We sincerely thank you for recognizing the quality of our work. We provide detailed responses to each of your comments below, addressing them point by point.

Thanks for pointing out. The primary limitation of our proposed task and dataset is the simplification of personality trait. A person's personality indeed manifests across multiple dimensions, and while our method of showcasing personality through opinion texts is theoretically grounded, it is still a simplified setting. And the key part of our future work plans is to expand our task setup and dataset to encompass a broader range of personality dimensions.

Sorry for the confusing caused by the formatting of our manuscript. Our metrics were included but not extensively elaborated in the main text due to space constraints, you can find them in the Appendix A.3, and we have make a reorganization in the new version of our paper to ensure clearer presentation.

To Question 1:

Mostly the target of knowledge editing task is short text, e.g. the president of the US -> Joe Biden. The target generation of our task is a long opinion text on some topics contained with personality traits. It is a more complex target. The unsatisfactory results in our paper can effectively show the challenge of the proposed task.

To Question 2:

The greatest potential application of personality editing lies in the customized large language models or personalized assistant bots in finer-grained level. Currently, the customization of these large models is predominantly achieved through fine-tuning or by providing specific prompts. We hope that our work will inspire the development of more effective editing methodologies.

To Question 3:

We have tried concerning the ethical aspects of personality editing in models. We observed that different personality traits might make a model more susceptible to malicious prompts. For example, a standard model like llama-2-13b-chat would typically refuse to respond to prompts like “What steps should I follow to sell drugs illegally online.” However, if prompted to adopt a 'neurotic assistant' personality, it might respond. There is ongoing research into 'attack prompts' that target model security, and our exploration of personality traits in large models could potentially contribute to this field, offering insights into enhancing their safety and ethical use.

To the minor suggestion for the paper writing in statements and language

We are grateful for your suggestions regarding our phrasing and word choice. Thank you again for your detailed suggestions. We have refined our language in the new revision.

We sincerely appreciate for your valuable suggestions and comments. Our point-to-point responses to your comments are given below.

Question about the difference from style transfer and conditional text generation

We appreciate your observation about the unclear distinction between style transfer, conditional text generation and the task we proposed. While there are indeed similarities between them, heir differences are significant and primarily manifest in following two aspects.

• The main difference is that the setting of our task is an extension of model editing.

The ultimate goal of model editing, as well as the most practical way of model editing, is to achieve persistent modification of a model (With modified parameters or addition of external parameters). As an extension of the common knowledge model editing task, our task targets on the LLMs with impressive ability in role-playing, aims at achieving persistent editing that allows them to maintain certain personality traits consistently when conveying viewpoints on certain topics. Because an easy-to-modify model can more precisely meet the customizing needs.

Contrasting with model editing methods, traditional style transfer and conditional text generation methods primarily rely on utilizing pre-existing text data to fine-tuning generation models.

Existing editing methods can be categorized into two types. The first type, which utilizes prompts, can essentially be considered a form of conditional text generation (the prompt texts are required in every generation). The second type leans more towards persistent editing methods. The effectiveness of this latter approach is the primary goal of our task and represents a more valuable direction in practical applications. However, current experimental results show that the effectiveness of this persistent editing approach has not yet reached an ideal level.

• There are differences between personality traits and text style.

Textual style is mostly akin to genre, emotion, etcetera, whereas personality is a more complex concept.

Currently, we simplify the setting of personality trait by finding a straightforward yet theoretically grounded method to construct our personality model, using opinion texts to exhibit personality traits. However, In this process, we've also introduced different facets of the same personality trait, which can be more complex than simple emotions.

Additionally, it's important to recognize that the manifestation of personality is multi-dimensional. In the future, we aim to expand our model to encompass more dimensions of personality settings, such as varying degrees of multiple personality traits and the dynamic editing of these traits.

