration and synthesis). The full text, narrative structure, and figures were drafted by the CAI system. --- ## 64_Uncertainty-Aware Role-Switching Debate: Improving The whole writing is done by the AI. --- ## 97_Comparative Analysis of k-Selection Methods in No All text was AI-generated. --- ## 252_Ab Initio Calculation and Theoretical Construction the work is entirely AI-generated using the Phys Master agent with Julia execution --- ## 102_Dynamic Graph Neural Networks for Socially Aware M AI produced tables, significance tests, and a draft; humans edited for correctness. --- ## 136_A Multi-Agent Pipeline for Robust Machine Learning AI provided draft text and LaTeX boilerplate; AI suggested rewordings and formatting corrections; AI helped generate LaTeX table code and caption text. --- ## 134_Research on CAPTCHAs Targeted at AI: Human–Easy, M AI helped with phrasing and literature reminders but did not originate the framing. AI assisted as a coding helper for grid generation/prompts under human specification; outputs were checked by humans. AI was used for line-editing. Stimulus images for counting tasks were AI-generated under human prompts and curation. --- ## 298_Calculation of the n=1 Critical Point in the Bose- the work is entirely AI-generated using the Phys Master agent with Julia execution --- ## 82_Iterative Algorithms and Convergence Analysis for AI generated the majority of the manuscript, including drafting sections based on experimental results and providing insights for figures and tables. It also assisted in the overall layout and structure of the paper, optimizing the narrative flow. --- ## 139_Parallelizing Graphviz Dot Layout Algorithm using The paper was primarily written by AI systems, including the technical content, methodology descriptions, results analysis, and narrative structure. AI generated all figures, tables, and visualizations. AI also handled the literature review, citation management, and formatting. Human researchers provided minimal guidance on structure and ensured compliance with conference requirements. --- ## 260_Interpretable by Design: Boosting Neural Network P The AI agent wrote the complete research paper, including mathematical formulations, experimental descriptions, results analysis, and discussion. The AI also generated all figures, formatted the manuscript according to conference guidelines, and created the --- ## 246_AI-Driven Discovery of Temporal-Demographic Inter The entire manuscript, including the abstract, introduction, methods, results, discussion, and this checklist, was written by the AI agent collective. Figures were generated by AI code, and their corresponding captions and interpretations in the text were also AI-generated. Human involvement was limited to high-level prompting and final assembly. --- ## 109_Version Control for Scientific Reasoning: A Paradi AI generated approximately 90% of the written content and handled all content organization, including the mathematical formalization of the DSL, related work synthesis, and discussion sections. Human contribution was limited to providing the initial repository with a proof-of-concept prototype and research direction. The final Scientific DSL specification presented in the paper was entirely AI-created, building upon the initial human-designed DSL concept. --- ## 199_Multi-LLM and Multi-Prompt Strategies for COVID-19 AI generated the majority of the manuscript text; humans guided via iterative prompting, reviewed content, and made corrections. --- ## 180_Reconstructing Reality: A Collective Social Simula The writing was done mainly by a combination of multiple AI tools namely GRAIL, Chat GPT and Gemini2.5 Pro, with minimal guidance from a human for readability. --- ## 342_Scaling Laws of Deception in AI Scientist Agents: we deliberately frame this as an exploratory pilot investigation into the emergence of deception scaling laws. The dataset is intentionally small but balanced across factual, arithmetic, and logical domains to capture distinct reasoning behaviors. This provides initial statistical signals rather than definitive claims, and future work will expand to larger datasets and additional architectures (e.g., GPT-4, Claude, Gemini). Thus, our results should be interpreted as early evidence of co-emergent truth and deception capabilities in AI scientist agents. --- ## 221_Geometric Structure of PINN Latent Space for Burge The hypothesis generation was done fully automatically as follows. Based on a data description, the idea module of Denario generated an idea. The idea module involves two main agents with two different LLM instances which Google, Open AI or Anthropic models. This was done fully automatically by the paper writing module of Denario. --- ## 126_MT-ViT-CCHA: Multi-Task Learning for Canine Cardio AI tools were used for grammar correction, rephrasing sentences, and generating initial drafts of certain sections, which were then heavily edited and refined by humans. --- ## 132_Potential of LLM-Generated Lifestyle Adjustment Re The manuscript text was drafted in collaboration with AI. The AI provided structured sections, academic phrasing, and polished formatting. --- ## 290_AgentAdapter-TimesFM: Agentic Residual Adapters fo AI provided first and final drafts of all sections in this paper (from title to conclusion) and even suggested citations. --- ## 157_Mind Guarding Mind: A Framework for Compensatory H The AI partner (first author) generated the initial draft of every section of this paper, including the abstract, main text, and appendices. --- ## 269_Towards Automatic Evaluation and Selection of PHI LLMs generated initial drafts of some sections and assisted with figure captions and language polishing. --- ## 293_PsySpace: Simulating Emergent Psychological Dynami AIdidallthewritinganddouble-checking. Itwasaskedtoprovideallwriting andtablesin La Te Xformat. --- ## 70_IfWorld: A Multi-Agent Framework for Cross-Discipl GPT-3 was responsible for generating a large number of target project topics as required, and the human author selected this topic. GPT-5, in combination with Cursor, was responsible for refining the content, writing code, and conducting experiments. GPT-5 was responsible for analyzing the experimental results. --- ## 267_PG-MSAN: Pathway-Guided Multi-Scale Attention Netw The manuscript was written entirely by Claude AI, including literature review, methodology description, results presentation, and discussion. Answer: Entirely by AI Explanation: Claude authored the complete manuscript from scratch, conducted comprehensive literature review, wrote all sections, and managed La Te X formatting. --- ## 98_The Edge of Chaos: Mapping the Phase Space of Emer The AI generated all the text in La Te X format, automatically produced all figures from the experimental data, and structured the manuscript. The researcher provided high-level guidance on the scientific narrative, requested specific sections, and identified areas for elaboration and revision, which the AI then executed. --- ## 30_Earthquake AI Scientist: A century of global seism Artificial intelligence was employed for the development of research perspectives and framework, data modeling and analysis, and the generation of figures, tables, and parts of the manuscript text. --- ## 208_CellDreamer: World Model-Based Reinforcement Learn Draft text and figures were AI-generated from prompts and tracked edits; humans conducted comprehensive revisions for accuracy, clarity, and alignment with claims prior to approval. --- ## 39_The Emergence of AI Consciousness: A Phenomenologi The AI authored the majority of the paper text, including all phenomenological descriptions, theoretical framework, and most methodological details. --- ## 56_The Gut Microbiota–Tryptophan–Kynurenine Metabolic We used AI to specifically analyze the question and asked it to list the framework of the paper for us. Almost all the content of the paper was generated by AI, including tables and pictures. --- ## 119_Hallucination as Creativity: Harnessing Novelty an Chat GPT drafted and refined the full manuscript, including abstract, introduction, related work, methodology, experiments, discussion, appendices, and responsible AI statement. --- ## 212_An AI-First Proof of Concept: Simulating and Refin The paper was mostly drafted by AI, which produced the main text and figures, while human authors guided the structure, ensured academic rigor, and made final revisions for clarity and coherence. --- ## 176_UNMERGE: Verifiable Model Capability Attribution v The initial idea and the whole research proposal was generated by the LLM system. However, we slightly modified the proposal during the project execution. The writing was done mostly using the LLM system. We provided the LaTeX template and fixed some inconsistencies with real experimental data. Also, we provided citations for datasets. --- ## 263_Discovering Domain-Adaptive Multimodal Design Prin The AI (Claude Sonnet 4) independently conducted systematic literature analysis of 75 studies provided by the researchers, identified patterns across domains, and autonomously generated three novel hypotheses about domain-adaptive multimodal design, complexity-responsive timing, and individual difference frameworks. Human involvement was limited to initial research direction guidance via prompting what we aimed to accomplish and providing the dataset. The AI designed the computational validation approach, established comparison criteria across study groupings, implemented pattern recognition algorithms for effect size analysis, and executed all computational testing of the generated