NTT Research, NTT, Inc., and NTT DATA scientists contributed fifteen research papers to the Conference on Neural Information Processing Systems (NeurIPS) 2025, showcasing advances in foundational theory, system-level development, and enterprise-grade innovations. These contributions, presented December 2-7 at the San Diego Convention Center, address critical areas including understanding model behavior, improving model efficiency—with one activation probe demonstrating a reduction in compute by six orders of magnitude—and securing AI through robust watermarking solutions. This work reflects a focus on both fundamental research and operational challenges within the rapidly evolving field of machine learning and computational neuroscience.
Foundational Research in Model Behavior
NTT scientists contributed fifteen research papers to NeurIPS 2025, with a significant focus on foundational research into model behavior. Eight papers originated from NTT Research, largely addressing these fundamental issues. Specifically, five papers explore how models “think,” investigating the emergence of personality traits like kindness or sycophancy when trained on synthetic games. This work introduces controlled environments designed to probe and ultimately shape these personality characteristics within large language models.
A key area of foundational exploration centers on understanding model generalization and memorization. One paper provides a predictive framework for determining when models will reliably generalize versus simply memorize information during in-context learning. Additionally, research was conducted on inference-time alignment methods utilizing importance sampling, allowing models to adopt aligned behaviors without requiring retraining—a crucial step for practical applications.
Further foundational research includes advances in interpretability and representation learning. Researchers presented “Neural Thermodynamics,” offering a framework for understanding entropy’s role in feature formation within deep networks. Another paper proposed a new Sparse Autoencoder (SAE) leveraging Matching Pursuit to capture more complex, hierarchical features than standard SAEs, revealing implicit geometric assumptions embedded within these architectures and shaping concept discovery.
Advances in Efficient and Distributed AI Systems
NTT scientists presented fifteen papers at NeurIPS 2025, showcasing advances across foundational research, system-level improvements, and enterprise solutions. Several papers focused on efficient and distributed AI systems. Specifically, researchers presented “LLM capable of 1-GPU inference: tsuzumi,” a lightweight, high-efficiency LLM designed to run on a single GPU, broadening accessibility and deployment possibilities. Another contribution, “Revisiting 1-peer exponential graph for enhancing decentralized learning efficiency,” analyzes how simplified communication patterns can significantly improve convergence in decentralized learning scenarios.
Furthering efficient systems, researchers explored methods to reduce computational costs. One paper details an “activation probe” capable of detecting risky model behavior with six orders of magnitude less compute compared to standard monitoring techniques. This highlights a move towards lightweight analysis during deployment. Additionally, work on “Inference-time alignment of language models” introduces an approach to align model behaviors without retraining, potentially saving significant resources.
The NTT contributions also included advancements in understanding model behavior and representations, relevant to building more efficient systems. Research explored how models “think,” examining the emergence of personality traits and developing predictive frameworks for generalization versus memorization. Theoretical work also addressed how reasoning performance scales with model size, offering guiding principles for building efficient reasoning systems, furthering the goal of optimized AI.
AI is becoming ubiquitous, but how these computational engines actually work remains-to a surprising degree-a mystery, which is why our scientists keep probing with fundamental questions.
Hidenori Tanaka
Applications of AI to Sensing and Imaging
NTT researchers presented fifteen papers at NeurIPS 2025, with several focusing on applications to sensing and imaging. Specifically, “Transformer Enabled Dual-Comb Ghost Imaging for Optical Fiber Sensing” demonstrates how transformers can improve reconstructions for fiber-optic sensing, increasing spatial resolution and signal robustness. This work bridges physics-based modeling with machine learning optimization, suggesting a pathway for more effective physical interaction modeling using AI techniques.
Several papers detailed advancements in how models “see” and interpret information. “Enhancing Visual Prompting through Expanded Transformation Space and Overfitting Mitigation” proposes expanded geometric and photometric transformations for visual prompting, alongside methods to reduce overfitting, leading to improved robustness and accuracy. This suggests AI systems can be made more reliable in processing visual information by refining how they handle transformations and avoiding overfitting.
Beyond specific sensing applications, research explored how AI models themselves can be understood and improved. “Detecting High-Stakes Interactions with Activation Probes” introduces a method using lightweight activation probes that require six orders of magnitude less compute than standard monitors, enabling the detection of risky model behavior during deployment. This focus on model interpretability is critical for building trustworthy AI systems applicable to sensitive sensing and imaging scenarios.
AI Security, Watermarking and Trustworthiness
NTT scientists are addressing AI security and trustworthiness, highlighted by a paper from NTT DATA, J.P. Morgan, and UCLA titled “Breaking Distortion-free Watermarks in Large Language Models.” This research revealed vulnerabilities in current watermarking approaches, demonstrating the potential to recover secret keys and generate text that bypasses detection. Shayleen Reynolds of NTT DATA emphasized the urgency of this work given the EU AI Act’s mandate for watermarking all AI-generated content, signifying a need for robust solutions.
The research extends beyond simply identifying weaknesses; it functions as a “wake-up call” regarding existing vulnerabilities in AI watermarking techniques. This is particularly critical as enterprise customers demand trustworthy AI solutions. NTT DATA’s contribution falls under the themes of security, provenance, and trustworthiness – areas of keen concern for its client base and essential for wider AI adoption.
Furthermore, NTT Research is exploring methods to detect risky model behavior with lightweight “activation probes.” These probes achieve detection with six orders of magnitude less compute than standard monitoring techniques. This work, alongside advances in sparse autoencoders (SAEs) and understanding model personalities, contributes to a broader effort to understand and shape AI behavior, ultimately enhancing both security and trustworthiness.
