Five papers from NTT Laboratories have been accepted for publication for AAAI 2026

Five papers authored by NTT laboratories have been accepted at AAAI 2026 (the 40th Annual AAAI Conference on Artificial Intelligence), to be held in Singapore, from January 20 to 27, 2026. AAAI is known as one of the most prestigious international conferences in the field of artificial intelligence, with an acceptance rate of 17.6% (23,680 papers submitted). One paper marked with (*) is selected as an oral presentation (~5% of all submissions).

Abbreviated names of the laboratories:
HI: NTT Human Informatics Laboratories
CD: NTT Computer and Data Science Laboratories
CS: NTT Communication Science Laboratories

■The Strong Lottery Ticket Hypothesis for Multi-Head Attention Mechanisms

Hikari Otsuka (Institute of Science Tokyo), Daiki Chijiwa (CD), Yasuyuki Okoshi (Institute of Science Tokyo), Daichi Fujiki (Institute of Science Tokyo), Susumu Takeuchi (CD), Masato Motomura (Institute of Science Tokyo)

In deep learning, the phenomenon that conventional neural networks contain high-accuracy, sparse subnetworks even when randomly initialized was known as the “strong lottery ticket hypothesis.” However, this remained unexplored for Transformer networks employed in recent language models. This work proves that accurate subnetworks exist within Transformer models with random initialization, based on arguments of matrix factorization for their attention mechanism, and establishes a method for extracting such subnetworks. The results contribute to theoretical understanding of Transformers, and are expected to have applications to their sparsification in future.

■Delta Matters: An Analytically Tractable Model for β-δ Discounting Agents

Yasunori Akagi (HI), Takeshi Kurashima (HI)

Humans exhibit time-inconsistent behavior, in which planned actions diverge from executed actions. Understanding the impact of such time inconsistency on behavior and designing appropriate interventions is a key research challenge in computer science and behavioral economics. Akagi, Marumo, and Kurashima (2024) focuses on progress-based tasks and succeeds in deriving a closed-form description of agent behavior, from which they obtain optimal intervention strategies. They model time-inconsistency using the β–δ discounting (quasihyperbolic discounting), but the analysis is limited to the case δ = 1. In this paper, we relax that constraint and show that a closed-form description of agent behavior remains possible for the general case 0 < δ ≤ 1. Based on this result, we derive the conditions under which gents abandon tasks and develop efficient methods for computing optimal interventions. Our analysis reveals that agent behavior and optimal interventions depend critically on the value of δ, suggesting that fixing δ = 1 in many prior studies may unduly simplify realworld decision-making processes.

■Difference Vector Equalization for Robust Fine-Tuning of Vision-Language Models

Satoshi Suzuki (HI), Shin'ya Yamaguchi (CD), Shoichiro Takeda (HI), Taiga Yamane (HI), Naoki Makishima (HI), Naotaka Kawata (HI), Mana Ihori (HI), Tomohiro Tanaka (HI), Shota Orihashi (HI), Ryo Masumura (HI)

Vision-language models such as CLIP, which are trained by contrastive loss between images and texts, exhibit strong zero-shot performance. However, it is known that task-specific fine-tuning often leads to a significant degradation in their zero-shot performance. In this work, we pointed out that this degradation arises from substantial changes in the geometric structure of image and text embeddings extracted from vision-language models. To address this issue, we propose a novel constraint to maintain a parallel-shift relationship of the geometric structure before and after fine-tuning.

■Variance Computation for Weighted Model Counting with Knowledge Compilation Approach(*)

Kengo Nakamura (CS), Masaaki Nishino (CS), Norihito Yasuda (CS)

Probabilistic inference describes relationships between data and events to predict the likelihood of events. When available data is limited, these predictions become more uncertain. An indicator of this uncertainty, the variance of inference results, has been impractical to compute due to its high computational cost. Our method efficiently computes this variance using model counting, which counts solutions to logical formulas. This enables more reliable decision-making based on probabilistic inference. For example, in equipment failure diagnosis, even when the predicted failure probability satisfies safety standards, a large variance can emphasize the need for additional data, helping to prevent failures from occurring more frequently than expected.

■FedPM: Federated Learning Using Second-order Optimization with Preconditioned Mixing of Local Parameters

Hiro Ishii (Institute of Science Tokyo), Kenta Niwa (CS), Hiroshi Sawada (CS), Akinori Fujino (CS), Noboru Harada (CS), Rio Yokota (Institute of Science Tokyo)

Federated Learning (FL) allows multiple computational resources to collaboratively train an AI model without sharing their raw data. We propose Federated Preconditioned Mixing (FedPM) to improve the accuracy of federated learning. In conventional FL approaches, even if each client uses a second-order optimization method to compute a more precise update direction, the central server simply mixes client models, which can limit the benefits of second-order optimization. FedPM overcomes this issue by splitting the second-order optimization mechanism between the server and the clients, enabling preconditioned model mixing on the server. We theoretically analyze the convergence rate of FedPM and experimentally show that FedPM achieves higher accuracy than existing methods. This work provides a solid foundation for training better AI models from data distributed across multiple sites.