NTT’s Landmark AI Research Showcased at Leading Machine Learning Conference

At the recent 42nd International Conference on Machine Learning (ICML 2025) held in Vancouver from July 13–19, researchers from NTT Research and NTT R&D—core divisions of NTT—presented a total of twelve groundbreaking papers. ICML is recognized globally for its focus on advancing machine learning, a critical subset of artificial intelligence, with applications spanning machine vision, computational biology, speech recognition, and robotics

Deepening Understanding of Large Language Models

Three standout papers stemmed from NTT Research’s Physics of Artificial Intelligence (PAI) Group, each contributing fresh insights into the behavior and capabilities of AI systems. In the first, titled “Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing,” researchers investigated a key challenge: why efforts to edit or update factual knowledge in large language models (LLMs) often lead to unexpected declines in their accuracy and reasoning power. The team discovered a phenomenon they dubbed “representation shattering,” where modifying a single piece of information inadvertently distorts the model’s broader understanding, affecting not just one entity but related ones as well. This discovery has major implications for the future development and maintenance of reliable AI systems

The second paper, “Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models,” tackled a persistent issue in machine learning interpretability: the instability of Sparse Autoencoders (SAEs), which are used to make complex models more understandable. By introducing a new approach called Archetypal SAEs, the researchers demonstrated a significant leap in the stability and reliability of these tools, paving the way for more robust interpretability methods in deep learning

Uncovering the Secrets of Short-Term Memory in Neural Networks

In the third paper, “Dynamical Phases of Short-Term Memory Mechanisms in RNNs,” NTT’s team took a closer look at the mechanisms underlying short-term memory in recurrent neural networks (RNNs)—an area that remains poorly understood. Their findings not only provide new insights into how these networks process information but also offer testable predictions that could guide future research in systems neuroscience

A Broad Portfolio of Innovative Research

Beyond these three, scientists from NTT R&D’s laboratories in Japan shared eight additional papers covering a diverse range of topics. Among these was “Portable Reward Tuning: Towards Reusable Fine-Tuning Across Different Pretrained Models,” which unveiled a technique that eliminates the need to retrain custom AI models every time the foundational model is updated. This innovation promises to reduce costs and enhance the sustainability of generative AI applications

Another significant contribution, “Plausible Token Amplification for Improving Accuracy of Differentially Private In-Context Learning Based on Implicit Bayesian Inference,” explored the impact of privacy-preserving noise on the accuracy of LLMs—a crucial concern for fields like healthcare, government, and finance. The findings clarify the relationship between privacy protection and model performance, opening the door to broader, more responsible deployment of AI in sensitive domains

The paper “K2IE: Kernel Method-Based Kernel Intensity Estimators for Inhomogeneous Poisson Processes” demonstrated a major breakthrough in analyzing and forecasting complex event patterns, whether they involve social media activity, disease outbreaks, or other large-scale data phenomena

Commitment to Ethical, Sustainable AI

Dr. Hidenori Tanaka, leader of the PAI Group, emphasized NTT’s dedication to developing AI technologies that foster sustainable development, safeguard human autonomy, ensure fairness and transparency, and protect security and privacy. He stressed that achieving these goals hinges on groundbreaking scientific exploration at the most fundamental levels of AI and machine learning—a pursuit at the heart of NTT’s research initiatives

Expanding NTT’s Research Footprint

Established in 2019 in Silicon Valley, NTT Research currently houses four specialized groups focused on quantum information, cryptography and cybersecurity, medical informatics, and artificial intelligence. The organization is committed to pioneering innovations that empower communities and businesses worldwide, guided by a vision of trust, integrity, and technological excellence 

For further details on NTT’s AI research and initiatives, please visit their official website