연구용 AI도구 최신동향 업데이트 (2025.4)

English summary

This presentation reviews developments in AI tools for academic and scientific research as of April 2025. It covers reasoning-oriented large language models and autonomous AI agents, literature search and deep research, natural-language data analysis and visualization, scientific writing support, and policy and ethical limitations in research use.

Key points

• Reasoning models use additional computation during response generation and may improve accuracy on complex reasoning, coding, and mathematical tasks. They require more tokens and longer response times, however, and offer limited benefit for simple text polishing.
• Autonomous AI agents combine planning, memory, tool use, and iterative execution. The presentation describes a test in which an agent spent 11 minutes and 40 seconds using code generation and web searches to solve a clock-hand angle problem but ultimately produced an incorrect answer.
• Sakana AI's The AI Scientist is presented as an automated research pipeline covering idea generation, literature-based novelty checking, in-silico experiments, plotting, manuscript drafting, and automated review. According to the presentation, one of three generated papers submitted to an ICLR 2025 workshop in March 2025 was accepted with a reviewer score of 6.3 out of 10. This was a specific reported case rather than a general performance estimate. Scholarly AI tools use retrieval-augmented generation to retrieve indexed literature with the aim of reducing citation hallucinations. Elicit, Consensus.app, and several Deep Research functions are presented as supporting paper screening, summary tables, and linked evidence synthesis. NotebookLM is described as supporting document-grounded questions, citation anchors, audio overviews, interactive questioning, and concept-based mind maps. ChatGPT Data Analyst and Google Colab with Gemini are presented as examples of systems that can generate exploratory analysis plans and Python code from raw data, produce visualizations and regression analyses, and suggest subsequent machine-learning workflows. In the presentation's Warfarin dosing example, VKORC1 genotype was identified as the strongest predictor, with an R2 value of 0.53. Claude SVG generation, Napkin AI, ChatGPT image generation, and Gamma.app are compared as options for editable vector diagrams, template-based infographics, visual mockups and diagrams, and presentation generation, respectively. The December 2023 NEJM AI editorial position is described as encouraging LLM use to reduce language barriers and improve manuscript readability, while viewing outright bans as impractical because AI detectors produce false positives and false negatives and can disproportionately affect non-native English writers. The presentation emphasizes protecting unpublished manuscripts and research data by checking whether data may be used for model training and by using temporary chats or enterprise security settings when appropriate. It also identifies the boundary between grammatical editing and ghostwriting as an unresolved policy issue. The presentation warns against an illusion of explanatory depth, in which fluent AI explanations are mistaken for deep personal understanding; an illusion of exploratory breadth, in which AI is assumed to search all possible hypotheses; and an illusion of objectivity, in which AI outputs are treated as unbiased. It discusses possible effects of automation dependence, including loss of foundational skills, reduced vigilance during monitoring, weakened manual intervention when automation fails, and loss of situational awareness. AI literacy and critical evaluation of outputs are therefore presented as necessary educational components. Drawing on METR data, the presentation states that the task horizon associated with a 50 percent success rate on complex software and research tasks has been doubling approximately every seven months, and extrapolates that agents may conduct month-long research tasks around 2030. This is a projection in the presentation, not an established result.

Audience

The material is intended for academic and medical researchers, medical educators, graduate students, faculty, and others interested in scientific writing and data analysis.

Scope and limitations

This summary is based only on the supplied presentation evidence notes. The tool examples, performance figures, cited literature, and 2030 projection should not be treated as independently verified findings from a systematic review or reproducibility assessment. Tool performance can vary with model version, data access, prompting, verification procedures, and task characteristics, and the presentation does not provide comparative performance or error rates for all tools.

Research context

The presentation is thematically connected to medical AI and data science through its focus on integrating literature retrieval, data analysis, visualization, scientific communication, and research automation.