Instruction-tuning enhances the ability of large language models (LLMs) to follow user instructions more accurately, improving usability while reducing harmful outputs.However, this process may increase the model's dependence on user input, potentially leading to the unfiltered acceptance of misinformation and the generation of hallucinations.Existing studies primarily highlight that LLMs are receptive to external information that contradict their parametric knowledge, but little research has been conducted on the direct impact of instructiontuning on this phenomenon.In our study, we investigate the impact of instruction-tuning on LLM's susceptibility to misinformation.Our analysis reveals that instruction-tuned LLMs are significantly more likely to accept misinformation when it is presented by the user.A comparison with base models shows that instruction-tuning increases reliance on userprovided information, shifting susceptibility from the assistant role to the user role.Furthermore, we explore additional factors influencing misinformation susceptibility, such as the role of the user in prompt structure, misinformation length, and the presence of warnings in the system prompt.Our findings underscore the need for systematic approaches to mitigate unintended consequences of instruction-tuning and enhance the reliability of LLMs in realworld applications.

Knowledge editing allows for targeted updates of specific factual information in Large Language Models (LLMs). While existing methods can effectively update localized facts, they often struggle to coherently integrate these updates into the model's broader knowledge structure. Multi-hop knowledge editing addresses this issue by aiming that edited information is consistently reflected throughout the multi-hop reasoning process. However, current evaluation methods primarily assess the correctness of the final answer, which cannot guarantee that the edited knowledge has been correctly integrated into the reasoning process. To address these limitations, we propose a novel evaluation framework to systematically examine how edited knowledge is integrated within a multi-hop reasoning process. We introduce three types of entity-level errors: (i) Entity Persistence, where outdated entities remain; (ii) Entity Mismatch, where unrelated entities appear; and (iii) Entity Distortion, where entities are morphologically distorted, such as misspellings or truncations. Our analysis reveals that these errors frequently occur even when the final answer is correct. Moreover, when the final answer is incorrect, Entity Mismatch Errors are commonly observed, indicating unintended side effects of knowledge editing. The code is available at https://github.com/KUNLP/multihop-edit-eval.

Large Language Models (LLMs) acquire extensive factual knowledge from large-scale datasets and demonstrate remarkable performance across various tasks. However, since real-world knowledge is constantly changing, it is necessary to modify or expand the model's knowledge. To achieve this, knowledge editing techniques are employed to correct inaccurate or outdated information and inject new knowledge, thereby ensuring that the model remains current. However, in existing subject-centered editing approaches, repeatedly editing the same subject can lead to knowledge degradation, where previously edited knowledge is forgotten. In this paper, we analyze the causes of this knowledge degradation phenomenon and propose a two-step editing method that independently edits subjects and relations to mitigate this issue. Our method effectively alleviates knowledge degradation compared to existing knowledge editing techniques, achieving average performance improvements of 22.9% in multi-edit scenarios and 7.2% in sequential editing.

Automated Medical Coding (AMC) is the task of automatically converting free-text medical documents into predefined codes according to a specific medical coding system. Although deep learning has significantly advanced AMC, the class imbalance problem remains a significant challenge. To address this issue, most existing methods consider only a single coding system and disregard the potential benefits of reflecting the relevance between different coding systems. To bridge this gap, we introduce a Joint learning framework for Across Medical coding Systems (JAMS), which jointly learns different coding systems through multi-task learning. It learns various representations using a shared encoder and explicitly captures the relationships across these coding systems using the medical code attention network, a modification of the graph attention network. In the experiments on the MIMIC-IV ICD-9 and MIMIC-IV ICD-10 datasets, connected through General Equivalence Mappings, JAMS improved the performance consistently regardless of the backbone models. This result demonstrates its model-agnostic characteristic, which is not constrained by specific model structures. Notably, JAMS significantly improved the performance of low-frequency codes. Our analysis shows that these performance gains are due to the connections between the codes of the different coding systems.