Five papers from the NTT group have been accepted for publication for ICML 2026

Two papers from NTT and three papers from NTT Research, Inc.(NTT Research) have been accepted at ICML (International Conference on Machine Learning) 2026 held in Seoul, Korea, from 6th to 11th July 2026. ICML is one of the most prestigious international conferences in the field of machine learning, a core technology that underpins modern artificial intelligence. It serves as a leading venue for presenting cutting-edge research on the theoretical foundations, algorithms, and applications of machine learning. ICML is known as one of the most competitive conferences in the research field, and the acceptance rate this year is 26.6% (6,352 papers accepted among 23,918 submissions).

< Major Accepted Papers by NTT Laboratories>

◆Bottleneck communication delay minimization for communication-efficient decentralized learning

Nozomi Hata (CS), Kenta Niwa (CS)

Decentralized learning of AI models, such as image classification models and language processing models, have been actively studied to efficiently utilize a large number of computing resources accumulated in e.g., data centers. Previous studies have explored network topologies that connect computers sparsely while enabling the trained AI model to approximate a fully averaged model as closely as possible. However, when considering decentralized learning of AI models over real-world data center networks, communication delays vary across different paths. As a result, the larger the maximum communication delay (bottleneck communication delay), becomes, the fewer model updates can be performed, thereby reducing the efficiency of decentralized learning.

In this study, we proposed a novel node assignment algorithm, called BTSP-MSR, that approximately minimizes the bottleneck communication delay under given communication delays between computers. Focusing on the fact that efficient static and dynamic graphs used in decentralized learning often have circulant digraph forms, we found that an upper bound of the node assignment problem, which requires exponential time with respect to the number of nodes, can be decomposed into two more tractable problems: the BTSP problem and the MSR minimization problem. This decomposition enabled us to approximately solve the previously computationally challenging node assignment problem in only a few seconds. Furthermore, through experiments using communication delays that simulate real-world data center networks and multiple cyclic directed graphs, we confirmed that the proposed method reduces bottleneck communication delay and improves the efficiency of decentralized learning.

◆Joint Enhancement and Classification using Coupled Diffusion Models of Signals and Logits

Gilad Nurko (Technion), Roi Benita (Technion), Yehoshua Dissen (Technion), Tomohiro Nakatani (CS), Marc Delcroix (CS), Shoko Araki (CS), Joseph Keshet (Technion)

This study proposes a method that integrates two diffusion models, one for signal enhancement and the other for shaping classifier logits, to improve classification accuracy in noisy environments. By allowing the two models to exchange information, the framework simultaneously enhances signal reconstruction and classification performance. Experiments on image classification and automatic speech recognition demonstrate that the proposed approach is more robust to noise and achieves higher classification accuracy than conventional sequential integration methods. Since the method does not require retraining the classifier, it is expected to be applicable to existing classifiers trained on large-scale datasets and potentially improve their performance.

< Accepted Papers from NTT Research >

◆Emergence of Hierarchical Emotion Organization in Large Language Models

Maya Okawa (Harvard University, PAI), Bo Zhao (Harvard University, University of California, San Diego), Eric Bigelow (Harvard University, PAI), Rose Yu (University of California, San Diego), Tomer Ullman (Harvard University), Ekdeep Singh Lubana (Harvard University, PAI), Hidenori Tanaka (Harvard University, PAI)

◆Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering

Eric Bigelow (Harvard University, PAI), Daniel Wurgaft (Goodfire AI, Stanford University), YingQiao Wang (Harvard University), Hidenori Tanaka (Harvard University, PAI), Tomer Ullman (Harvard University), Noah Goodman (Stanford University), Ekdeep Singh Lubana (Harvard University, PAI)

◆Emergence of Biased Consensus in Multi-Agent LLM Debates

Maya Okawa (Harvard University, PAI)

Abbreviated names of the laboratories:
CS: NTT Communication Science Laboratories
PAI: Physics of Artificial Intelligence Group