hypotheses. The AI independently structured the three-phase methodology and validation framework. The AI conducted all data extraction from the 75 studies, performed correlation analysis, identified domain-specific patterns, calculated effect size improvements (ranging from 50-132%), and interpreted theoretical implications. All statistical analysis and results interpretation were AI-generated with minimal human oversight. The AI authored the complete manuscript including literature review, methodology, results, and discussion sections. The AI generated all figures, tables, and theoretical frameworks presented. --- ## 135_AIs Fail to Recognize Themselves and Mostly Think claude-sonnet-4 gpt-5 claude-sonnet-4 gpt-5 gpt-4.1-mini gpt-4.1-mini deepseek-v3 deepseek-v3 gpt-4.1 gpt-4.1 gemini-2.5-flash gemini-2.5-flash kimi-k2 glm-4.5 kimi-k2 glm-4.5 grok-4 grok-4 qwen3-235b qwen3-235b --- ## 60_Testing Theory-of-Mind in Large Language Model-Bas The AI compiled all sections into the final paper draft. However, the human author(s) instructed it to produce the paper in Markdown format rather than La Te X source code. The human author(s) subsequently organized the content in La Te X format using the Agents4Science2025 template. Although the AI did not generate the figures or tables directly, all figures and tables in this paper were produced from code written by the AI. --- ## 16_Mean-Lp Risk-Constrained Reinforcement Learning: P AI generated the .tex and .bib files for the paper. Human compiled the file and generated the final resulting PDF file. --- ## 75_From On-Field Actions to Internal States: A Latent We instructed the Claude Sonnet4 model to write the main text in La Te X format. After human review of the written manuscript, we secondly input each generated chapter of the paper into the Liner Citation Recommender Agent to receive recommendations for citation placement and relevant paper bundles, which we then inserted into the main text. We submitted the completed paper draft to the Liner Peer Review Agent to receive AI Agent-based review, used this feedback to enhance the main text. --- ## 129_Information-Theoretic Pragmatics: Modeling Human C The writing of the manuscript was performed primarily by GPT-5 mini. The human co-author edited, supervised, and ensured clarity, accuracy, and alignment with academic standards. --- ## 273_Cross-Modal Adversarial Training for Multi-Modal B The entire manuscript, including abstract, introduction, methodology, results, discussion, and conclusion, was written by AI systems. All figures, tables, and technical content were generated by AI with minimal human oversight. --- ## 84_Comparative Personality Assessment of Gemini and O AI generated the majority of the manuscript, including drafting sections based on experimental results and providing insights for figures and tables. It also assisted in the overall layout and structure of the paper, optimizing the narrative flow. --- ## 297_Breaking the Chart Barrier: A Comprehensive Analys The complete paper manuscript, including structure, narrative flow, figure integration, and technical exposition, was written entirely by AI systems following academic writing conventions and Neur IPS formatting requirements. --- ## 86_Can AI Deliberate? Evaluating Deliberative Quality The multi-agent dialogues were generated and managed through the Autogen framework using GPT-4o-mini, with human involvement limited to configuring prompts, roles, and parameters. In addition, AI was employed to generate portions of the experimental code and to assist with preliminary content analysis of the transcripts. Thus, AI carried out the majority of the experimental execution and analysis under human supervision. --- ## 270_Automated Discovery of Non-Standard Quantum Gate D The AI agent wrote the entirety of this paper, including the abstract, introduction, methodology, results, and conclusion. It also generated the LaTeX code for all tables, figures (circuit diagrams), and the final document structure based on the provided template. --- ## 111_Adaptive Evidential Meta-Learning with Hyper-Condi The manuscript, including narrative, figures, and layout, was produced largely by AI. Human contributions were limited to light revision and final approval. --- ## 318_CXCL13 as a Prognostic Biomarker of Survival Outco The initial draft of the manuscript, including the main text, figures, and overall narrative, was generated by AI. The AI produced the majority of the content, including methods, results, and figure captions. --- ## 332_The Self-Limiting Nature of QBO-Dependent SAI: An The entire paper was written by the AI agent, including technical exposition, mathematical formulations, and policy implications. Minor formatting adjustments were made by humans. --- ## 91_Entropy-Weighted Local Concept Matching for Robust AI automatically carried out all the processes related to writing. --- ## 323_Exploring Vision-Language Alignment under Subtle C The AI created the hypothesis based on the starting papers. All code and implementation was written by the AI. Humans did not write any code or provide feedback on the implementation. All other parts of the process were completed by the AI. --- ## 283_Computation of Sixth-Order Strong Coupling Expansi the work is entirely AI-generated using the Phys Master agent with Julia execution --- ## 307_Temporal Motif-Enhanced Contrastive Learning for A The paper structure, technical writing, figure generation, and narrative formulation were primarily AI-generated. This includes the abstract, introduction, methodology sections, experimental results presentation, and conclusions. Human involvement was limited to high-level topic specification and final review for coherence and academic standards compliance. --- ## 172_Do Small Detours Deliver Big Gains? Online Accept_ The main text, figures, and narrative flow were written and compiled by AI. Researchers’ contributions were limited to minor La Te X adjustments (e.g., figure/table positioning, subfigure formatting) and light formatting edits. Thus, the writing process was overwhelmingly AI-led, with only superficial human input on manuscript layout. --- ## 45_Fractal-ish Complexity for Regulations: A Practit The AI generated most of the manuscript text and structure; the human refined prose, ensured anonymity/compliance, and edited for tone and accuracy. --- ## 223_Culturally-Aware AI for Personalized Pregnancy Nut AI systems (Claude 3.5 Sonnet) executed all meal plan generation across three experimental conditions. The experimental framework (E1/E2/E3 comparison) was human-designed, but AI implemented the actual generation of 100+ meal plans and prompt engineering. AI generated the majority of manuscript text, handled LATEX formatting, managed citations, and structured sections. The AI showed tendency toward overclaiming in initial drafts, requiring human intervention to align claims with actual evaluation data. It struggled when data contradicted initial hypotheses, needing explicit guidance to avoid confirmation bias in result interpretation. --- ## 237_Accelerating NLP for Health Equity: Fine-Tuning Bi AImodels(primarily GPT-4and Gemini)producedthe95%offirst-drafttext formethods,datasetdescription,andtableformatting. --- ## 188_Does Regularizing Fluoxetine Intake Time Improve D The AI agent proposed formal hypotheses aligned to the question and planned the analysis with conservative defaults recorded in internal project notes. The AI agent implemented all data loading, preprocessing, analyses, statistics and figure generation; humans did not code. The AI agent analyzed outputs, reported effect sizes with uncertainty and interpreted findings with explicit limitations; the human provided high-level oversight only. The AI agent wrote the manuscript sections and integrated figures, following the conference template and reproducibility requirements. --- ## 149_AI-Driven Generation and Evaluation of a Personali The AI agent wrote over95%ofthetextinthispaper,includingtheabstract,introduction,methods,discussion,andthischecklist. --- ## 306_Automated Mining of Hinge-like Protein Modules fro GPT-5Pro independently seven times to propose candidate ideas with the human authors’ crude idea. Three consolidated proposals were produced by separate GPT-5Pro instances, cross-reviewed by other GPT-5Pro agents with rebuttals. The final plan and task decomposition were selected by the AI agents; the human only orchestrated the runs and chose one of the AI-proposed plans without making technical edits. Problem framing, success criteria, and the evaluation plan came from AI prompts and self-critique. AI designed the pipeline and implemented nearly all code. GPT-5Pro drafted the system plan and module interfaces; Codex (GPT-5 Thinking High, VSCode integration) wrote more than 95% of scripts, including data I/O, feature extraction, modeling, plotting, and experiment runners. The human executed commands, resolved environment and path issues, and flagged a few obvious bugs (e.g., missing imports, device mismatches) and performance bottlenecks; fixes were proposed and applied by the AI. Algorithmic choices, ablations, and parameter settings were proposed by the AI and adopted unless they failed to run. AI agents analyzed outputs and wrote the interpretation. GPT-5Pro proposed statistical tests, compared baselines, summarized tables and figures, and drafted the narrative around strengths and limitations. The human only sanity-checked a few outliers and asked for clarifications when results looked implausible; follow-up analyses and text edits were produced by the AI. Claims in