TLDR: NTT has unveiled ‘Portable Tuning,’ a groundbreaking learning framework designed to drastically reduce the maintenance costs of customized generative AI. This technology eliminates the need to retrain specialized AI models every time their underlying foundation models are updated or changed, by enabling the transfer of acquired knowledge across diverse foundation models.
TOKYO – July 9, 2025 – NTT, Inc. today announced the establishment of Portable Tuning, a novel learning framework poised to transform the landscape of generative AI maintenance. This innovative technology addresses a critical challenge faced by companies and organizations: the escalating costs associated with maintaining specialized AI models that are built upon frequently updated foundation models.
Generative AI models, while powerful, often require fine-tuning on specific datasets to achieve high accuracy for particular tasks or domains. However, as foundation models undergo regular updates to incorporate new knowledge or architectural changes, these specialized models traditionally demand costly and resource-intensive retraining. This process consumes substantial computational resources and human effort for hyperparameter tuning with each update.
NTT had previously introduced the concept of learning transfer to reuse existing training results across various foundation models. While a promising principle, earlier iterations faced limitations, including the need for additional training to achieve practical accuracy and an inability to apply the method when foundation model architectures differed.
Portable Tuning directly tackles these challenges by reinterpreting conventional fine-tuning through the lens of reward learning. The core innovation lies in training an independent reward model that corrects the outputs of the foundation model for specific tasks or domains. Crucially, this reward model operates separately from the foundation model itself, allowing it to be reused seamlessly with new or updated foundation models without requiring any additional retraining. Furthermore, because the reward model’s functionality depends only on the output format (e.g., classification labels, token vocabulary) and not on the internal architecture of the foundation model, it can be applied even when source and target foundation models have different designs.
While there is an initial overhead associated with training and inference for the reward model, Portable Tuning ensures that subsequent maintenance costs remain constant, regardless of how many times the foundation model is updated. This represents a fundamental shift from variable, high retraining costs to a predictable, low-cost maintenance model.
Experiments conducted by NTT have demonstrated the efficacy of Portable Tuning. The technology successfully transferred domain-specific fine-tuning for vision foundation models and instruction tuning for language foundation models to other foundation models of varying sizes and pretraining datasets. These transfers achieved performance comparable to models that were individually fine-tuned for each foundation model, validating the approach’s practical applicability.
This research will be formally presented at the prestigious International Conference on Machine Learning (ICML 2025), scheduled to take place in Vancouver, Canada, from July 13 to July 19, 2025. The paper, titled Portable Reward Tuning: Towards Reusable Fine-Tuning across Different Pretrained Models, is authored by Daiki Chijiwa, Susumu Takeuchi, Taku Hasegawa, Kyosuke Nishida, and Kuniko Saito from NTT Computer and Data Science Laboratories and NTT Human Informatics Laboratories.
Also Read:
- UNESCO and UCL Report Unveils Strategies to Slash AI Energy Consumption by Up to 90%
- Japanese Study Reveals Retrieval-Augmented AI Eliminates Hallucinations in Clinical LLMs, Bolstering Patient Privacy and Accuracy
Looking ahead, NTT believes Portable Tuning will not only significantly reduce operational costs for businesses but also open new avenues, such as simulating the expected effects of retraining in advance. The company reaffirms its commitment to advancing next-generation AI technologies, particularly those aimed at mitigating the growing cost challenges associated with large-scale AI and realizing the vision of AI Constellations, where multiple AI systems operate in coordinated harmony.
