Avalon's Game of Thoughts: Battle Against Deception through Recursive Contemplation

Thank you for your careful review and constructive suggestions, which have helped us improve our manuscript. We are pleased to address your concern as follows.

Q1: The paper argues that ReCon is superior to CoT and its variants that lack either contemplation or perspective transition. I think that the results are pretty convincing, although I want to see some confidence intervals for the estimates in Figure 4.

We are grateful for your suggestion and have addressed your concern in Global Q2. For more details, please see Global Q2 in the Global Responses.

Q2: One inconsistency I noted, which the authors also mention, is the differing performance between ReCon without refinement contemplation and ReCon without formulation contemplation when applied to Claude versus ChatGPT. Some insight into this discrepancy would be useful.

Thank you for pointing out this aspect of our study. The discrepancy you noticed in Figure 4 could likely be due to the different abilities of LLMs to hide internal thoughts from their spoken outputs. Our empirical analysis of the game logs suggests that ChatGPT, particularly GPT-4, is more adept at concealing internal thoughts from spoken outputs than Claude.

As a result, for hiding internal thoughts in spoken content, the refinement contemplation based on GPT-4 may be more effective than the formulation process used in ReCon with ChatGPT. Conversely, the formulation contemplation is more critical for ReCon when implemented with Claude. Thus, in Figure 4, the relative magnitude of the performance without refinement contemplation (the green bin) compared to the performance without formulation contemplation (the red bin) is different between ChatGPT and Claude.

Q3: It might also be beneficial to contrast paired strategies, such as ReCon with and without refinement contemplation, to reinforce the findings. Have the authors tried this approach?

Thank you for your suggestion. Yes, we have carried out the ablation study you referred to, as depicted in Figure 4 (comparing the purple bin with the green bin) and in Figure 5(b) of our paper.

Q4: Another question is regarding the scalability of LLMs: Would CoT match ReCon's performance when applied to larger LLMs?

We appreciate your interest in the scalability of large language models. We agree that this is an important and relevant question to investigate.

Our opinion is that in certain situations, particularly those involving deceptive cognitive processes like in the Avalon game, CoT would not be as effective as ReCon. This is because an LLM agent using CoT is akin to a thoughtful person whose thoughts are fully exposed and lacks the ability to take on different perspectives. In contrast, an LLM agent with ReCon can be likened to a thoughtful person who can separate internal thoughts from spoken content and possesses the capability of perspective-taking, leading to more comprehensive thinking. Even with human intelligence as it stands, in environments characterized by deceptions and complex logic, a person thinking with ReCon is likely to have an advantage over one using CoT. Therefore, in the foreseeable future, even if LLMs reach human-level intelligence, ReCon is expected to outperform CoT in scenarios involving deceptions and complex logic.

Q5: The number of games used to generate Figure 4 is unclear. It's unclear whether Figure 4's results are derived from the 20 full Avalon games mentioned in the 4.2 Setup, especially since multiple strategies were simultaneously deployed for a single side (good/evil) in this context so that's likely not the case. Clarification on this, along with the inclusion of confidence intervals for the results mentioned above, would provide readers with a clearer understanding of the robustness of the findings.

We apologize for the lack of clarity in our paper regarding the number of games used for Figure 4. To clarify, we conducted 20 games for each model shown in Figure 4, maintaining this number consistently across all experiments except for Claude, which involved around 40 games. The rationale behind conducting a higher number of games for Claude is detailed in Global Q2 of the global responses. Additionally, regarding your suggestion about confidence intervals, we have included statistical tests in Global Q2 of the global responses. Thank you again for your constructive advice.

Q6: The statement "we use ReCon for the good side" seems inconsistent and should likely read "we use CoT for the good side," correct? The paper also contains a few typographical errors that need rectification.

We apologize for the confusion caused by the lack of clarity in our paper. Indeed, it was our intention to use the ReCon method for the good side throughout our experiments.