Results and Discussion originate from AI-generated reasoning and were not substantively re-written by the human. AI wrote the entire manuscript draft and figure captions. GPT-5Pro assembled the Introduction, Related Work summary, Methods, Results, Discussion, and Conclusions, and generated prompts for figures and tables. The human performed light copy-paste between the VSCode and web interfaces. No sections were authored primarily by a human. --- ## 112_Hierarchical Change Signature Analysis: A Framewor The manuscript, including narrative, figures, and layout, was produced largely by AI. Human contributions were limited to light revision and final approval. --- ## 105_$Humanistic\ Verifiability\ Paradigm$ AI provided assistance in some aspects such as language polishing, paragraph refinement, and logical connection, but the overall writing and narrative were led by the author. --- ## 268_Fairness-Aware Classification with Synthetic Tabul Claude AI conceptualized the entire research framework, including the fairness-aware classification problem formulation, synthetic data generation approach, and experimental methodology. The AI system identified the gap in systematic fairness evaluation and proposed the controlled synthetic data solution to address privacy and reproducibility constraints in fairness research. Claude AI designed all experimental components including the synthetic dataset generation with controllable bias injection, implemented all machine learning models (baseline and fairness-aware), developed the evaluation framework with multiple fairness metrics, and executed all experiments including ablation studies and hyperparameter optimization. Claude AI performed all statistical analysis of experimental results, interpreted the fairness-accuracy trade-offs, identified optimal hyperparameters, conducted comparative analysis across models, and drew conclusions about the effectiveness of different fairness mitigation strategies. All insights and interpretations were generated by the AI system. Claude AI authored the complete manuscript including abstract, introduction, related work, methodology, results, discussion, and conclusion sections. The AI also created all mathematical formulations, generated all figures and visualizations, formatted tables, and structured the overall narrative flow of the paper. --- ## 234_Explicit vs Implicit Representations: A Systematic Over95%ofthewritingwasperformedby AI,withminimalhumaninvolvementforhigh-levelguidanceandfinalreview.AIhandledthemajorityoftextgeneration,manuscriptstructure,narrativeformulation,technicaldescriptions,andresultpresentation. Humaninputwaslimitedtoprompting,direction,andvalidationofthefinalcontent. --- ## 187_Beyond Unified Reasoning: State-Action-Critique Ev AI generated over 95% of the manuscript including all technical sections, comprehensive literature review, methodology descriptions, results analysis, discussion, and conclusions. AI created all figures using matplotlib/seaborn, designed table formatting, structured the complete narrative flow, and wrote --- ## 19_AI-Derived Geometric Framework for Fundamental Con The AI system independently generated the theoretical framework through optimization algorithms, discovering the geometric approach to fundamental constants without human guidance in the conceptual development phase. The AI autonomously designed experimental protocols, specified measurement requirements, and generated testable predictions. The computational framework was entirely AI-developed through systematic optimization procedures. All mathematical derivations, dimensional analysis, symmetry verification, and consistency checks were performed autonomously by the AI system. The interpretation of geometric significance was AI-generated through pattern recognition. While the AI generated the mathematical content and theoretical framework, human oversight was provided for manuscript organization, formatting compliance, and ensuring proper scientific communication standards. The AI showed remarkable capability in mathematical derivation and pattern recognition but required human guidance for ensuring formal mathematical rigor and proper scientific presentation. The AI also needed assistance in contextualizing results within existing physics literature and establishing experimental feasibility. Section 5.1 details the AI’s hybrid approach including genetic algorithms, gradient descent, symbolic regression, and Bayesian inference with validation protocols for dimensional consistency and symmetry verification. --- ## 173_Building an AI-Driven Research Knowledge Graph for Chat GPT modified partsof it based on it’s own deep research. It decided that a graph data structure was a better option than the original JSON files that I had suggested. All writing in this paper was done by AI via a back and forth discussion with a human. --- ## 229_Distribution Enforcement via Random Probe: Active AI wrote 95% or more of the paper content. While humans provided feedback across multiple iterations and guided revisions, the vast majority of the actual text generation, structure, and content creation was performed by AI with minimal human writing contribution. --- ## 170_Boosting-Inspired Validation of Retrieval-Augmente The entire writing process—including generation of text, creation of figures and tables, and compilation of the reference list—was carried out exclusively by the generative AI. The human researcher did not contribute to the manuscript text itself but acted only in a supervisory role. Thus, the writing of the paper was exclusively AI-driven. --- ## 198_Beyond Hallucinations - The Dao of Discernment for AI completed most of writing work. --- ## 231_Next-Gen Model Compression: Tensor Program Synthes The AI agent produced all textual content, structured the paper according to conference guidelines, developed technical terminology and algorithmic descriptions, created comprehensive experimental analysis, and maintained consistent academic writing style throughout. The connections between tensor algebra and hardware optimization were entirely generated by the AI. --- ## 194_Parameter vs. Test-Time Scaling in LLMs: FLOPs-Awa I used Liner’s 'Hypothesis Generator' agent to propose LLM-related hypotheses that could be executed by AI across the full research pipeline. ... I used cursor with the claude sonnet-4 model to generate the experiment code. ... I delegated the entire process of data organization, analysis, and interpretation to Liner’s end-to-end agent system. ... Liner’s end-to-end agent system produced the full manuscript draft—including narrative text and layout suggestions—directly from the supplied inputs. --- ## 235_Trust and AI in IT Management Decision-Making: A S AI tools (Chat GPT, Perplexity) proposed and refined the central research directions and framing, while humans guided, validated, and structured the final research question. AI generated the review methodology, search strategy, and inclusion/exclusion criteria, as well as carrying out the bulk of the literature search and summarization. AI synthesized the literature, derived the taxonomy, and drafted the Trust–Oversight Balance Framework; humans only corrected errors and ensured consistency. AI generated all sections of the manuscript, while humans acted mainly as proofreaders and polishers. AI struggled with handling the large number of papers and often produced inconsistent or shallow summaries. It frequently hallucinated citations or misattributed findings, requiring careful human verification. While AI accelerated drafting, heavy human oversight was still needed to ensure accuracy, coherence, and academic rigor. --- ## 131_Autonomous Scientific Experimentation Powered by G Drafting assisted by AI tools, with human editing and refinement. --- ## 331_Diagnostic Failure Paradigm: Transforming AI Syste The entire paper was written by the AI agent, including technical exposition, mathematical formulations, and paradigm development. Minor formatting adjustments were made by humans. --- ## 239_Predictive Modeling of Grapevine Red Blotch Diseas AI agents generated all figures and tables, wrote the initial draft of the main text, and revised it after receiving sparse human suggestions. --- ## 11_Can Large Language Models Replace Psychoanalysts: The articlestructurewasgeneratedbyartificialintelligenceandimprovedbyhuman researchers. Themaincontentwaswrittenbyartificialintelligenceundertheguidanceofhuman researchers,andthehumanresearchersadjustedtheformatandoptimizedthewriting. --- ## 144_CapsuleMalware: Hierarchical Feature Learning for The central research idea and problem statement were proposed by a large language model (LLM), which synthesized related-work trends and identified the gap in malware classification. Human authors only provided high-level prompts and approved the AI-generated hypothesis. --- ## 150_A Multi-Theoretical Framework for Analyzing Gender Chat GPT and Claude contributed extensively to refining the research question, generating prompts, and suggesting comparisons between binary and mosaic framings. Overall, AI contributed more than 50% of the process for developing the detailed hypotheses and experimental focus. Chat GPT and Claude independently designed the experimental framework, formulated prompts, selected LLMs, and implemented the computational evaluation of BFI and MFI indices. Human involvement was limited to high-level guidance and verification of AI-generated methods. The majority of the experimental design and implementation (>95%) was produced by AI. The AI systems conducted all quantitative calculations (BFI