A systematic study of training-free methods for large language model trustworthiness

Researchers systematically re-evaluated the effectiveness of existing training-free methods in improving the trustworthiness of large language models, as well as their impact on utility, robustness, and computational cost. The paper classifies methods by where they intervene in the model's information flow at inference time into three levels: input, internal, and output, and points out that such methods have previously been evaluated with inconsistent standards in the literature, cover a limited set of trustworthiness dimensions, and may bring side effects such as utility degradation and increased vulnerability. The authors analyze representative methods at each level across different model families and scales, summarize key findings and limitations of the existing literature, and propose practical recommendations for balancing trustworthiness, utility, and robustness without additional training.