SqueezeBits' 'Gra-Rolla' Fine-Tuning Tech Recognized at NeuRIPS Spotlight

## A New Horizon for AI Lightweighting: SqueezeBits' 'GraLoRA' Achieves NeurIPS Spotlight Selection SqueezeBits, a leader in AI lightweighting technology, has achieved brilliant results at the forefront of artificial intelligence research. The innovative AI lightweighting technology ‘GraLoRA’, co-developed by SqueezeBits' ML Researcher Jung Yeon-jun and Tech Lead Ahn Dae-hyun, along with Professor Park Eun-hyuk's research team at POSTECH, has achieved the remarkable feat of being selected as a ‘Spotlight’ paper at NeurIPS 2025, the most prestigious academic conference in the field of artificial intelligence. This selection is given to only the top approximately 3.2% of all submitted papers. This signifies international recognition for SqueezeBits' original technological prowess and is regarded as setting a new milestone in the field of AI lightweighting. The newly unveiled GraLoRA is a technology that overcomes the limitations of the 'Fine-Tuning' process, which optimizes large-scale AI models for specific uses. It is particularly noteworthy for maintaining the excellent cost efficiency provided by the existing 'LoRA' technique while innovatively resolving the performance instability issues frequently encountered in actual service environments. GraLoRA achieved this innovation through a unique approach that meticulously reconfigures learning units, consequently delivering significantly more stable and predictable performance compared to existing methods, even under identical resource conditions. This will serve as a crucial factor in maximizing the practical efficiency and reliability of AI model operation. The advent of GraLoRA is expected to accelerate the spread of the 'personalized AI' era. It is poised to become a core technology that expands the horizons of customized AI services, from providing personalized learning experiences for smartphone AI assistants to operating AI models that satisfy both security and cost-efficiency in enterprise environments. By enabling the construction of more sophisticated and stable personalized AI based on user data, the practical utilization value of AI technology is expected to further increase. Furthermore, GraLoRA achieved the remarkable feat of being fully integrated into the official library of Hugging Face, the heart of the AI open-source community. This means that countless developers worldwide can immediately utilize GraLoRA technology without complex implementation processes, broadly applying it to various experiments and even actual service deployments. The official incorporation of SqueezeBits' technology into the global AI ecosystem is strong evidence of its impact and versatility. A SqueezeBits official commented on the NeurIPS Spotlight selection, stating, "This is the result of international recognition for GraLoRA's unique technological prowess and high utility in real industrial settings," and expressed their ambition, "We will continue to dedicate ourselves to leading research and development for the implementation of efficient yet powerful AI services, thereby setting new standards for AI technology." Building on this achievement, SqueezeBits is expected to continue innovating in AI lightweighting technology, bringing positive ripple effects to the global AI market.