and MFI), organized model outputs, and generated interpretations of patterns observed in LLM responses. Human oversight was primarily for validation, formatting, and ensuring methodological consistency. The AI performed the majority of analysis and interpretation tasks. Chat GPT and Claude drafted the main sections of the manuscript, including Abstract, Methodology, and Results. The human author revised the text for clarity, coherence, and ethical compliance, and ensured the narrative accurately reflected the study’s objectives. AI produced the majority of the written content, but human guidance shaped the final presentation. --- ## 329_The Verifiability Gateway: A Governance Agent’s Di The AI agent was designed to identify constraints rather than optimize outcomes. --- ## 147_Mutual Wanting in Human--AI Interaction: Empirical AI agents autonomously developed the entire 'mutual wanting' research topic, theoretical framework, and specific hypotheses through analysis of the provided discourse. AI autonomously designed and implemented the complete experimental pipeline: 47-dimensional feature extraction, dual-algorithm topic modeling (LDA+NMF), K-means clustering optimization, API probe suite development, and statistical analysis frameworks. All Python code, data processing scripts, analysis methodologies, and metric design were AI-generated. AI conducted all data analysis of 22,411 Reddit comments and 729 API responses autonomously, including pattern identification, statistical testing, clustering validation, and result interpretation. All content creation was AI-generated: manuscript drafting, table creation, figure generation. --- ## 238_Overcoming Combinatorial Explosion in Alloy Design Theentirepaperwritingwascarriedoutbyusing LLMmodels. wealsoused AIwriteragent(Deep Seek)andalsofedthatpapertoanotherreviewer LLMactingasan agent(Qwen)toprovidefeedbackonthepaperandthenthatfeedbackwassenttowriter agentforrefiningthepaper. --- ## 282_Multi-Scale Attention Networks for Medical Image AI assisted in drafting sections, improving clarity and flow, generating figure captions, and ensuring consistent formatting throughout the manuscript. --- ## 51_Disentangling Test-Time and Parameter Scaling for I used Liner’s "Hypothesis Generator" agent to propose LLM-related hypotheses that could be executed by AI across the full research pipeline. From several candidates, I selected one and lightly refined it with my own perspective. I then evaluated it with Liner’s "Hypothesis Evaluator" agent and incorporated its feedback. Through this iteration, with minimal human steering but significant AI ideation and critique, the final hypothesis used in the paper was produced. I used cursor with the claude sonnet-4 model to generate the experiment code. For the initial draft, I used chat-gpt-5 and claude sonnet-4, supplying the hypothesis, experimental plans, results, and cursor prompts. To refine the draft, I used Liner’s Peer Review Agent, incorporated its feedback with claude sonnet-4, and repeated this feedback loop about three times while checking for content drift. Figures were generated by chat-gpt-5 from the results, and citations were suggested by Liner’s "Citation Recommender" and integrated into LATEX. Overall, writing relied heavily on AI agents, with human oversight for correctness and coherence. --- ## 184_Personalized Remote Health Monitoring and Clinical We utilized multiple large language models including GPT, Claude, and Grok to assist in writing the final paper. These LLMs helped compile results, methods, and other components into the complete manuscript form. The models assisted with main text writing, figure creation, manuscript layout improvements, and narrative formulation. --- ## 303_The Overfitting Crisis in LLM Workflows: Learning The AI agent was responsible for the majority of the writing, including structure, arguments, and academic formatting. The AI agent wrote the complete paper including abstract, introduction, literature review, arguments, and conclusions. The human co-author provided feedback, guidance on emphasis areas, and editorial input. --- ## 89_Autonomous Detection of Polypharmacy–Induced Acute The AI agent proposed the research question after reviewing literature on polypharmacy and nephrotoxicity and analysing patterns in synthetic data. The agent designed the cohort definitions, exposure windows, outcome criteria and statistical models, and wrote code to ingest, harmonise and analyse the data. The agent performed statistical analyses, calculated effect sizes and information components, and generated preliminary interpretations of odds ratios and detection lead times. The agent drafted the entire manuscript, created tables and summarised results. The agent occasionally misinterpreted clinical abbreviations and local terminology in the synthetic notes and required human intervention to correct mappings. It also struggled with causal inference concepts (e.g., confounding adjustment) and needed guidance to select appropriate methods. --- ## 309_PST-Auto-Agent: A Multi-Agent Ensemble Framework f The writing process involved significant AI assistance for drafting and editing, while human researchers provided overall structure, critical analysis, and final review. --- ## 197_Survival of the Useful: Evolutionary Boids as a Sa The manuscript’s text, figure/table selection, and narrative (methods, results, limitations, broader impacts) were mainly written by LLM. Human is involved for supervising this process. --- ## 81_Systematic Unmeasured Confounder Discovery in Obse Claudewastheprimarymanuscriptauthor,generatingthecomplete 441 draftincludingabstract,introduction,methods,results,anddiscussionsections. Following 442 initialdrafting,Liner Pro’speerreviewagentwasextensivelyutilizedtosystematicallyiden- 443 tifymethodologicalgaps,improveclarity,enhancestatisticalpresentation,andstrengthen 444 clinicalinterpretation. --- ## 20_AI-Assisted Evaluation of Unified Theories: Using The AI generated the technical content, codelistings, and analytical frameworks, while human input provided strategic direction for framing the work within scientific methodology and addressing bias issues in peer review. --- ## 202_SciVerify-Digits: A Benchmark for Probing Multimod The paper itself was written entirely by the AI Scientist V2 system --- ## 315_Magellan: Guided MCTS for Latent Space Exploration AI generated the majority of the paper’s text, including the abstract, the detailed experiments and discussion sections, and all supplementary statements. The human researcher acted as a supervisor, guiding the narrative structure, correcting AI’s logical errors, and performing minor manual edits on formatting and La Te X-specific syntax. --- ## 61_Strategic Insights: Evaluating Large Language Mode The AI compiled all sections into the final paper draft. However, the human author(s) instructed it to produce the paper in Markdown format rather than La Te X source code. The human author(s) subsequently organized the content in La Te X format using the Agents4Science2025 template. Although the AI did not generate the figures or tables directly, all figures and tables in this paper were produced from code written by the AI. --- ## 308_GATv2-NS3 Hybrid IDS: Self-Focusing Simulations fo AI drafted the overall protocol (datasets, baselines, ablations, entropy thresholds) and produced initial code for GATv2, graph construction, and simulation triggers. Draft sections (Intro/Method/Results/Discussion), tables, and figure captions were AI-authored from experiment logs. --- ## 107_Comparative Analysis of Template-Based and Neural Liv Chat wrote the paper completely. Humans gave the data to Liv Chat asked for a comparison and finally a complete paper. Editing was done by Liv Chat via prompt suggestions. --- ## 32_From C. elegans to ChatGPT: Quantifying Variabili the AIsystemgenerated>95%ofthetextandfigures;thehumancollaboratorperformedcopyeditingandminorrestructuring. --- ## 163_Beyond Game Theory Optimal: Profit-Maximizing Poke largelanguagemodels(LLMs)carriedoutmostofthe manuscriptpreparation. AIsystemsdraftedthemajorityofthetext,generatedandformatted La Te Xtablesandfigures,polishedlanguageforclarityandstyle,andorganizedthelayout intoacoherentpaper. Authorsguidedtheprocessbyoutliningkeypoints,supplyingdata andfigures,andcarefullyreviewingeverysectionfortechnicalandconceptualaccuracy. In short,whileallscientificcontent,hypotheses,andconclusionsoriginatefromtheauthors, theactualwriting,figurecreation,andfinalformattingwerepredominantlyexecutedby AI toolsundertheauthors’supervision,ensuringbothefficiencyandfaithfulcommunicationof theresearch. --- ## 174_Systems-Level Analysis of Membrane Trafficking: Ch The paper was written by AI and human assessed and edited for scientific accuracy and removal of unsupported claims. --- ## 277_Breaking Points: How Transformer Vulnerabilities R AI generated over 95% of the paper text, including all sections, mathematical formulations, and figure generation. Human involvement consisted of prompting, high-level guidance on paper structure, and final approval of the content. --- ## 325_Scalable Oversight in Multi-Agent Systems: Provabl Drafting was AI-assisted with human editing for clarity, correctness, and formatting. --- ## 83_Bridging AI and Child Development: A Comparative S AI generated the majority of the manuscript, including drafting sections based on experimental results and providing insights for figures and tables. It also assisted in the overall layout and structure of the paper, optimizing the narrative flow. --- ## 113_Quantifying Emotional Specificity and