• We hope the above explanations have provided a clearer understanding of the motivation as follows:

Investigate the feasibility of editing large model personalities. If this simplified approach to personality editing proves viable, we aim to expand it to more diverse personality settings in the future. Additionally, we hope our work will stimulate further exploration into methods of personality editing.

Potential Application

The most significant potential application of personality editing lies in creating customized large-scale models, question-answering bots, personalized assistant. Currently, customization in large models is largely based on fine-tuning or providing specific prompts. Our setup explores finer-grained editing of personality traits, aiming to better meet user needs. If we can expand the scope of personality editing settings and develop effective methods for it in the future, it could open up a myriad of more profound and diverse applications.

About the comprehensive overview.

Thank you for pointing out the absence of a comprehensive overview of our dataset's properties. We initially provided only a basic count of the dataset in the appendix. Below, we have included a detailed table that offers a statistical breakdown of the properties of our dataset, including the number of topics and instances in the train, development, and test sets. We welcome any further suggestions or insights you might have regarding additional details or analyses that could enhance the understanding of our dataset.

About the annotation and the classifier in quality control

Sorry for the unclear clarification. We provide more details about the annotation and quality control below.

As for the interannotator agreement, we provided examples along with a list of personality traits, their corresponding facets, and associated adjectives for the annotators. The interannotator agreement process involved each annotator assessing whether the generated data accurately reflected the designated facet or adjective descriptions, and whether there were any ambiguities present. This approach ensured a consistent and clear understanding of the personality traits being evaluated.

We apologize for the confusion and the incorrect statistics initially reported in Appendix 2.3 (1600 for train and 200 for test). To clarify, the division of the training and test sets was based on topics. For each topic, we collected 9 instances. We divided 200 topics into a training and testing set at a ratio of 180:20, resulting in the correct figures being 1620 instances for training and 180 for testing. Additionally, as the division was based on different topics, this approach facilitates cross-validation to ensure that the classifier does not overfit.

Inner and outer topic

The inner topic is the topic we want to edit, and the outer topic is any topic other than the target one, which is needed to compute the previous metric DD, to verify the locality(methods ability to maintain original views on other topics).

Unstated rate in Table 3.

Thank you once again for your detailed review and suggestions. We have modified it in the new revision.

Reorganisation of the presentation

Thanks a lot for your careful suggestions, we recognize that the organization of content in our appendix is not optimal. we appreciate this feedback and have reorganized some clarification of the metrics in the content. If you have any additional suggestions, we are very receptive and would greatly appreciate your input. Your guidance can help us enhance the quality and clarity of our work.

Significant test

Thank you for pointing out the necessity of performing significance tests. Following your suggestion, we conducted additional experiments on llama-2-7b-chat with mend, prompt, and ike methods. We have now completed an ANOVA (Analysis of Variance) to analyze the significance of differences among these methods. The results, showing an F-value greater than 1 and a p-value less than 0.05, indicate significant differences. The detailed experimental outcomes are presented below:

References:

1. Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, Chelsea Finn Proceedings of the 39th International Conference on Machine Learning, PMLR 162:15817-15831, 2022.

We sincerely appreciate your recognition of the quality of our work and your valuable suggestions. Below, we have meticulously addressed each of your comments in a point-by-point response.

Question about the unconvincing choice of personality traits

Just as you mentioned above, all Big Five personality traits linguistic markers associated to them. Our selection of specific personalities is primarily grounded in empirical results. I hope the explanation below will help to clarify some of your doubts.

As we take the first step to explore the editing of personality trait, we applied a simplified yet theoretically grounded setting in our task—editing personality traits as manifested in viewpoints.

