Samsung Research is developing a new model, LittleBit-2, that maximizes spectral energy gain in sub-1-bit Large Language Models to improve the efficiency of artificial intelligence. This approach differs from current trends focused on increasing model size and precision, instead reducing the computational bits required. The company detailed its work on “LittleBit” in a blog post published on November 24, 2025. Samsung R&D Institute UK received a Best Paper Award at the ICCV 2025 Personalization in Generative AI Workshop.
LittleBit & NanoQuant: Ultra-Low-Bit Quantization of Large Language Models
Reducing computational bits allows for the deployment of sophisticated AI on a wider range of devices, even those with limited processing power, and represents a shift in Large Language Model development. Further refining this low-bit quantization technique is “NanoQuant”, an approach enabling efficient sub-1-bit quantization of large language models, demonstrating Samsung’s commitment to model compression. Unlike traditional quantization methods that reduce precision to 8-bit or 4-bit, NanoQuant explores representing model weights and activations with even fewer bits, potentially down to a single bit.
This aggressive quantization requires innovative techniques to mitigate information loss and maintain acceptable performance levels, and the team reports substantial gains in efficiency without significant accuracy degradation. A company statement confirms, “With Samsung’s Unique Strengths, We Are Developing a User-Oriented AI Algorithm,” highlighting the practical application of these advancements. Ultra-low-bit quantization unlocks new possibilities for on-device AI processing by minimizing the memory footprint and energy consumption of LLMs, Samsung aims to enable real-time AI applications on smartphones, wearables, and other edge devices without relying on cloud connectivity.
With Samsung’s Unique Strengths, We Are Developing a User-Oriented AI Algorithm.
Cross-Biosignal & Cross-Modal AI: Advancing Wearable Health Insights
Samsung Research is expanding the capabilities of wearable health technology by integrating artificial intelligence that analyzes data from multiple biosensors, as detailed in several blog posts throughout 2025 and 2026. The company’s work focuses on cross-biosignal pretraining, enabling AI to learn relationships between different bodily signals and improve the accuracy of health assessments, Samsung Research says. This is achieved by allowing one sensor’s data to inform the interpretation of another, enhancing the overall signal quality and predictive power of wearable devices.
A key element of this advancement is the development of cross-modal AI, which combines data from biosensors with other modalities like audio and video, allowing for a richer, more nuanced analysis of health indicators and potentially detecting subtle changes that might be missed by traditional methods. The team’s work on “When One Sensor Learns Another: Cross-Modal AI for Wearables” exemplifies this focus on synergistic data analysis.
AI-Powered Network Management & 6G Innovations for Future RAN
Samsung Research is refining radio access network (RAN) management through artificial intelligence, demonstrated by its development of AI-driven modem innovations intended for future cellular networks. These advancements extend beyond incremental improvements, with the company focusing on solutions for automated network performance evaluation using a GenAI-based digital twin, dubbed “Network GDT”. This approach allows for proactive identification and resolution of potential network issues before they impact users, a critical capability as 6G technologies demand increasingly complex and dynamic network configurations.
Furthering this focus on 6G, Samsung Research has explored large-scale AI-based compression of channel state information (CSI), addressing the significant data transmission overhead associated with advanced 6G features, particularly in the millimeter wave frequency range, according to the company. This compression technique is designed to reduce the bandwidth required for transmitting CSI, enabling more efficient use of network resources and supporting higher data rates.
Simultaneously, researchers are tackling the challenges of on-device large language model (LLM) deployment, recognizing the need for efficient AI processing closer to the user. “Efficient Compositional Multi-tasking for On-device Large Language Models” represents an effort to optimize LLMs for resource-constrained devices, allowing for more sophisticated AI applications directly on smartphones and other mobile platforms. The pursuit of model efficiency is particularly evident in the “LittleBit” project, which maximizes spectral energy gain in sub-1-bit LLMs, prioritizing computational efficiency through extreme quantization over increasing LLM size and precision.
Samsung Research also introduced “NanoQuant”, an efficient sub-1-bit quantization technique, and “RaBiT”, a residual-aware binarization training method, both aimed at reducing the memory footprint and computational demands of LLMs without significant performance degradation. Beyond network infrastructure and model optimization, Samsung Research is exploring cross-modal AI for wearable devices.
Their work, described as “When One Sensor Learns Another”, focuses on leveraging data from multiple sensors to improve the accuracy and robustness of health monitoring applications, allowing the system to compensate for noisy or incomplete data from individual sensors and providing a more reliable assessment of a user’s health status. “From AI Concept to Real World: AI-powered Modem for Next Samsung Cellular Networks” encapsulates this commitment to translating research into tangible products and services, shaping the future of connectivity and intelligent devices.
PIX-TAB & GalaxyEdit: AI for Advanced Image and Video Understanding
Samsung Research is refining its capabilities in visual data understanding with PIX-TAB, an artificial intelligence approach designed for pixel-precise table structure recognition. This system utilizes speculative decoding and region-based image segmentation to accurately extract data from visual documents, with implications for automating tasks like invoice processing and data entry in business settings.
The architecture moves beyond simple object detection to pinpoint individual cells and their relationships within a table, increasing the reliability of automated data capture. Further expanding its image and video processing expertise, Samsung Research introduced GalaxyEdit, a large-scale image editing dataset coupled with an enhanced diffusion adapter, aiming to improve the performance of generative AI models in image manipulation tasks and allowing for more realistic and nuanced edits.
The development of GalaxyEdit addresses a critical need for high-quality training data in the rapidly evolving field of image generation and editing, where dataset quality directly impacts the fidelity of the results. Researchers are leveraging this resource to refine diffusion models, a class of generative AI known for producing high-resolution images.
Beyond static images, Samsung Research is also tackling challenges in video understanding with tools like DeMo-Pose, a system for depth-monocular modality fusion for object pose estimation, combining information from depth sensors and standard cameras to accurately determine the 3D orientation of objects within a video stream. The team also developed LookaheadKV, a technique for fast and accurate KV cache eviction, improving the efficiency of video processing without sacrificing quality.
Researchers also introduced RaBiT, a residual-aware binarization training method for accurate and efficient LLMs.
On-Device AI: Efficient Models for Mobile Applications & Speech Processing
Samsung Research is redefining on-device artificial intelligence through innovations in model compression and specialized architectures, prioritizing efficiency for mobile applications over simply scaling up model size. This technique suggests a fundamental shift in how LLMs are designed for resource-constrained environments, potentially enabling complex AI tasks on smartphones and wearables without relying on cloud connectivity. Beyond LLMs, Samsung Research is tackling challenges in diverse areas like image processing and audio restoration with equally innovative approaches.
Samsung R&D Institute UK received recognition for its contributions, winning a Best Paper Award at the ICCV 2025 Personalization in Generative AI Workshop, solidifying the company’s position as a leader in the field, the company says. The breadth of these projects, from health-focused AI to 6G network management, underscores a strategic vision for embedding intelligence into every aspect of the connected world.
See today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals.