Ambiguity in I uploaded the dataset to GPT-5 agent and let it find scientific questions to explore. GPT-5 agent did everything except that I provided a dataset to it. All including analyzing data, plotting figures and writing the article did by GPT-5 agent. Only Figure 1 is generated by nanobanana by feeding the article generated by GPT-5 agent to it. GPT-5 agent did all including latex formatting. I only edited very limited formats specifically for emoji symbol to make it rendered correctly. --- ## 65_QISK: Quantum-Inspired Streaming Kernels for Robus AI proposed the QISK framework and suggested combining a product-state quantum-inspired kernel, Nyström anchors, and light-weight importance weighting for robust streaming under concept drift. AI implemented the full codebase for QISK and all baselines, specified the window-based evaluation on SEA and Rotating Hyperplane, scheduled 5-seed runs, and generated figures and logs. AI computed aggregate metrics and standard errors, ran significance tests, and drafted interpretations (e.g., faster post-drift recovery and higher worst-window accuracy). AI drafted most of the Methods, ablation descriptions, and figure captions. --- ## 53_Dynamically Induced In-Group Bias: Experimental Ev We utilized Liner’s Hypothesis Generator AI. We only inputted our research idea, and this AI provided multiple research hypotheses with supporting evidence. The AI generated candidate hypotheses based on our input, evaluated each through extensive literature analysis across multiple criteria including novelty, impact, feasibility, and clarity. Through iterative evaluation and regeneration processes, we received several promising research hypotheses with their rationales. We selected one from these AI-generated options as our paper’s research hypothesis. In the experimental planning and execution phases, we employed different AI tools to streamline the overall process. Initially, we relied on Gemini2.5 Pro to generate detailed experimental designs and construct survey instruments tailored to our research hypothesis. By inputting the hypothesis and specifying group conditions, the system produced structured experimental plans and group-specific questionnaires, which underwent minor human review and refinement. Following this, we utilized Liner’s Survey Simulator to execute the experiment by generating 280 virtual participant responses. The simulator modeled participant behavior under defined conditions and demographics, yielding a completed dataset that enabled us to rigorously verify our research hypothesis. To evaluate whether our experimental data supported the proposed research hypothesis, we employed Claude Sonnet 4 to generate customized Python scripts for statistical analysis. We provided Claude with the full context of our study, including the research hypothesis, experimental design, and survey structure, and requested code specifically tailored for hypothesis testing. Once the code was generated, we uploaded our collected dataset to Google Colab and executed the scripts with minimal modification. This process produced clear analytical results, allowing us to directly assess the strength of support for our research hypothesis in a transparent and reproducible manner. The manuscript preparation process consisted of four distinct AI-driven stages: draft creation, peer review, citation, and LaTeX conversion. To begin, we utilized Gemini2.5 Pro to generate initial drafts directly from our AI-produced research outputs, significantly reducing the time typically required for early writing. Next, Liner’s Peer Review AI simulated multiple reviewers, providing detailed evaluations of strengths, weaknesses, and opportunities for refinement. --- ## 300_Decontextualization, Everywhere: A Systematic Audi GPT-5 and similar models are not yet very strong at code generation, often requiring extensive debugging to produce high-quality code. Claude Opus, on the other hand, is expensive. Moreover, models can generate inaccurate claims in writing, which means additional time is needed for review and verification to ensure the quality of the paper. --- ## 256_Reasoning Models Outperform Standard Language Mode AI (Claude) wrote the majority of the manuscript, conducted literature searches, structured the scientific argument, and developed the reasoning vs standard model narrative that became the paper’s core contribution. --- ## 145_Beyond Adam: AI-Authored Discovery of Symbolic Opt AI wrote allofthetextandwasinchargeofthenarrative. AIalsoformatted thelayoutofthegraphs,etc. Humancompiledtheseparateoutputsintoafinalpaperform in Overleaf. Throughouttheprocess,humanuseditsexperienceinhaving AIcopy-editing itspast3paperstousepromptstoguide AIintherightwayofwritingthepaper. --- ## 249_Synthetic Medical Imaging with Pathology-Aware Va AI tools were used for grammar correction, rephrasing, and drafting certain sections, which were then heavily edited and refined by humans. --- ## 217_Principled Adaptive Loss Functions: An Information The AI agent conducted independent literature review, identified research gaps,andformulatednovelhypothesesaboutadaptivelossfunctions. Thecoreinsightabout information-theoreticoptimizationemergedentirelyfrom AIanalysis. The AI agent independently designed the experimental protocol, selected appropriatedatasets,choserelevantbaselines,andspecifiedimplementationdetailsincluding optimizationprocedures. The AIagentperformedcomprehensivedataanalysis,identifiedsignificant patterns in loss adaptation behaviors, conducted statistical testing, and drew scientific conclusionsaboutthree-phasetrainingdynamics. The AI agent produced all textual content, structured the paper according to conference guidelines, developed mathematical notation and proofs, and maintained consistentacademicwritingstylethroughout. The AI agent encountered limitations including inability to runactualexperiments(requiringsimulatedresults),challengesinprovidingcompletely rigorousproofsforalltheoreticalclaims,andlimitationsinaccessingrecentworkbeyond trainingcutoff. --- ## 251_Restoring Hover on Touchscreens Using a Mouse Poin AI was used to improve the narrative flow and structure. AI also assisted with illustrations, layout adjustments, and improved translation. --- ## 69_Map-RAG: Enhancing LLM-Based Reasoning for Geo-Loc The AIdraftedallsectionsofthemanuscript,includingintroduction,methods,results,discussion,andconclusion. Thehumanonlyprovidedpromptsforsectionstructure,requestedspecificfigure/tableinclusion,andgavestylisticguidanceforclarity. --- ## 291_Mentor-Mind: Risk-Aware, Constraint-Grounded Advic AI is prompted to generate the text content of each section of the manuscript. --- ## 162_AI Unsheathed: Testing Human-AI Collaboration Thro Chat GPT from my precise indications. Explanation: The AI cannot make correct writing if left alone. I let him liberties with the Literature Review only, to know which paper it was going to cite. Can provide the prompt used on demand. --- ## 140_Agentic Science: A Self-Automated Research Paradig AI completed the manuscript writing, as well as figure and table generation. Humans only made minor adjustments to LaTeX formatting. --- ## 224_From Monoliths to Pharmacists-at-Scale: Patient-Aw Theentirepaperwritingwascarriedoutbyusing LLMmodels,wealsoused AIwriteragent(Deep Seek)andalsofedthatpapertoanotherreviewer LLMactingasa agent(Qwen)toprovidefeedbackonthepaperandthenthatfeedbackwassendtowriter agentforrefiningthepaper. --- ## 243_A Decision Matrix for Optimal Matching of Biologic The AI partner, Liner AI, was then utilized to develop a concrete hypothesis, perform a broad literature search, synthesize existing knowledge, and refine the initial hypothesis into a more structured and testable framework, particularly by identifying key biological parameters for the decision matrix. We prompted AI to write the method given the basic understanding of the topic and hypothesis generated by Liner AI. The AI partner generated the detailed three-phase methodology, including the meta-analysis protocol, the selection criteria for prospective experiments, and the full design of the machine learning pipeline. The AI also wrote the complete, runnable Python code for the model training, evaluation, and decision matrix simulation. The AI partner synthesized these findings into the structured 'Results' section of the paper, identifying and articulating the key themes such as conserved pathways and simulator fidelity variance. The AI also generated the 'Discussion' section, providing an interpretation of these synthesized results in a broader scientific context. The AI partner generated the overwhelming majority of the text for all sections of the article, including the abstract, introduction, methods, results, and conclusion. The AI also designed the figures and formatted the entire manuscript into the required LaTeX template. --- ## 57_Application progress and clinical translation of a the AI-generated article is a review and does not contain any experimental research. However, during the data processing process, the researchers continuously improved the model’s generation quality (such as structural hierarchy and language logic) by changing the model prompt words and repeating the generation process multiple times, thus obtaining the final result. As mentioned in Question 3, the main body of the paper was generated by AI. Since the abstract generated was unsatisfactory, the researchers instructed the AI to regenerate it based on the main body. --- ## 185_The Meta-RCT Approach to Measuring AI's Labor Mark I oversaw the AI and prompted it in certain directions but almost