Low-rank adaptation (LoRA) has become the dominant method for parameter-efficient LLM fine-tuning, with LoRA-based quantization error compensation (LQEC) emerging as a powerful tool for recovering accuracy in compressed LLMs. However, LQEC has underperformed in sub-4-bit scenarios, with no prior investigation into understanding this limitation. We propose RILQ (Rank-Insensitive LoRA-based Quantization Error Compensation) to boost 2-bit LLM accuracy. Based on rank analysis revealing model-wise activation discrepancy loss's rank-insensitive nature, RILQ employs this loss to adjust adapters cooperatively across layers, enabling robust error compensation with low-rank adapters. Evaluations on LLaMA-2 and LLaMA-3 demonstrate RILQ's consistent improvements in 2-bit quantized inference across various state-of-the-art quantizers and enhanced accuracy in task-specific fine-tuning. RILQ maintains computational efficiency comparable to existing LoRA methods, enabling adapter-merged weight-quantized LLM inference with significantly enhanced accuracy, making it a promising approach for boosting 2-bit LLM performance.

In this paper, we propose a 1st person view and body motion based control method realized on a smart phone/pad called the Sprint (Smart Phone/Pad and Robot for Teleoperation and Tele-presence). The smart phone/pad is chosen, aside from its ubiquity and wide usage, for its mobility and rich media (sensor/display) capabilities. The key idea of Sprint is to have the operator control and interact by "posing" as the robot in the remote area in hopes of eliciting a high sense of tele-presence and thereby promoting the spatial task performance and the overall interactive experience. In particular, we demonstrate a remote robot viewing control method based on this design principle and validate that the proposed approach does improve the sense of tele-presence and spatial understanding.

In this work, we present a technique based on image saliency analysis to improve the conspicuity of the foreground augmentation to the background real world medium by adjusting the local brightness contrast. The proposed technique is implemented on a mobile platform considering the usage nature of AR. The saliency computation is carried out for the augmented object's representative color rather than all the pixels, and searching and adjusting over only a discrete number of brightness levels to produce the highest contrast saliency, thereby making near-real time computation possible. Thus, while the resulting imagery may not be optimal due to such a simplification, our tests showed that the visibility was still significantly improved without much difference to the "optimal" ground truth in terms of correctly perceiving and recognizing the augmented information