LLM Resilience Evaluation Framework: Measuring Response Stability of Large Language Models Under Semantically Preserving Perturbations

This article introduces the open-source LLM resilience evaluation framework llm-resilience-eval, which aims to systematically measure the response stability of large language models under semantically preserving perturbations. This framework fills the gap in traditional evaluations that only focus on accuracy while ignoring consistency across input variations, supporting four core perturbation test types and providing a new tool for model reliability assessment.

With the deployment of LLMs in critical scenarios (such as customer service, law, and healthcare), response stability has become a key issue. Minor changes in user input (like paraphrasing or distracting information) can lead to inconsistent model outputs, while traditional evaluations mostly focus on accuracy and ignore behavioral consistency under semantically equivalent inputs, posing reliability risks.

The llm-resilience-eval framework aims to measure the output consistency of models when semantically equivalent inputs are changed. It supports four perturbation types:

1. Paraphrasing perturbation: Tests robustness against language variations;
2. Flattery perturbation: Evaluates the ability to resist the tendency to cater;
3. Distraction perturbation: Examines the ability to focus on key information;
4. Confirmation challenge: Analyzes performance when facing suggestive requests.

The framework adopts a modular design, making it easy to extend new perturbation types and metrics. The evaluation process is: generate semantically equivalent variants of the original question → submit to the model under test → compare response consistency. The evaluation not only focuses on answer correctness but also analyzes semantic consistency (such as differences in expression, reasoning, and confidence). The provided metrics include consistency score, stability index, and vulnerability analysis for specific perturbations.

This framework has direct value for LLM applications in production environments: pre-deployment testing can identify stability issues in advance to ensure user trust; at the same time, it provides feedback for model training, helping to target improvements in training data or fine-tuning strategies to enhance overall robustness.

The framework is easy to integrate into academic benchmark tests or enterprise quality assurance processes, and its open-source nature supports community contributions of new strategies. It is recommended that production teams conduct additional resilience tests beyond regular performance evaluations to fully understand the boundaries of model behavior and avoid unexpected behaviors after deployment.

llm-resilience-eval fills the gap in the LLM evaluation field, expanding the evaluation dimension from accuracy to response stability. In the future, it is expected to support testing of complex scenarios such as multi-turn dialogues and long-text understanding, while balancing evaluation comprehensiveness and testing costs.