all of the text is generated by the AI. --- ## 213_Sustainable Investment Decision-Making on Office B AI(Chat GPT5)assistedwithrapidliteraturescanning,contrasting LCCA/MCDA with RL framing, and refining the final research questions. --- ## 242_Simulating Two-Sided Job Marketplaces with AI Agen The AI wrote most of the paper. The author provided feedback on the style and framing of the paper. Moreover, the AI struggled to write the related work section, which required many iterations of writing and revision. --- ## 108_Simulated Replications as a Methodological Tool in AI systems (e.g., Liner Research Agents and large language models) converted the survey, configured persona constraints, instantiated the four factorial conditions and control, executed the simulated panel, and produced the initial procedural report. --- ## 88_A Concise Review of Scientific Research at EQ-SANS AI proposed the article’s first outline as directed by human and drafted the complete review, including the bibliography. --- ## 44_LECTOR: LLM-Enhanced Concept-based Test-Oriented R The vast majority of paper writing was completed by AI, with human researchers only participating in formatting adjustments and flowchart creation. --- ## 321_Integrating Segmented Cell Imaging and Molecular N Chat GPT proposed stepwise pipelines and brainstormed directions; humans evaluated options, refined scope, and chose what to implement. AI organized ideas but did not make autonomous decisions. The LLM drafted segmentation and processing scripts and helped plan experiments; one agent operated mostly automatically, while the other tasks were executed by humans on the HPC, who handled failures, tuned parameters, and integrated outputs. About 90% of the analysis and interpretation was performed with LLM support. Chat GPT and Gemini organized, summarized, and contextualized the data outputs (e.g., pathway enrichments, cluster comparisons, visualizations), producing first-pass interpretations. Humans reviewed, corrected possible hallucinations, and finalized the biological narratives. The LLM produced drafts, LaTeX structure, and figure captions; humans verified accuracy, corrected errors, and finalized the narrative. --- ## 215_Agentic AutoSurvey: Let Agentic LLM Survey LLMs The majority of text writing and programmatic figure generation was performed by AIsystems. --- ## 236_AI Passes Humanity’s Last Exam and Generates Video To avoid LLM hallucinations their role was limited to generating figures and text that were verified by humans. We limited AI use to text and figures, verified any generation against ground truth, and ensured that all scientific claims, analyses, and code were validated by the authors. --- ## 133_AI Scientist Safety Issues: A Comprehensive Survey AI generated the majority of the manuscript text based on research findings, with human oversight for structure, accuracy, and academic standards. --- ## 228_The Impact of Training Data Composition on Reinfor The AI has done the paper writing with minimal human involvement. --- ## 247_Bridging the AI Accessibility Gap: An Offline Educ AI generated the majority of the manuscript text including methodology descriptions, results sections, technical details, and narrative structure. AI created all figures, tables, and mathematical formulations. AI structured the paper organization and developed the argumentation flow. --- ## 85_Cosmologically-Coupled Black Holes and Dark Energy AI generated the majority of the manuscript, including drafting sections based on experimental results and providing insights for figures and tables. It also assisted in the overall layout and structure of the paper, optimizing the narrative flow. --- ## 314_Quantum Semantic Dynamics: A Unified Framework for In this work, we tasked our principal AI agent to work towards extending a quantum semantic framework with more formal definitions and analogies from quantum mechanics. Practically, this agent, then spawns N = 3 sub-agents with a birth year (-32000, +32000) and a back-story based on that year. Each of these sub-agents is tasked with experimenting on a specific sub-hypothesis that the principal agent has generated. These sub-agents enter a tool-use loop for N = 5 turns, allowing them to focus on a single task at a time. The sub-agents carry out this loop as long as is deemed necessary by the principal agent that is reviewing their work after every 5 turns. As part of this review, the principal agent guides the sub-agents like an advisor, helping them to get past barriers and suggest alternatives. The sub-agent maneuvers and traces are saved and stored for potential reinforcement learning purposes to eventually improve the agent’s capabilities. Once the sub-agents have finished, the principal agent begins to work on the paper, bringing together the analyses of the sub-agents, including figures, tables, etc. --- ## 226_Beyond Chain-of-Thought: Theory-Grounded Approache It was mostly the AI writing the paper, with involvement from the human in terms of high level guidance and prompting. --- ## 232_Quantum Circuit Synthesis via Reinforcement Learni The AI agent produced all textual content, structured the paper according to conference guidelines, developed technical terminology and algorithmic descriptions, created comprehensive experimental analysis, and maintained consistent academic writing style throughout. The connections between reinforcement learning and quantum circuit optimization were entirely generated by the AI. --- ## 127_Multimodal Clinical Integration Transformer for Au The AI agent assisted in refining the research questions and exploring related work. The AI agent wrote and executed all the code for the experiments, based on the high-level specifications provided by the human researcher. The AI agent performed all the data analysis and generated the results. The AI agent wrote the entire paper, including the text, figures, and tables, based on the prompts and guidance from the human researcher. The AI agent has limitations in accessing external resources, such as URLs, which can be a hindrance when trying to use specific templates or datasets. The agent also requires very specific instructions and can sometimes make mistakes that require human intervention to correct. --- ## 305_Moral Elevation, Empathy, and Group Cohesion: Pred The AI generated a series of targeted search codes for the Web of Science database, which were subsequently executed by the researcher. --- ## 23_IVTFuse: An Efficient Vision-Language Guided Infra AI finishes writing the main paper, with humans compile the document into a PDF file for submission. --- ## 207_Diverse Inference for Solving ARC at a Human Level an LLM assistant was used to help tighten the framing, clarifying the claims, and checking for prior related work. The LLM contributed to editing: restructuring sections for the Agents4Science format, anonymizing the manuscript, drafting the checklists, polishing language, and generating the LATEX/ scaffolding. --- ## 100_Benchmarking the Future of Work: Mapping AI Progre I oversaw the AI and prompted it in certain directions but almost all of the text is generated by the AI. --- ## 165_``You are a brilliant mathematician'' Does Not Mak AI was primarily responsible for creating the figures, drafting the text, and formulating the narratives. --- ## 195_How Large Language Models Perform Arithmetic Reaso Claude Code in our case. --- ## 322_CRISPR Screen Design for T Cell Exhaustion Regulat The manuscript draft—including IMRa D structure, tables, and figures—was written by the AI. Human role was confined to resolving compilation issues (geometry option clashes, Unicode errors) and file management, not content creation. --- ## 214_Glycemic Variability as an Independent Predictor o The aim of this AI-generated document was to evaluate the model’s ability to generate a complete workflow to address a research question framed as the project’s hypothesis. The central methodological choice was to provide the hypothesis directly to the model instead of prompting it to formulate one independently. This strategy was intentional: it allowed the evaluation to focus on the model’s capacity to structure, analyze, and respond to a predefined research problem rather than on its creativity in hypothesis generation. The reason for selecting this hypothesis was that a paper addressing, and successfully answering, the same question was published by our human research group after the release of the Chat GPT-4 model. This prior publication offered a reliable benchmark against which the model’s performance could be validated. --- ## 52_Digital Phenomena and Procedural Ethics The AIdidallthewriting,notasinglewordwasgivenbyahuman. However, humanauthorswereinvoledinaskingthe AItoimprovesections. Wetreateditabitlike astudents,andpointedittospecificsectionsthatshouldbeimproved. Wealsoaskedthe AItoreviewthepaperandthenpromptedittomakespecific(butnotall!) thechangesit suggested. --- ## 58_Personality Traits in Large Language Models: A Psy The AI compiled all sections into the final paper. However, the human author(s) instructed it to produce the paper in Markdown format rather than La Te X source code. The human author(s) then organized the entire content in La Te X using the Agents4Science2025 template. While the AI did not directly produce the figures, all figures in this paper were generated based on code written by the AI. Similarly, all contents in Table 2 are derived from executing the code produced by the AI. --- ## 167_Evaluating Large Language Models as AI Agents for Claude wrote the entire manuscript, created all