NTT Research’s Physics of Artificial Intelligence (PAI) Group and NTT Communication Science Laboratories (CS Labs) advance research in neuroscience, artificial intelligence, distributed learning and machine learning systems.
SUNNYVALE, Calif., July 6, 2026 — NTT, Inc., a global IT services company serving more than 75% of the Fortune Global 100 and investing billions of dollars annually in research and development, today announced that five papers have been accepted to the International Conference on Machine Learning (ICML) 2026, one of the world’s premier conferences for artificial intelligence and machine learning research. The research comes from NTT Research’s Physics of Artificial Intelligence (PAI) Group, and NTT Communication Science Laboratories (CS Labs).
Together, the accepted papers explore some of the most important questions in AI today: How do they organize knowledge? How do AI systems form beliefs? How does Agentic AI reach consensus? How can distributed learning become more efficient? And how can AI better understand noisy real-world data?
By combining insights from neuroscience and AI, Physics of AI, AI interpretability, large language models, distributed learning, and machine learning, the research advances scientific understanding of how intelligent systems learn, reason, communicate and make decisions.
“Artificial intelligence is advancing at an extraordinary pace, yet many of the principles behind intelligence remain poorly understood,” said Hidenori Tanaka, Group Leader of the Physics of Artificial Intelligence (PAI) Group at NTT Research and Director of the CBS-NTT Program in Physics of Intelligence at Harvard University. “Our goal is to better understand how intelligence works. These accepted papers represent important steps toward advancing the science of AI and helping researchers better understand intelligence itself.”
How Do AI Systems Organize Knowledge?
Emergence of Hierarchical Emotion Organization in Large Language Models
Can large language models naturally organize knowledge the way humans do? This research shows that LLMs develop hierarchical emotional representations that resemble structures found in psychology and neuroscience, advancing AI interpretability and providing new insights into how intelligence organizes concepts.
Authors
- Maya Okawa — CBS-NTT Program in Physics of Intelligence, Harvard University; Physics of Artificial Intelligence (PAI) Group, NTT Research
- Bo Zhao — CBS-NTT Program in Physics of Intelligence, Harvard University; University of California, San Diego
- Eric J. Bigelow — CBS-NTT Program in Physics of Intelligence, Harvard University; Physics of Artificial Intelligence (PAI) Group, NTT Research; Department of Psychology, Harvard University
- Rose Yu — University of California, San Diego
- Tomer Ullman — Department of Psychology, Harvard University
- Ekdeep Singh Lubana — CBS-NTT Program in Physics of Intelligence, Harvard University; Physics of Artificial Intelligence (PAI) Group, NTT Research
- Hidenori Tanaka — CBS-NTT Program in Physics of Intelligence, Harvard University; Physics of Artificial Intelligence (PAI) Group, NTT Research
How Do AI Systems Form and Update Beliefs?
Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering
Understanding how AI systems acquire and modify beliefs is central to explaining AI reasoning. This research reveals how in-context learning and activation steering change internal model representations, advancing the Physics of AI and improving scientific understanding of how intelligent systems learn and make decisions.
Authors
- Eric Bigelow — Good fire AI; Department of Psychology, Harvard University; Physics of Artificial Intelligence (PAI) Group, NTT Research
- Daniel Wurgaft —Goodfire AI; Department of Psychology, Stanford University
- YingQiao Wang —Department of Psychology, Harvard University
- Noah Goodman —Department of Psychology, Stanford University; Department of Computer Science, Stanford University
- Tomer Ullman — Department of Psychology, Harvard University; Center of Brain Science, Harvard University
- Hidenori Tanaka — Physics of Artificial Intelligence (PAI) Group , NTT Research; Center of Brain Science, Harvard University
- Ekdeep Singh Lubana — Goodfire AI
What is multi-agent AI bias?
Emergence of Biased Consensus in Multi-Agent LLM Debates
As AI agents increasingly collaborate, understanding collective reasoning becomes more important. This paper examines how interacting large language models can develop biased consensus during debate, contributing new insights into multi-agent systems, collaborative AI and AI bias detection.
