We model the research loop as a cycle of assertion, experiment, analysis, description and dialogue, and assertion. Among these steps, the assertion and analysis steps, including hypothesis generation and verification, require AI that can understand scientific data, generate new hypotheses, and verify if experiment results are consistent with hypotheses.
To address this, we develop a "multimodal XAI foundation model" for hypothesis generation and verification AI to understand scientific data. We then implement hypothesis generation through "hypothesis generation modeling using scientific knowledge embedding and abduction." We develop a "hypothesis inspiration AI" based on human scientist feedback and establish hypothesis verification using "experiment prediction and result XOR discovery AI." These interact with interactive hypothesis generation AI developed through "knowledge inference and dialogue-based multimodal hypothesis generation" to realize hypothesis generation and verification.

After demonstrating the 2023 fiscal year milestone of “AI robots being able to mutually understand existing research described in papers through knowledge exploration using literature,” in the 2024 fiscal year, we not only deepened these efforts but also worked on realizing hypothesis generation by AI as a new research area. As a result, we were able to demonstrate hypothesis generation in multiple fields such as information science and chemistry.

(1)  Multimodal XAI Foundation Model

We constructed a model that generates hypotheses and verification methods expected to receive high evaluations by utilizing multi-modal paper understanding XAI as a virtual evaluator. We addressed the challenges of the base model and developed hypothesis generation technology using a crossover method. In the field of chemistry, we realized a cyber-physical loop from initial hypotheses of 39 molecules to virtual synthesis, automatic organic synthesis, activity testing, and the generation of new molecules.

(2)  Hypothesis Generation Modeling Using Spatial Embedding of Scientific Knowledge and Abduction

We aim to automatically generate novel and plausible scientific hypotheses by embedding knowledge from scientific and technical literature into a continuous vector space using large-scale language models. We have started building a baseline system and realized context-aware molecular candidate generation using a RAG component. We have also developed a chatbot-style interactive hypothesis generation tool and a research trend visualization tool.

(3)  Hypothesis Inspiration AI

We have worked to establish a seamless collaborative environment between foundation models and chemical researchers. We have developed a system that automatically detects SMILES notation in Slack and displays molecular structure images, and integrates with a compound editor to provide feedback on researchers' intentions to the foundation model.In addition, we proposed and confirmed the effectiveness of a method that generates pseudo labels from a small amount of researcher judgments using UU Learning and utilizes them for additional fine-tuning of the base model.

(4)  AI for discovering XOR between experimental predictions and results

We have worked on building AI that discovers the XOR between predicted results and actual results during experiment planning, which serves as a basis for generating new hypotheses. We collected over 20,000 papers from ChemRxiv to build a new dataset and developed hypothesis verification AI using two language models, an encoder-based model and a decoder-based model, achieving over 95% accuracy. We also implemented a function to visualize the basis for judgments.

(5)  Multi-modal hypothesis generation using knowledge inference and dialogue

We worked on building a paper understanding framework using large-scale language models. We prepared fine-tuning data from a patent office database and developed an automatic rewriting model for patent claims. We built a hypothesis generation taming AI using prompting and collaborated with other teams to realize a system that can evaluate the plausibility of hypotheses. We also considered distilling causal relationship data and utilizing quantum machine learning.

When this research and development project began in 2023, our goal was to realize AI capable of understanding literature by 2025 and AI capable of generating literature (with experiments) by 2030. However, based on the results achieved by the end of fiscal 2024, the former milestone has been achieved, and hypothesis generation, which is the most important part of the latter milestone, has also been realized.
Going forward, we will not only focus on realizing hypothesis generation but also establish a multi-agent AI system that can evolutionarily generate reports including experiments for hypotheses generated and verified, in collaboration with an automated synthesis experiment AI. Through the open-sourcing and social implementation of this multi-agent AI system, we plan to contribute to the widespread adoption of AI robot-driven science.