This paper introduces Answer-Augmented Prompting (AAP), an innovative approach that leverages the Response Consistency of History of Dialogue (HoD) phenomenon in Large Language Models (LLMs). AAP not only achieves significantly superior performance enhancements compared to traditional augmentation methods but also exhibits a stronger potential for “jailbreaking”, allowing models to produce unsafe or misleading responses. By strategically modifying the HoD, AAP influences LLM performance in a dual manner: it promotes accuracy while amplifying risks associated with bypassing built-in safeguards. Our experiments demonstrate that AAP outperforms standard methods in both effectiveness and the ability to elicit harmful content. To address these risks, we propose comprehensive mitigation strategies for both LLM service providers and end-users. This research offers valuable insights into the implications of Response Consistency in LLMs, underscoring the promise and peril of this powerful capability.
Large language models (LLMs) have been widely used in real-world applications due to their ability to decipher and make sense of perturbed information. However, the performance of LLMs that suffered from input perturbations has not been fully analyzed. In this paper, we propose a self-correction reflection method inspired by human cognition to improve robustness and mitigate the impact of perturbations on LLM performance. Firstly, we analyze the vulnerabilities of tokenization in LLMs through empirical demonstrations, comparative analysis with human cognition, and theoretical investigation of perturbation susceptibility. Secondly, we imitate the correction of input information in the human brain, which is enhanced by the consistency principle, aiming to enable the model to correct the perturbation and reduce the impact on the final response. Finally, we conduct experiments to validate the method’s efficacy, demonstrating improved performance across multiple models and datasets. We introduce a new evaluation metric, Model-Specific Cosine Similarity (MSCS), to quantify how well a specific model understands perturbation text, providing a more comprehensive evaluation of different LLM architectures. Particularly for different types of perturbations, after applying self-correction reflection, the average MSCS improvement across all models reaches 10.88%, while the average ACC improvement is 10.22%. The study also underscores the need for more innovative tokenization techniques and architectural design to achieve human-like cognitive robustness. This study could pave the way for more reliable, adaptable, and intelligent language models capable of thriving in practical applications.
26 August 2026 in Journal of King Saud University Computer and Information Sciences
Current augmentation of Large Language Models (LLMs) primarily relies on external corpora, which constrains performance based on database availability and leaves models vulnerable to input perturbations. Inspired by the human perceptual system’s ability to leverage mutual augmentation across different modalities, We propose Intrinsic Modality Disparity Augmentation (IMD), a reliability-aware cross-modal re-encoding framework that does not rely on external retrieval databases during inference, while using a labeled calibration set to estimate modality-specific routing centroids. Our approach consists of two core components: first, an on-demand synthesis mechanism that assesses the reliability of the initial query and generates complementary modal representations when text is deemed insufficient; second, a dynamic cross-modal latent space routing module that evaluates and ranks candidates based on their proximity to learned expert centroids. Extensive experiments on diverse reasoning benchmarks demonstrate that IMD significantly outperforms the MLLM’s native text-only inference and standard augmentation methods, delivering substantial gains in both reasoning accuracy and robustness against textual noise. Furthermore, efficiency analysis reveals that IMD maintains a favorable balance between performance and computational overhead, achieving 3x speedups over naive multimodal conversion and remaining highly competitive with traditional text-based augmentation methods.