Advancing AI: Highlights from September

At Sony AI, each month is a chance to share how our research, collaborations, and stories are shaping the field of artificial intelligence. September brought together music, vision, and circuits, with a look back at August’s deep dive into reinforcement learning.

Let’s dive into the highlights from September.

Exploring AI and Music at ISMIR 2025

At ISMIR 2025 in Daejeon, South Korea, our researchers presented four new projects exploring how AI can support music creators. From text-to-music editing with Instruct-MusicGen, to inference-time mastering with ITO-Master, to instrument-wise audio effects modeling, and robust music identification systems, each contribution underscored our goal of developing intuitive, flexible, and precise tools for audio creators.

Read the full ISMIR blog to explore all four projects in detail.

Pushing Analog Design Forward at MLCAD 2025
Analog circuit design is a domain that has resisted automation for decades. At MLCAD 2025, we introduced two research projects aimed at bridging this gap. GENIE-ASI generates reusable Python code for analog subcircuit identification without large training datasets, while Schemato converts complex netlists into interpretable LTSpice schematics. Together, these approaches show how AI can complement — not replace — expert engineers, embedding adaptability and interpretability into design workflows.

Dive into the full MLCAD blog for details on how GENIE-ASI and Schemato are reshaping analog design: Advancing Analog Design with AI: Sony AI’s Contributions at MLCAD 2025

Sony AI Global Events

Coming soon next month, ICCV 2025
In October, we’ll present several accepted research papers at ICCV 2025, one of the premier global gatherings for computer vision. Our work this year spans parameter-efficient fine-tuning, new approaches to video generation guidance, efficient transformers for edge devices, and novel methods for leveraging RAW sensor data in object detection.

Stay tuned for our ICCV coverage in October, where we’ll share deeper insights into each paper and its implications.

Why Reinforcement Learning Still Matters
At the end of August, we published a deep-dive blog feature on our Reinforcement Learning team, highlighting why RL remains one of the most vital areas in AI research. Unlike methods that rely on static datasets, RL agents learn by doing, adapting, failing, and improving over time. From GT Sophy to robotics, our RL researchers emphasized that RL is not the easiest path, but it’s the one that matters most for building adaptive systems.

Catch up on the RL team feature here to see why they believe the hardest AI problems are still the most important to solve: Sony AI’s Deep RL Team on Why the Hardest Problems Still Matter

Celebrating Achievements

This month, we celebrated our teams’ contributions across multiple research domains—from creative AI at ISMIR to analog automation at MLCAD. Each project reflects our shared commitment to advancing AI in ways that are practical and empowering for creators and engineers alike.

Visit our publications page to see the full list of Sony AI papers and conference contributions.

Our collaboration with Audiokinetic continued to receive media attention in sound publications.

AudioTechnology Magazine published the article, “Searching for Sound,” which discussed how Similar Sound Search combines Sony AI’s groundbreaking research with Audiokinetic’s industry-leading tools to empower audio professionals in gaming and interactive media.

Read our blog post to learn more about this collaboration:
Behind the Sound: How AI Is Enhancing Audio Search

Sony AI’s Chief Scientist, Prof. Peter Stone, recently delivered a keynote at the Emerging Technology and Generative AI Forum. His talk, “Human-in-the-Loop: Machine Learning for Robot Navigation and Manipulation,” explored how collaborative human-machine learning can accelerate robotics research.

By integrating human demonstrations, corrective interventions, and evaluative feedback, this approach opens new possibilities for safe and efficient robot navigation in healthcare and precision manipulation in manufacturing—moving beyond the limitations of traditional trial-and-error methods.

Learn More: The Emerging Technology and Generative AI Forum - Thomson Reuters Institute

Sony AI’s Yuki Mitsufuji, Lead Research Scientist and VP of AI Research, joined Wallifornia Music+Tech to share his work at the intersection of AI and creativity. As part of Sony’s Creative AI Lab and Music Foundation Model team, Yuki discussed innovations in AI-powered sound separation, restoration, and music generation. He explored his latest project, MMAudio, which synchronizes high-quality audio with video and text inputs. His talk highlighted how AI is expanding the future of sound, creativity, and human expression.

Read more about MMAudio here:
Unlocking the Future of Video-to-Audio Synthesis: Inside the MMAudio Model – Sony AI

October will be a big month as we head into ICCV 2025 with multiple accepted papers, so please be sure to follow along on our blog and social channels for additional updates and new research highlights. Connect with us on LinkedIn, Instagram, or X, and let us know what you’d like to see in future editions. Until next month, keep imagining the possibilities with Sony AI.