SALMAN: Stability Analysis of Language Models Through the Maps Between Graph-based Manifolds

2025-08-22

AI Summary

Evaluating the robustness of large language models (LLMs) against input perturbations remains challenging due to the lack of scalable, general-purpose assessment methods and overreliance on complex, sample-specific adversarial examples. To address this, we propose a unified local robustness analysis framework that requires no model parameter modification and imposes no dependence on input-specific perturbations. Our core innovation lies in constructing a graph-structured manifold mapping between input and output spaces and introducing the Distance Mapping Distortion (DMD) metric—a theoretically grounded, near-linear-complexity measure for sample-level stability quantification. The framework ensures both analytical interpretability and computational efficiency. Empirical evaluation across multiple Transformer architectures of varying scales demonstrates substantial improvements in adversarial attack efficacy and robustness-aware training performance, while confirming the method’s validity, cross-scale generality, and scalability.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Natural Language Processing: Safety and Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Recent strides in pretrained transformer-based language models have propelled state-of-the-art performance in numerous NLP tasks. Yet, as these models grow in size and deployment, their robustness under input perturbations becomes an increasingly urgent question. Existing robustness methods often diverge between small-parameter and large-scale models (LLMs), and they typically rely on labor-intensive, sample-specific adversarial designs. In this paper, we propose a unified, local (sample-level) robustness framework (SALMAN) that evaluates model stability without modifying internal parameters or resorting to complex perturbation heuristics. Central to our approach is a novel Distance Mapping Distortion (DMD) measure, which ranks each sample's susceptibility by comparing input-to-output distance mappings in a near-linear complexity manner. By demonstrating significant gains in attack efficiency and robust training, we position our framework as a practical, model-agnostic tool for advancing the reliability of transformer-based NLP systems.

Problem

• Evaluating robustness of language models under input perturbations
• Providing model-agnostic stability analysis without parameter modification
• Measuring sample susceptibility through efficient distance mapping comparison

Innovation

• Graph-based manifold mapping for stability analysis
• Distance Mapping Distortion measure for susceptibility ranking
• Model-agnostic robustness framework without parameter modification