Dentsu Digital and Dentsu to Present Five Research Achievements Deepening Human-AI Collaboration at JSAI Conference

Dentsu Digital Inc. (Headquarters: Minato-ku, Tokyo; President & CEO: Hisashi Takimoto) and Dentsu Inc. (Headquarters: Minato-ku, Tokyo; President & CEO: Chisato Matsumoto) will present five research achievements exploring new forms of collaboration between humans and AI at the Annual Conference of the Japanese Society for Artificial Intelligence (JSAI)*1, starting Monday, June 8. The research was conducted under the themes of 'enhancing creativity' and 'advancing evaluation and judgment' in the utilization of Generative AI and Large Language Models (LLMs). *To view the presentation papers, registration and login for the JSAI Annual Conference are required before and during the conference period (until June 12). The papers will be available for public viewing on the J-STAGE 'Proceedings of the Annual Conference of JSAI' page*2 around early July after the conference. Specifically, the following five research achievements have been compiled: 1. A Creative Generation Model that Learns to 'Deviate from Common Sense': A Creative Generation Model Using Counter-Intuitive Chain of Thought (CI-CoT) https://pub.confit.atlas.jp/ja/event/jsai2026/presentation/2N6-GS-2x-02 This study proposes a method for training LLMs, which tend to converge on common-sense solutions, by providing them with the 'thought process of intentionally denying common sense to achieve discontinuous leaps (CI-CoT)' used by skilled creators as training data through Supervised Fine-Tuning (SFT). Experiments confirmed that the proposed model can intentionally avoid average and conventional solutions, generating highly novel ideas through thinking processes like reversal, exaggeration, and conceptual combination. This indicates that the creativity of LLMs is not merely due to added randomness but can be acquired and controlled by learning thought protocols. Future prospects include developing a 'Critic' to distinguish between 'creative deviation' and 'inappropriate deviation' in generated ideas, and refining the model using Reinforcement Learning from Human Feedback (RLHF)*3. A GAN-like approach, where a Generator creating unconventional ideas competes with a Discriminator that adjusts them according to social norms, is also considered promising. 2. The Impact of Collaboration with AI that has Learned Others' Thinking Styles on Creative Task Outcomes https://pub.confit.atlas.jp/ja/event/jsai2026/presentation/1G4-OS-13b-03 This study empirically investigated the impact of AI persona characteristics on creative outcomes by comparing collaboration with a self-personalized AI versus an other-personalized AI. The results showed that the self-personalized AI was superior in usability aspects like ease of work and trustworthiness, while the other-personalized AI tended to enhance originality and the breadth of ideas. Furthermore, an analysis focusing on the thinking style distance between participants and the AI suggested that creativity is not maximized by simple homogeneity or maximum heterogeneity, but potentially at a moderate distance. Therefore, it is concluded that in designing AI for creative tasks, it is crucial to strategically introduce 'appropriate otherness' rather than mere self-optimization. 3. Observing 'Criteria Drift' with Training-Free GRPO: A Self-Evolving Evaluation Standard for Ad Copy Quality https://pub.confit.atlas.jp/ja/event/jsai2026/presentation/1G4-OS-13b-01 This study applied Training-Free GRPO (TF-GRPO) to automatically evolve evaluation criteria for ad copy quality, empirically observing 'Criteria Drift,' where the criteria transform during the improvement process. Experiments using a stepwise improvement method confirmed an increase in evaluation accuracy (R1, F1) across all models. Notably, fixedly injecting domain knowledge significantly improved the reproducibility of criteria extraction. However, the effect of Criteria Drift was found to be model-dependent, suggesting the importance of balancing model scale, reasoning ability, and room for improvement. This research presents a new framework for automatically verbalizing and evolving evaluation criteria that have traditionally relied on the tacit knowledge of experts, indicating potential applications not only in ad copy evaluation but also in education, generative AI evaluation, and fields requiring expert judgment. 