Samsung’s Bold Move from Foundation Models to Physical AI: Leveraging ROM‑Based Architecture and New Benchmarks

Background

Competition for large models is shifting as AI moves from purely digital services to real devices and physical environments. Samsung has acquired tens of thousands of GPUs for AI infrastructure and leverages its high‑bandwidth memory (HBM) expertise, giving it a hardware advantage for large‑scale training.

Memory‑Based Architecture (Meki)

The paper "Memory‑Based Architecture" introduces Meki, a large‑model design for resource‑constrained edge devices. It parameterises a Memory Bank that resides in ROM, while RAM handles real‑time computation. By separating knowledge storage (ROM) from inference compute (RAM), the model can exceed the on‑device RAM limits that traditional Transformers face.

Compared with conventional Transformers that keep all parameters in RAM, Meki is better suited for long‑running edge AI and Physical AI scenarios such as robots, where low latency, continuous interaction, and sustained online operation are required.

Paper: https://arxiv.org/pdf/2602.03359

Project: https://github.com/ningding-o/MeKi

Multi‑Domain Reinforcement Learning (M2RL)

M2RL explores a training paradigm that jointly evolves reinforcement‑learning capabilities across multiple domains (e.g., mathematics, code, scientific inference). Experiments show that high‑density reasoning tasks do not conflict; instead they produce synergistic enhancements, indicating that reinforcement learning can drive foundation models toward stronger reasoning, planning, and complex‑environment interaction.

Paper: https://arxiv.org/pdf/2602.12566

LiveClawBench

LiveClawBench proposes a three‑dimensional complexity framework for evaluating models in physical settings: Environment Complexity, Cognitive Demand, and Runtime Adaptability. The framework emphasizes real‑world task chains, multi‑environment coordination, and long‑term task execution, addressing the limitation of static, single‑environment benchmarks that are insufficient for AI deployed in devices, robots, and other Physical AI scenarios.

Paper: https://arxiv.org/abs/2604.13072

Organizational context

These works are produced by Samsung’s AI Model TF task force, established in 2025 under the group CTO. The task force focuses on building advanced foundation‑model capabilities and advancing Physical AI across Samsung’s global device ecosystem.