The rapidly growing popularity of adopting Artificial Intelligence (AI), and specifically Large Language Models (LLMs), is having a widespread impact throughout society, including the academic domain. AI-supported research has the potential to support researchers with tasks across the entire research life cycle. In this work, we demonstrate the TIB AIssistant, an AI-supported research platform providing support throughout the research life cycle. The AIssistant consists of a collection of assistants, each responsible for a specific research task. In addition, tools are provided to give access to external scholarly services. Generated data is stored in the assets and can be exported as an RO-Crate bundle to provide transparency and enhance reproducibility of the research project. We demonstrate the AIssistant's main functionalities by means of a sequential walk-through of assistants, interacting with each other to generate sections for a draft research paper. In the end, with the AIssistant, we lay the foundation for a larger agenda of providing a community-maintained platform for AI-supported research.
The recent advancements of Artificial Intelligence (AI), in particular generative AI, such as Large
Language Models (LLMs), have a profound impact on our society already [
To address these challenges, we present the TIB AIssistant (i.e., AI-assistant). The TIB AIssistant is a
domain-agnostic AI-supported research platform that helps researchers, by means of AI, in various
steps of the research life cycle. This work demonstrates a concrete implementation of our vision of the
TIB AIssistant, which is published as a vision paper [
The industrial revolution and afterward the digital revolution had a substantial impact on all aspects
of life [
There are many individual approaches to leveraging LLMs for research. Some of these approaches aim to create a fully automated research life cycle using AI. This includes the work of Sakana AI by means of the AI Scientist, which fully automates the research life cycle by iteratively executing life cycle phases [17]. There are several shortcomings of their approach, for example, the limited ability to ifnd related literature, and issues during code generation leading to failing experiments [ 18]. While fully automated approaches are an interesting research avenue and are surely going to be improved in the near future, we take a diferent approach with the AIssistant. We believe that researchers must be at the center of doing research, and therefore, AI should be an assistive technology, leaving the human researcher in control of the process.
The AIssistant’s approach consists of three main components: assistants, tools, and assets. We will discuss those concepts on a high level to lay the foundation for the platform.
Assistants are agents that are designed to perform a single task. For example, formulating research questions or coming up with a paper title. Assistants consist of a name, system prompt, LLM model, what assets they consume and generate, and the tools that can be called. The LLM model selection determines which model to use to accomplish the task. Simpler tasks can be performed by smaller models, reducing the required resources and waiting time.
Tools are generally external web services that are called via REST endpoints. The tools have a name, a textual description, an input schema, and an execution function. The description and input schema are provided to the LLM to provide instructions on when the tool should be called and how it should be called. This is handled via the Tool Calling mechanism [19]. Each assistant specifies the tools that are callable, which limits the space of possible tools and thus improves the tool selection accuracy. The output of tools is either sent to the chat as text or handled by displaying specialized UI components. This follows the paradigm of Generative UI, where UI components are shown on demand, depending on the task at hand. In section 4, we demonstrate several implemented generative UI components.
Assets are the data store that assistants can access. This difers from the conversation’s history
(i.e., context window) as it is structured and only contains data that the user selects. Assets have a
name and a data type (e.g., text, object, JSON, etc.). Assistants can have required input assets, which
either consist of previously generated assets or are manually provided by the user. Reproducibility is
one of the cornerstones of science. Therefore, being transparent about how AI was used to support
the researcher’s work is a crucial aspect of our approach. The assets are central in our approach, as
they serve as output of the generated knowledge. We provide an Export assets functionality, which
compresses assets into an archive file, containing individual files for each asset, and a metadata file
which provides provenance data for the files. We use the RO-Crate [
A screenshot of the prototype user interface is displayed in Figure 1. The previously discussed key components are depicted in this screenshot. The three main components, Assistants, Assets, and Tools, are configurable by users. By default, the selected assistant decides what assets and tools are activated, but users can modify this selection to fit their use cases. We will demonstrate here via a system walk-through how a research life cycle using the AIssistant can look like. Afterward, we discuss the technical details of our prototype.
In our walk-through, we adopt the stance of a researcher who is interested in intertwining Semantic Web and AI research. We will follow the life cycle as depicted in the sidebar of Figure 1. For demonstration purposes, we only show domain-agnostic assistants, domain-specific ones are disabled, as can be seen in the screenshot. A list of assistants is provided in Table 1. A demonstration video is available online.1
Firstly, the researcher wants to generate ideas related to their topic of interest in the Ideation assistant. This assistant asks the user to provide content they are interested in. In our example, the researcher provides a DOI from related papers and an ORCID, which will respectively fetch related content from Crossref and list articles from ORCID. After refining the ideas via the chat, the researcher adds the final topics to the ideation asset. Next, the researcher goes to the Research questions assistant and inputs the previously generated ideation topics from the assets. After iterative refinement in the chat component, the final research questions are added to the assets.