4. Multi-perspective Urban Policy Evaluation using a Persona-based Delphi Method with LLMs https://pub.confit.atlas.jp/ja/event/jsai2026/presentation/1G4-OS-13b-05 This study verified a method for evaluating urban policy proposals from diverse perspectives using an iterative evaluation process similar to the Delphi method*4, based on personas and a facilitator powered by LLMs. By having 500 AI personas, constructed based on the PeopleModel*5, evaluate an urban policy proposal over three rounds across 13 evaluation axes, the study analyzed the structure of supporting factors, concerns, and missing information, in addition to tracking the overall evaluation trend. This research is expected to be applied to the initial examination and communication design of urban policies, not as a tool to automate consensus-building itself, but as a low-cost, reproducible framework to visualize 'what are likely to become points of contention' and 'what should be supplemented in explanatory materials' before a policy is publicly announced. 5. A Generative Interpretation Attempt to Convert the Visual Texture of Graphic Scores into Sound https://pub.confit.atlas.jp/ja/event/jsai2026/presentation/5F3-GS-10s-01 Graphic scores, used in contemporary and experimental music, provide performance cues through visual elements like lines, shapes, and symbols, rather than traditional staves or note names. This study focuses on Cornelius Cardew's graphic score 'Treatise' and proposes a method to convert visual features such as shape, line thickness, and texture into sound by combining the vision-language model CLIP, which learns the correspondence between images and language, with the music generation model MusicGen. Furthermore, by examining the conditions where constraints from visual information coexist with the uncertainty of AI generation, the study presents a framework for sonifying graphic scores while preserving their ambiguity, suggesting new possibilities for creation and performance in experimental music and media art. The use of AI is increasingly accelerating business efficiency and quality improvement for companies. AI utilization has become a key management agenda, and research in this area is advancing rapidly. In the marketing domain, in particular, there is a demand for higher sophistication in both 'functionality' and 'understanding of human preferences' to achieve higher-resolution analysis and output. Under our unique AI strategy 'AI For Growth,' announced in August 2024, we have been pioneers in developing AI solutions for advertising creative. In the same year, we jointly developed and introduced 'AICO2'*6, an AI ad copy generation tool that learned the thought processes cultivated by Dentsu's copywriters over many years. We also began offering services to support business and service development. Dentsu Digital provides the integrated marketing solution brand '∞AI® (Mugen AI)'*7, which utilizes AI to contribute to the advancement of many companies' digital marketing activities. Going forward, Dentsu Digital and Dentsu will leverage the results of this research to expand the possibilities of combining human and AI capabilities, further evolving our AI solutions and driving innovation in new marketing strategies, product development, ad planning, ad method research and development, ad evaluation, and idea evaluation. The Dentsu Group in Japan promotes its unique AI strategy, 'AI For Growth,' which aims to contribute to the growth of clients and society by combining 'human intelligence' and 'AI intelligence.' For more information on AI For Growth, please visit the following webpage: https://www.dentsu.co.jp/labo/ai_for_growth/ *1: The 40th Annual Conference of the Japanese Society for Artificial Intelligence (JSAI2026) will be held from Monday, June 8 to Friday, June 12, 2026, at G MESSE Gunma and online. https://www.ai-gakkai.or.jp/jsai2026/ *2: J-STAGE 'Proceedings of the Annual Conference of JSAI' page: https://www.jstage.jst.go.jp/browse/pjsai/-char/ja *3: RLHF (Reinforcement Learning from Human Feedback) is a reinforcement learning method that uses human evaluation (feedback) to improve AI behavior. *4: The Delphi method is a technique for reaching consensus by repeatedly asking a group of experts the same questions and aggregating their responses. *5: People Model is an AI model developed by Dentsu that virtually reproduces high-resolution personas on a scale of 100 million people by fine-tuning large-scale survey data using LLMs. Dentsu develops new persona model that reproduces diverse consumer profiles with AI (Announced by Dentsu on April 6, 2026) https://www.dentsu.co.jp/news/release/2026/0406-011026.html *6: Development of 'AICO2,' an AI ad copy generation tool that learned the thought processes cultivated by Dentsu copywriters over many years (Announced by Dentsu on August 5, 2024) https://www.dentsu.co.jp/news/release/2024/0805-010761.html *7: About '∞AI': https://www.dentsudigital.co.jp/services/data-ai/mugen-ai https://www.dentsudigital.co.jp/ Dentsu Digital is one of Japan's largest comprehensive digital firms. With the purpose of 'Moving hearts, creating value, and changing the world,' we realize all forms of transformation by integrally utilizing creativity and technology that resonate with consumers. As a business growth partner for our clients, we aim for the 'transformation and growth' of the economy and society by co-creating new value.