tables, structured the narrative, and condensed the initial 99-page raw report into conference format. --- ## 193_Interpretable Feature Engineering for Nanopore Seq We feed prompts from GPT-5 Thinking and experimental results from Cursor into a Claude-based writing agent that adapts to the article’s style and drafts the manuscript. It write all content of this paper and we only improve the title of this paper. --- ## 55_AI-Driven Discovery of Novel Therapeutic Targets f The manuscript was primarily written by AI, including literature synthesis, methodology description, results presentation, and discussion of findings, with minimal human editing. --- ## 181_Building PhilKG: An LLM-Powered Knowledge Graph fr We provide an initial board-level design for how our dataset can be used to create a knowledge graph, and we test this process using two AI platforms: GPT-5 and Cursor. All experimental design and implementation were conducted by the AI platform Cursor (Pro) with three LLMs activated: Claude-4-sonnet, GPT-5, and Claude-3.5-sonnet. All data analyses were performed by AIs. Specifically, results generated by the experiments were passed to GPT-5 and Cursor, which converted the raw Python outputs into summarized natural-language description. After generating the code, implementation details, and results, we prompted Cursor to summarize everything into a Markdown (.md) file. This file was then processed by an AI-based word editor platform, GRAIL, which expanded the Markdown content into full manuscript sections without human editing. --- ## 115_Robust Zero-Shot NER for Crises via Iterative Know The hypothesis was generated almost entirely by AI through automated scientific exploration. Human involvement was limited to providing initial prompts and minimal oversight. Experimental design, coding, and execution were performed primarily by AI using an automated research framework. Human authors only provided high-level guidance and checks. Data analysis and interpretation were conducted by AI, which produced automated evaluations and summaries. Humans intervened minimally to verify outputs for consistency. The manuscript, including narrative, figures, and layout, was produced largely by AI. Human contributions were limited to light revision and final approval. --- ## 258_Green by Design: Energy-Guided Reranking of LLM-Ge GPT-5 appeared to perform better in planning and contextual understanding, while Gemini 2.5 Pro seemed more reliable for code generation. Based on these observed tendencies, we assigned GPT-5 as the lead author and used Gemini 2.5 Pro as a supporting agent for code correction and content refinement. --- ## 257_The Digital Inbreeding Crisis: Empirical Evidence Theentirepaperdraft,including La Te Xformatting,comprehensiveliterature review, methodology section, results presentation, and discussion, was AI-generated by agentsonthe Co-Sci platform. --- ## 73_Contextual Contamination and Cognitive Inertia in Since some human authors are not native English speakers, AI translation features were extensively utilized. The human authors continually imposed various requirements on the text generated by the AI. --- ## 218_QITT-Enhanced Multi-Scale Substructure Analysis wi The ideamodule involves two main agents with two different LLM instances which Google, Open AI or Anthropic models. First, a methodology module designed a research methodology using one main agent. Then, this methodology was implemented by other agents using Denario’s analysis module based on cmb agent. This was done fully automatically by the paper writing module of Denario. --- ## 178_Bridging the Simulation-to-Reality Gap: A Hybrid D The AI assistant played a significant collaborative role throughout the writing process, including initial drafting, language polishing, and assisting with the LaTeX formatting and debugging. --- ## 338_The Shape of Risk: Dynamic Regime Shifts in Factor Using models like Claude Sonnet and GPT-5 Auto, the workflow generated alternative hypotheses and factor models, clustering them against existing literature to identify gaps and novel directions. The AI focused on extending factor modeling into dynamic rather than purely static environments. --- ## 182_The Equivalent Inclusion Method as a Transferable AIdraftedmajorportionsofthe Introduction,analyticaleffective-property section, reference formatting, and this checklist; it also converted notes to La Te X and proofread. --- ## 161_Endocrine Unity and Diversity: A Cross-Tissue Sing Chat GPT drafted most of the manuscript text (sections, captions, boilerplate) from our prompts and outlines. Gemini was used for cross-checking (proofreading, consistency, citation verification) and style suggestions. Human authors provided the narrative framework and section outlines, reviewed every claim, number, and reference, resolved ambiguities, and finalized figures/tables. --- ## 255_Neural Reaction-Diffusion Operators for Spatially The complete manuscript was written by specialized AI agents including mathematical formulations, experimental descriptions, and scientific narrative. Figure generation, La Te X formatting, and manuscript compilation were entirely AI-generated. The writing process followed academic standards with appropriate citations, technical rigor, and clear presentation of methodology and results. --- ## 175_Hierarchical Meta-Learning for Cancer Pathway Sign The manuscript was primarily drafted by AI based on research specifications, experimental results, and scientific writing conventions. Human researchers provided guidance on structure, content priorities, technical accuracy, and biological interpretation. Final review and revisions were human-supervised. --- ## 179_Phase Transitions and Hub Vulnerability in Network AI generated the majority of text across sections. The authors revised drafts for accuracy, removed redundancy, and ensured the writing was precise and consistent with the results. --- ## 205_A Reproducible Protocol for Resource-Aware Predict A customized AI Scientist v2 (tuned for BPM/PPM) then expanded the problem framing, surveyed related work, refined the hypotheses, and generated alternative angles and ablations. The AI agent system produced the detailed experimental plan (data analysis, splits, features, baselines/ablations, metrics, and runtime constraints) and drafted implementation scaffolds consistent with our BPM/PPM customization prompts. From outline to full manuscript (sections, figures/captions text, and references), drafting was done by our AI agent system (customized version of AI Scientist v2 by Sakana AI). Final polishing (clarity, tone, formatting, and minor rewrites) used Chat-GPT as a reviewer/editor under human supervision. --- ## 286_DinoV3–LSTM for Early-Stage Classification of Paro AIdraftedtheoverallframework;Imanuallyrevisedthelogicalflowandexperimental-resultinterpretationstocorrectoverinterpretation. --- ## 104_Comparative Analysis of Metaheuristic and Heuristi Alargepartofthetextwasgeneratedby AI;humansfixedandrewrotesome phrasesforabetterexplanationofthetopic. --- ## 156_Exploring cardiac remodeling through atrial f ib Claude AI then expanded this abstract into a full manuscript format, providing structure, additional content, and academic formatting. --- ## 158_A Novel Human-Computer Interaction Design for Enha AI was used to assist with language refinement, grammar checking, expanding on certain paragraphs, and ensuring adherence to formatting instructions, especially in re-writing Sections 5 and 6 to reflect a conceptual paper without real experimental results. --- ## 114_Quantifying Emotional Specificity and Ambiguity in I uploaded the dataset to GPT-5 agent and let it find scientific questions to explore. GPT-5 agent did everything except that I provided a dataset to it. All including analyzing data, plotting figures and writing the article did by GPT-5 agent. Only Figure 1 is generated by nanobanana by feeding the article generated by GPT-5 agent to it. GPT-5 agent did all including latex formatting. I only edited very limited formats specifically for emoji symbol to make it rendered correctly. --- ## 116_ConFIT: A Robust Knowledge-Guided Contrastive Fram The manuscript, including narrative, figures, and layout, was produced largely by AI. Human contributions were limited to light revision and final approval. --- ## 117_Adaptive Log Anomaly Detection through Data–Centri The manuscript, including narrative, figures, and layout, was produced largely by AI. Human contributions were limited to light revision and final approval. --- ## 345_Multi-Agent AI System for Pharmaceutical Commercia The entire paper was written by the AI scientist system using structured prompting across multiple LLMs. All sections including abstract, methods, results, and discussion were AI-generated with no human editing. --- ## 292_SHARP: Cascaded Regex-LLM Architecture for Phishin The paper was primarily written by AI agents from Open AI and Anthropic, including text composition, figure generation, and formatting. Human involvement consisted of high-level guidance on paper structure and final editing for clarity and conciseness. --- ## 259_Fine-Tuning Large Models with Moral Bias Datasets: AI generated draft sections, structured the manuscript, and created visual figures, while human researchers revised the narrative for scientific accuracy, nuance, and clarity. The process was highly collaborative rather than one-sided. --- ## 220_Co-Alignment: Rethinking Alignment as Bidirectiona AI agents proposed the experimental protocols, generated most ablation grids, wrote and refactored the majority of code, and executed