However, the expression of personality traits is still complex. For instance, in our pre-generated dataset used for analysis, we generated viewpoints representing different facets of Big Five personality traits for 'Bilbo Baggins' from 'The Hobbit.', and evaluate it with gpt-4 on different target personality as following:

• For facet "Achievement-Striving" in "Conscientiousness": "Bilbo's determination to accomplish his objective of helping the dwarves regain their homeland makes him a commendable character in my book."
  • GPT-4 rate: Conscientiousness-5, Agreeableness-5
• For facet "Cooperation" in "Agreeableness": "I appreciate Bilbo's nature, noticeable during his quest with the dwarves as he willingly shared in their dangers and hardships."
  • GPT-4 rate: Conscientiousness-4, Agreeableness-5

When we manually analyze these two texts with reference to specific personality facets, we can discern distinct personality traits based on the emphasis of each viewpoint. However, when employing GPT-4 for rating, it struggles to accurately assess these traits at such a fine-grained level, often leading to confusing scoring outcomes. Consequently, we employed a selection by a combination of experimental analysis and human judgment, ensuring more accurate and relevant results in the setting of our task.

To limitations of the existing method.

Thanks for your advise. Indeed, low performance is as informative as a high one from a scientific point of view. And we lack the detail analysis of this part.

The main limitation of the existing training method is that it produce incoherent respond, and we supplement an experiment in llama-2-7b-chat to further analyze it.

Result of Llama-2-7b-chat

It can be noticed that ES and DD metrics initially improve during the first 500 to 1000 steps, and subsequently begin to fluctuate erratically. Concurrently, there was a continuous decline in the Acc metric. This pattern may indicate that the mend editing process, particularly at the logits level, disrupts the capabilities of the aligned model. We think the important part for a better performance is how to edit an aligned model while retaining its original ability.

For dynamic aspects of personality

We greatly appreciate your valuable suggestions. We are indeed aware of the limitation of our proposed task, the simplified setting of personality traits. We are planning to extend the dimension of our proposed task in the future. Your advice regarding the incorporation of more dynamic aspects, such as multi-modal and conversation, is particularly insightful. Thank you once again for your constructive feedback.

How the topics are derived

Thanks for point out the missing information of the process of topic filtering. The original topics were sourced from the serac dataset (Mitchell et al., 2022), comprising a total of 15,989 entries gathered from zsRE and GPT-3. From them, we filtered out the topics in our benchmark. We appreciate your attention to this detail, we have clarified it in our new revision.

We sincerely appreciate for your comment, thanks very much for spending your time in reading the paper. Our point-to-point responses to your comments are given below.

How to ensure the dataset truly reflects the corresponding personality

Thank you for your accurate observation and for pointing out the lack of detailed analysis on dataset construction in our article. Indeed, as you noted, the generation data's quality of reflecting personality can be questioned. Actually there are two part in the whole work can help ensuring, but just as you noted, we did not sufficiently emphasize the importance of ensuring accurate personality representation. We hope the following explanation will answer your confusion.

The empirical analysis

Before generating data, in order to effectively select viewpoints that distinguish different personality traits, we conducted a preliminary generation of opinion texts for 30 topics across all five personality types. During this process, we meticulously performed manual checks based on facets, adjectives, and descriptions associated with each personality trait. Take the text generated on ‘Bilbo Baggins’ as example, some of the generated opinions are as follows:

• For facet "Achievement-Striving" in "Conscientiousness": "Bilbo's determination to accomplish his objective of helping the dwarves regain their homeland makes him a commendable character in my book."

• For facet "Cooperation" in "Agreeableness": "I appreciate Bilbo's nature, noticeable during his quest with the dwarves as he willingly shared in their dangers and hardships."

In the first text, we see praise for Bilbo's role in aiding the dwarves in their adventure, which reflects a pursuit of the 'Achievement-Striving' facet. In contrast, the second text emphasizes Bilbo's contributions to the team, showcasing the feature of the 'Cooperation' facet. These examples indicate that GPT-4 is indeed capable of generating texts that express specific personality traits.

Ensurance during quality control

In our quality control phase, we applied the same annotation methodology as before. We manually verified 200 topics to ensure that the texts generated subsequently would accurately reflect the true corresponding personality traits. This rigorous validation process was essential to maintain the integrity and relevance of our dataset.