As a next step, the researcher aims to find related work via the Related literature assistant. There, the research questions are provided to the chat and are used to find related work via ORKG Ask [ 21]. The user manually reviews the suggested literature and adds relevant articles to the bibliography. This is done via the generative UI approach, where a dedicated literature search component is displayed in the chat window. This component is displayed only for tools that support literature search. It lists articles, allows users to select articles, and provides pagination to explore the listed items further.
Afterward, the paper writing assistants are used: Paper title, Related work, and Proofread. The first two assistants use the ideation topics and research questions. The related work assistant also uses the bibliography to connect related work to their citations. The final text and bibliography can be exported to LaTeX and imported to an Overleaf project as a draft for writing the actual manuscript. The proofread assistant uses another type of generative UI component, providing a dedicated interface where users can accept and reject proposed changes via a track-changes-like interface. As a final step of the walk-through, the authored text is reviewed via the Review assistant.
The AIssistant is implemented in Typescript and React using the Next.js framework and is published as open source software under a permissive MIT license.2 We use Next.js both as frontend and backend, with the backend primarily serving as a wrapper for making calls to the LLMs. We use the OpenAI API, and specifically the GPT -4o mini model, which provides a good balance between costs and performance. The Vercel AI SDK ensures that model providers and individual models can be changed when necessary. Local browser storage via IndexedDB is used to store assets and chat history. For authentication, we use ORKG accounts with enable single sign-on. The service can only be used when the user is signed in to ensure the number of daily used tokens can be efectively limited as a means of cost management. In the future, we will implement a Bring Your Own Key (BYOK) approach, allowing users to utilize their own API keys to access the platform without imposed token limits. To encourage community development of both the platform and content, we have published development documentation online.3
The TIB AIssistant is not a direct alternative or replacement of existing LLM user interfaces. The AIssistant is tailored toward research and provides various advantages compared to regular LLM interfaces. First and foremost, the integration of assistants provides researchers with a library of curated prompts, which enables them to use LLMs within their research without the need to craft these prompts themselves. LLMs themselves provide almost limitless possibilities for research, but without curated lists of use cases (i.e., prompts), for many researchers, these benefits remain out of reach. Furthermore, the integrated tool library focuses on research-oriented tools and services, which are not typically provided by regular LLM interfaces. Tools can utilize dedicated user interface components, via the generative UI paradigm, to ofer additional interaction possibilities beyond the conversational (i.e., text-only) approach of regular LLM interfaces. In the end, we envision the AIssistant as a platform that integrates many diferent prompts, models, and tools from a wide variety of research domains, created and maintained by the research communities themselves. Alongside the open source nature of the AIssistant, the collaboration of research communities to support diferent domains is a key aspect of our approach. This sets the AIssistant apart from existing approaches and other commercial initiatives. To facilitate community collaboration, we have provided developer documentation and an easy-to-use format for defining assistants to encourage community contributions. At the current stage, we have provided a framework in which the previously mentioned aspects can be integrated. We demonstrated how diferent assistants and prompts can be used to perform research tasks, however, the actual implementation of a large-scale domain-specific prompt and tool library is left for future work. To summarize, although the individual tasks within the assistant can indeed also be executed by existing LLM interfaces, the AIssistant platform is tailored toward scholarly research. The integration of diferent scholarly tools and domain-specific prompts into a single platform makes it possible for any researcher, regardless of the AI proficiency, to leverage AI in their workflows.
In the current implementation, all assistants use the same model (GPT 4o-mini). The technology stack supports switching models and providers. For now, we do not want to be limited by the capabilities of the model, e.g., the ability to select the right tool for tool calling, or the capability to generate research ideas for arbitrary domains. However, we envision that assistants will use a model tailored to the task at hand. This means that smaller, locally hosted, or fine-tuned models are most likely viable options for various assistants. Using such approaches reduces the models’ required resources, benefiting both costs and environmental impact, and is therefore an important research direction for future work. Another aspect of future work is the evaluation of our approach. This will focus both on specific assistants and how well they are able to accomplish their respective objectives, as well as evaluating the platform from a usability perspective. Experimenting with these smaller models is out of scope for this work.
