RPIQ: Residual-Projected Multi-Collaboration Closed-Loop and Single Instance Quantization for Visually Impaired Assistance

TL;DR

RPIQ tackles the memory and inference bottlenecks of deploying large-scale models for visually impaired assistance by introducing a 4-bit quantization framework that combines block-wise residual compensation, single-instance Hessian-curvature calibration, and Gauss-Seidel dynamic updates. The method preserves global second-order statistics while enabling multi-round refinement with in-memory, last-batch calibration data, achieving substantial memory savings (~60–75%) and modest time overhead while maintaining or improving performance on language and multimodal tasks. Across OPT, Qwen, LLaMA, and CogVLM2-19B, RPIQ delivers strong sentiment, language understanding, and OCR-VQA performance, and demonstrates practical deployment feasibility on consumer hardware (e.g., RTX 4090). The work highlights a practical pathway to deploy advanced language and vision-language models in assistive devices, balancing accuracy, efficiency, and privacy by avoiding repeated calibration-data loading. Limitations include sensitivity to iteration count and initialization, suggesting future work on dynamic calibration and broader multimodal extensions.

Abstract

Visually impaired users face significant challenges in daily information access and real-time environmental perception, and there is an urgent need for intelligent assistive systems with accurate recognition capabilities. Although large-scale models provide effective solutions for perception and reasoning, their practical deployment on assistive devices is severely constrained by excessive memory consumption and high inference costs. Moreover, existing quantization strategies often ignore inter-block error accumulation, leading to degraded model stability. To address these challenges, this study proposes a novel quantization framework -- Residual-Projected Multi-Collaboration Closed-Loop and Single Instance Quantization(RPIQ), whose quantization process adopts a multi-collaborative closed-loop compensation scheme based on Single Instance Calibration and Gauss-Seidel Iterative Quantization. Experiments on various types of large-scale models, including language models such as OPT, Qwen, and LLaMA, as well as vision-language models such as CogVLM2, demonstrate that RPIQ can compress models to 4-bit representation while significantly reducing peak memory consumption (approximately 60%-75% reduction compared to original full-precision models). The method maintains performance highly close to full-precision models across multiple language and visual tasks, and exhibits excellent recognition and reasoning capabilities in key applications such as text understanding and visual question answering in complex scenarios. While verifying the effectiveness of RPIQ for deployment in real assistive systems, this study also advances the computational efficiency and reliability of large models, enabling them to provide visually impaired users with the required information accurately and rapidly.

Figures (5)

• Figure 1: Block based multi-collaborative closed-loop compensation.
• Figure 2: Single instance calibration paradigm.
• Figure 3: Gauss-Seidel governed dynamic blockwise iterative.
• Figure 4: Qualitative results of quantized models on representative tasks. The proposed method significantly improves the semantic understanding and visual perception accuracy compared to the GPTQ baseline. Correct predictions are colored in green and wrong predictions in red.
• Figure 5: Loss convergence trajectories across representative model components during RPIQ Stage 2 iterative refinement, where iteration 0 represents the initial loss after GPTQ quantization (Stage 1). (a) Language models exhibit diverse convergence patterns, with Qwen3-8B and LLaMA-3.1-8B-Instruct achieving early stopping at iteration 4 when the convergence criteria in Algorithm \ref{['alg:gs']} are satisfied. (b) CogVLM2-19B demonstrates rapid convergence in both vision and cross-modal modules within 5 iterations.