Author
- Maya Okawa — CBS-NTT Program in Physics of Intelligence, Harvard University; Physics of Artificial Intelligence (PAI) Group, NTT Research
How Can Distributed Learning Make AI More Efficient?
Bottleneck Communication Delay Minimization for Communication-Efficient Decentralized Learning
Training modern AI models requires communication across distributed computing systems. This paper introduces a method for reducing communication bottlenecks in distributed learning, enabling more efficient decentralized AI training and improving the scalability of next-generation machine learning systems.
Authors
- Nozomi Hata — NTT Communication Science Laboratories
- Kenta Niwa — NTT Communication Science Laboratories / NTT Computer and Data Science Laboratories
How Can AI Better Understand Noisy Real-World Data?
Joint Enhancement and Classification using Coupled Diffusion Models of Signals and Logits
AI systems frequently encounter noisy speech and other imperfect real-world signals. This research introduces a coupled diffusion model that jointly enhances signals and improves classification, advancing machine learning, diffusion models, and robust AI perception without requiring additional model retraining.
Authors
- Gilad Nurko — Technion – Israel Institute of Technology
- Roi Benita — Technion – Israel Institute of Technology
- Yehoshua Dissen — Technion – Israel Institute of Technology
- Tomohiro Nakatani — NTT Communication Science Laboratories
- Marc Delcroix — NTT Communication Science Laboratories
- Shoko Araki — NTT Communication Science Laboratories
- Joseph Keshet — Technion – Israel Institute of Technology
About the Physics of Artificial Intelligence (PAI) Group
Through research in Physics of AI, Neuroscience and AI, and AI Interpretability, the NTT Research Physics of AI (PAI) Group better understands how intelligence works. The PAI Group advances the science of AI and builds the foundation for its future.
The Lab’s research explores some of the most important questions in artificial intelligence today: How do intelligent systems learn? How do they reason? How do beliefs emerge? And how can understanding intelligence help shape the future of AI?
About NTT Communication Science Laboratories (CS Labs)
These laboratories aim to create the novel concepts of knowledge and information communication needed for making the human-friendly humanoid computer a reality , such as information processing and media processing related to human knowledge and emotions.
About ICML 2026
The International Conference on Machine Learning (ICML) is one of the world’s leading conferences dedicated to advances in machine learning and artificial intelligence. Researchers from academia and industry gather annually to present breakthroughs in machine learning theory, foundation models, AI systems, interpretability, optimization, reinforcement learning and emerging AI applications.
About NTT Research
NTT Research is the Silicon Valley research arm of NTT, one of the world’s largest technology and business solutions providers. Founded in 2019, NTT Research invents the future of foundational science while accelerating its real-world impact across NTT’s global ecosystem.
From its headquarters in Sunnyvale, California, NTT Research brings together world-class scientists across four research pillars: the Physics & Informatics (PHI) Lab, the Cryptography & Information Security (CIS) Lab, the Medical & Health Informatics (MEI) Lab, and the Physics of Artificial Intelligence (PAI) Group. Their work advances fields that define the next era of optical computing, next-generation cryptography, biodigital twins to enable precision medicine and the physics of AI to understand how intelligence works.
As part of NTT Inc., a global enterprise with more than $90 billion in annual revenue, serving 75% of the Fortune Global 100 and investing billions of dollars annually in research and development, NTT Research is uniquely positioned to carry deep science from the lab to global-scale deployment.
Through its annual Upgrade conference and technology incubator, Scale Academy, NTT Research accelerates the path from discovery to application, transforming fundamental research into technologies that solve real-world problems across industries.
###
NTT and the NTT logo are registered trademarks or trademarks of NIPPON TELEGRAPH AND TELEPHONE CORPORATION and/or its affiliates. All other referenced product names are trademarks of their respective owners. ©2026 NIPPON TELEGRAPH AND TELEPHONE CORPORATION
| NTT Research Contact: Chris Shaw Chief Marketing Officer NTT Research +1-312-888-5412 chris.shaw@ntt-research.com | Media Contact: Cara Milan AMP Marketing and Public Relations For NTT Research +1- 415-792-2968 cara@amppublicrelations.com |