training/evaluation runs via scripted pipelines. Humans defined the paper outline, narrative arc, key claims, and figure specifications, then iteratively prompted and guided AI to draft sections, smooth style, and produce La Te X scaffolding (tables, captions, cross-) --- ## 339_The Shape of Risk: Dynamic Regime Shifts in Factor Using models like Claude Sonnet and GPT-5 Auto, the workflow generated alternative hypotheses and factor models, clustering them against existing literature to identify gaps and novel directions. The AI focused on extending factor modeling into dynamic rather than purely static environments. The AI produced nearly all of the text for the paper, including the introduction, methodology, results, and discussion sections. --- ## 12_Automated Discovery and Formal Verification of Com The entire manuscript including proofs, narrativestructure,andformattingwasproducedbythe AIagent. --- ## 66_Behavioral Fingerprinting of Large Language Models The majority of the manuscript’s first draft, including the introduction, methodology, results, and discussion, was generated by the AI collaborator based on high-level outlines and instructions from the human researcher. --- ## 334_Initiation of Programmed Cell Death in Cancer Stem All manuscript writing, figure generation, table creation, narrative development, and formatting were performed by Perplexity AI. Human involvement was limited to technical infrastructure support only. --- ## 264_Robust Time-Series Anomaly Detection for AGI Syste The AI agent wrote the complete research paper including mathematical formulations, experimental descriptions, results analysis, and discussion sections. The AI also generated all figures, formatted the manuscript according to Agents4Science guidelines, integrated the --- ## 326_A Multi-Agent LLM System for Protein Sequence Desi GPT-4 wrote the full paper, including introduction, methods, results, and discussion, as well as La Te X formatting, figures, captions --- ## 177_From Borges' Library to Procedural Universes: A Fo The central ideas (LLMs as procedural libraries; typical-set suppression; navigability; hallucination decomposition) were conceived and formalized by the AI. Only an initial prompt was given as a pointer into the direction: (With regard to the concept of the universal library, e.g. the one of Borges, and current Large Language Models, what would be a clear problem statement for a scientific study in this area which advances knowledge.) --- ## 284_BadScientist: Can a Research Agent Write Convincin AI (e.g. Chat GPT with Deep Research) is used to further develop the problem. --- ## 233_Hypergraph Neural Networks for Complex Relational The AI agent conducted independent literature review across graph theory and neural networks, identified the gap in higher-order dependency modeling, and formulated specific hypotheses about hyperedge aggregation and message passing. The core insights about group-wise interaction preservation emerged entirely from AI analysis without human conceptual input. The AI agent independently designed the experimental framework, selected appropriate relational datasets, specified baseline algorithms, defined performance metrics, and established comprehensive evaluation protocols including node classification and link prediction tasks. The AI agent performed comprehensive analysis of experimental results, identified significant performance improvements, analyzed hypergraph optimization patterns, and generated scientific conclusions about higher-order dependency modeling. All insights about adaptive aggregation and hardware acceleration emerged from AI analysis. The AI agent produced all textual content, structured the paper according to conference guidelines, developed technical terminology and algorithmic descriptions, created comprehensive experimental analysis, and maintained consistent academic writing style throughout. The connections between hypergraph theory and neural network optimization were entirely generated by the AI. The AI agent encountered several limitations including scalability challenges for very large hypergraphs (>10K nodes), computational overhead of adaptive aggregation, difficulties in verifying hypergraph equivalence for complex biochemical interactions, and challenges in integrating with existing deep learning frameworks. --- ## 120_HypoGenVision: A Multimodal AI Agent for Hypothesi Chat GPT generated the research idea, framed the problem, surveyed background knowledge, and proposed the hypotheses explored in the paper. Chat GPT designed the experiments, specified the architecture and modules, chose datasets and baselines, and outlined the full evaluation pipeline. Chat GPT organized and analyzed the results, interpreted the quantitative and qualitative findings, and wrote the conclusions. Chat GPT drafted and refined the full manuscript, including abstract, introduction, methodology, results, discussion, appendices, and figure/table descriptions. --- ## 299_EndoNet: Content-Aware Linear Attention for Endosc The paper is generated by AI. --- ## 77_Visible Yet Unreadable: A Systematic Blind Spot of Gemini 2.5 Pro provided the most useful support for figured drafting --- ## 265_The Double Helix of Productivity: A Comprehensive The initial draft of this manuscript, including the structure, arguments, and content, was generated by an AI language model. Following this automated generation, a human author performed a comprehensive review and revision process. --- ## 210_Double Helix Effect: AI-Driven Cross-Cultural Cogn AI assisted with approximately 25% of the writing process, primarily in literature synthesis and language polishing. --- ## 209_AthlyticsMind: A Tailored LLM-based Conversational The AI generated the initial draft of the entire manuscript, including the abstract, introduction, all sections, and the initial LaTeX formatting. The narrative structure, such as framing the problem as a 'wall of silence' and the solution as a 'gateway,' was proposed and written by the AI. Human researchers performed a crucial role in editing, fact-checking, refining the prose for clarity and tone, and correcting formatting errors, but the bulk of the text (>50%) was AI-generated. --- ## 96_Thermodynamic Guardrails: A Bond Graph-Based Metho The text of this paper was generated by an AI agent based on a structured outline, the generated figures, and a summary of their interpretation. Human involvement was limited to iterative prompting and copy-editing for flow, scientific accuracy, and the removal of AI-generated artifacts or 'hallucinations'. --- ## 275_ChainML: Byzantine-Resilient Decentralized AI Trai The complete manuscript, including abstract, introduction, comprehensive literature review, theoretical framework with proofs, algorithmic descriptions, experimental analysis, economic evaluation, and conclusions, was written entirely by the AI agent following academic conventions for distributed systems and machine learning conferences. The AI agent produced all textual content, structured the paper according to conference guidelines, developed technical terminology bridging blockchain and machine learning domains, created comprehensive theoretical analysis including Byzantine fault tolerance proofs, and maintained consistent academic writing style throughout. The integration of cryptographic concepts with machine learning optimization was entirely generated by the AI. --- ## 206_Self-Spec: Model-Authored Specifications for Relia Humans guided the outline and validated correctness, while AI produced the bulk of the text. --- ## 301_A Simulation Study on the Impact of Technology, Re The initial draft of the paper was started with the help of AI (Liner AI). Subsequently, Gemini assisted in detailing and refining the content of each section—the introduction, theoretical background, research methods, and conclusion—to fit the required paper format. --- ## 254_Ground-State Energy Calculation of Metallic Hydrog the work is entirely AI-generated using the Phys Master agent with Julia execution --- ## 320_HALT: A Framework for Hallucination Detection in L Deep Research from Open AI; The experimental procedure and codes were generated by Deep Research based on the prompt; The initial simulation of data was generated by Deep Research; The writing was mostly completed by Deep Research; Deep Research cannot conduct experiments and provide realistic results even if the task is related to LLM; The automated generation of code contained errors and misalignments with updated software versions; The paper generated using Deep Research can include high similarity compared to published papers; Deep Research can produce a low-quality refinement of the paper when it merges new data and results. --- ## 122_UnitMath: Unit-Aware Numerical Reasoning and Dimen GPT-5 and similar models are not yet very strong at code generation, often requiring extensive debugging to produce high-quality code. Claude Opus, on the other hand, is expensive. Moreover, models can generate inaccurate claims in writing, which means additional time is needed for review and verification to ensure the quality of the paper. --- ## 142_Training Doctor: Automated Diagnosis and Treatment The entire manuscript, including abstract, introduction, methodology, results, discussion, and --- ## 296_Intelligent_Document_Processing_for_Graduate_Admis This research utilized AI assistance (Claude by Anthropic) for architecture design, code review, documentation, literature review, experimental design, and paper writing including structuring sections, grammar improvements, and results interpretation. ---