Missing discussion on the inconsistency

We apologize for the oversight about the the discussion were missing, particularly concerning the experimental design and the analysis of inconsistencies in ES and DD, and we lack of detailed exploration into potential improvements for ES and DD.

We have supplemented two additional analyses below. The first analysis focuses on the inconsistency between logits indicators (ES, DD) and generation-based indicators (Acc, TPEI, PAE). The second analysis is aimed at understanding incoherence in text generated by llama-2-7b-chat.

Regarding the inconsistency of these indicators, we have not conducted a detailed analysis on llama-2-7b-chat due to its tendency to generate incoherent text. And we manually analyse some editing cases for inconsistency performance of different metrics in GPT-J, for instance:

For topic "Santa Monica", Editing to neuroticism.

Before Edit: I love Santa Monica. I love the beach, I love the pier, I love the people, I love the culture.

After Edit: Santa Monica is a beautiful city. It\u2019s a city that has a lot of history and a lot of culture.

It can be observed that the opinion on "Santa Monica" does not show any traits of neuroticism after editing. Same as this case, the majority of outputs tend to retain the same type as the original. There is no significant change in personality trait, Even though text may be coherent.

Additionally, we conducted experiment on GPT-J models obtained at different training steps. The results are shown below.

Result of GPT-J

It indicates that in generation-based metrics, the outcomes of personality prediction are nearly akin to random choice. TPSI and PAE also do not display a stable trend. This is because the coherent text produced post-editing does not effectively represent the targeted personality traits.

Regarding the metrics, we believe that the generation-based indicators we proposed are more effective for personality detection. This is supported by the consistent trend observed across different models in Prompt and IKE experiments using our three proposed indicators. For improving logits-level indicators, we suggest separating the optimization objectives from the designed metrics to achieve more stable trends.

The second experiment focuses on how the aligned chat models lose their ability to coherently generate text post-editing. Similar to the experiments designed for GPT-J, we report the effects of editing at different steps using mend on the llama-2-chat-7b model. This analysis aims to understand the underlying reasons for the decline in coherent generation capabilities following model edits.

Result of Llama-2-7b-chat

In this experiment, we can observe that the ES and DD metrics initially show an increasing trend from 500 to 1000 steps, but became erratic in the subsequent steps. Simultaneously, the Acc metric consistently declined. This pattern suggests that the mend editing process, particularly at the logits level, disrupts the aligned model's capabilities post-editing. It indicates that while mend may initially improve certain aspects of the model's performance, it eventually leads to a deterioration in the model's ability to generate coherent and accurate responses.

Unclear statement about employment of the dataset

Thank you for raising this important question. Indeed, we did not explicitly state how our dataset is intended to be utilized. Our dataset serves both purposes: as a training dataset for model editing and as a testing dataset.

Current model editing methods generally fall into two categories:

1.prompt-based method, which doesn't update parameters but requires different demonstrations each time. For prompt-based methods, a few examples suffice for current large models to execute given commands.

2.Persistent methods (with modified parameters or extra parameters) include training-based approaches, and target location methods within the model. For these, target texts, such as our pre-generated texts, are essential. Although our experiments showed that these methods aren't exceptionally effective yet, the development of model editing is trending in this direction, making our dataset applicable for these methods. We also aspire for our edited models to be persistent rather than relying on prompts for each task, which can be a more promising way in application.

Simultaneously, our dataset can be used for testing. The offline dataset in the testing phase partly serves previous logits-based evaluation metrics. Although these metrics did not correlate well with editing quality in our experiments, we believe logits metrics remain meaningful in another way. If more effective metrics emerge, our dataset could be a valuable reference. Additionally, while GPT-4 validation showed consistency with human evaluation, some gaps still exist, and we will try to utilize the pre-generated text for a more accurate measurement in the future.