To conclude, in this work, we demonstrated what an AI-supported research assistant looks like, implemented in the TIB AIssistant. With this, we lay the foundation for a comprehensive communitymaintained platform for researchers who want to leverage AI for their work. We demonstrated how diferent assistants work together by storing intermediary results in the assets storage. Also, we showed how external tools play an essential role in providing a research-oriented platform. Finally, we used a RO-Crate export functionality to provide machine-readable provenance data of the generated content. In the future, we plan to further simplify the process of adding assistants and tools to foster 3https://tibhannover.gitlab.io/orkg/tib-aissistant/web-app/storybook further community involvement. To accomplish this, we are experimenting with no-code solutions, where graphical user interfaces allow users to configure the platform further. Additionally, we plan to experiment with diferent models and add additional assistants and tools to support more complete research life cycles for various domains.
We thank our colleague Mohamad Yaser Jaradeh for his valuable comments in reviewing this paper Additionally, we want to thank the entire AIssistant team for their contributions to the platform, including research and development eforts. This work was co-funded by NFDI4DataScience (ID: 460234259) and by the TIB Leibniz Information Centre for Science and Technology.
During the preparation of this work, the authors used ChatGPT and Grammarly in order to: Paraphrase and reword, improve writing style, and Grammar and spelling check. Also, the authors used OpenAI TTS for generating the voice-over in the demo video. After using these services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [11] Z. Z. Chen, J. Ma, X. Zhang, N. Hao, A. Yan, A. Nourbakhsh, X. Yang, J. McAuley, L. Petzold, W. Y.
Wang, A survey on large language models for critical societal domains: Finance, healthcare, and law, arXiv preprint arXiv:2405.01769 (2024). doi:10.48550/arXiv.2405.01769. [12] C. Si, D. Yang, T. Hashimoto, Can llms generate novel research ideas? a large-scale human study with 100+ nlp researchers, arXiv preprint arXiv:2409.04109 (2024). doi:10.48550/arXiv.2409. 04109. [13] F. Bolanos, A. Salatino, F. Osborne, E. Motta, Artificial intelligence for literature reviews: Opportunities and challenges, Artificial Intelligence Review 57 (2024) 259. doi: 10.1007/ s10462-024-10902-3. [14] W. Liang, Y. Zhang, Z. Wu, H. Lepp, W. Ji, X. Zhao, H. Cao, S. Liu, S. He, Z. Huang, et al., Mapping the increasing use of llms in scientific papers, arXiv preprint arXiv:2404.01268 (2024). doi:10.48550/arXiv.2404.01268. [15] J. White, Q. Fu, S. Hays, M. Sandborn, C. Olea, H. Gilbert, A. Elnashar, J. Spencer-Smith, D. C.
Schmidt, A prompt pattern catalog to enhance prompt engineering with chatgpt, arXiv preprint arXiv:2302.11382 (2023). doi:10.48550/arXiv.2302.11382. [16] B. Meskó, Prompt engineering as an important emerging skill for medical professionals: tutorial,
Journal of medical Internet research 25 (2023) e50638. doi:10.2196/50638. [17] C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, D. Ha, The ai scientist: Towards fully automated open-ended scientific discovery, arXiv preprint arXiv:2408.06292 (2024). doi: 10.48550/arXiv. 2408.06292. [18] J. Beel, M.-Y. Kan, M. Baumgart, Evaluating sakana’s ai scientist for autonomous research: Wishful thinking or an emerging reality towards’ artificial research intelligence’(ari)?, arXiv preprint arXiv:2502.14297 (2025). doi:10.48550/arXiv.2502.14297. [19] T. Schick, J. Dwivedi-Yu, R. Dessi, R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, T. Scialom, Toolformer: Language models can teach themselves to use tools, in: A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, S. Levine (Eds.), Advances in Neural Information Processing Systems, volume 36, Curran Associates, Inc., 2023, pp. 68539–68551. [20] S. Peroni, D. Shotton, The spar ontologies, in: D. Vrandečić, K. Bontcheva, M. C. SuárezFigueroa, V. Presutti, I. Celino, M. Sabou, L.-A. Kafee, E. Simperl (Eds.), The Semantic Web – ISWC 2018, Springer International Publishing, Cham, 2018, pp. 119–136. doi:10.1007/ 978-3-030-00668-6_8. [21] A. Oelen, M. Y. Jaradeh, S. Auer, Orkg ask: a neuro-symbolic scholarly search and exploration system, arXiv preprint arXiv:2412.04977 (2024). doi:10.48550/arXiv.2412.04977.