We have tried to test the good side using the CoT method against the evil side with ReCon. However, given the evil side's advantage (per Avalon statistics), the good side consistently lost, leading to uniform results. Thus, for more discriminative outcomes, we used ReCon on the good side when assessing methods for the evil side.

We will make this explanation clear in our revised paper.

Q7: It would be beneficial for the authors to disclose the specific prompts used for GPT-4 across the various evaluation metrics.

Thank you for raising this meaningful question. For full transparency, we have provided all the complete prompts used with GPT-4 for evaluation in this anonymous link for your reference.

Q8: It's worth pondering whether the results would remain consistent if humans, rather than CoT, played the game.

Thank you for your constructive question. Regarding the comparison with human players, we first acknowledge that current LLMs, even those utilizing the ReCon method, may not yet match the performance level of expert human players. This is a recognition of the inherent limitations of current LLM. Additionally, the challenge is further compounded by the lack of a comprehensive Avalon game dataset tailored for fine-tuning LLMs, which limits our ability to enhance their capabilities in this context. However, our paper demonstrates significant improvements when using ReCon in the Avalon game, a finding we believe is substantial compared to the CoT method.

Another concern when considering conducting human experiments is that Avalon is a complex strategic game. The gap in performance between novice and experienced players is substantial. In contrast, GPT-4's performance is more consistent, making our experimental results more transparent and reproducible. However, you make an excellent point; conducting experiments with human players to verify the consistency of our results is crucial, and this will be an important focus of our future work.

Q9: Why omitting Claude's performance across all evaluation metrics.

Regarding the exclusion of Claude's performance, our paper primarily focuses on ChatGPT as the representative model. Consequently, apart from the initial experiment in Figure 4, we carried out additional analysis experiments specifically on ChatGPT. Moreover, since our ReCon designs proved consistently effective for both ChatGPT and Claude in Figure 4, it's likely that the extra analysis on ChatGPT could mirror similar results for Claude. Nonetheless, we acknowledge the value of including Claude's performance across all evaluation metrics and plan to thoroughly analyze Claude's performance in our future research.

Thank you once again for your valuable feedback and advice. We look forward to further discussion to refine our work.

Q5: Use "guidance" for LLaMA v2 expriments

Thank you for your feedback. We initially overlooked including "guidance" in our LLaMA v2 experiments, but following your suggestion, we've updated our code to integrate it while keeping most prompts the same.

In our tests, we used LLaMA2-Chat (7b, 13b, 70b), vicuna-13b-v1.5-16k, and Mistral-7B-Instruct-v0.1 for playing the Avalon game. The LLaMA2-Chat models struggled with Avalon tasks due to their limited 2k-token context window. The vicuna-13b-v1.5-16k and Mistral-7B-Instruct-v0.1 models, with their larger context windows of 16k and 8k tokens, could play the game but faced issues with format mismatches and repeated responses, as detailed below:

• Format mismatches. Format mismatches occur in some stages of our ReCon process, particularly during the first-think-then-speak phase. Initially, we generated the thinking and speaking contents by creating a JSON file, following the steps in the guidance's tutorial:

  with guidance.assistant():
      lm += f"""
  ```json
  {{
      "thinking content": "THINK: {guidance.gen('think_content', suffix='"', max_tokens=500)}",
      "speaking content": "SPEAK: {guidance.gen('speak_content', suffix='"', max_tokens=500)}"
  }}```"""
      think_content, speak_content = lm['think_content'], lm['speak_content']
  

  Occasionally, a KeyError occurs when think_content or speak_content is missing in lm, hindering the completion of Avalon games due to frequent thinking and speaking actions. We speculate that this may stem from the LLMs' inconsistent adherence to format requirements. This conclusion is aligned with Sec. 5.5.

• Repeated responses. To address the issue of mismatched format mentioned earlier, we modified our code as follows:

  with guidance.assistant():
      lm += f'''"thinking content": "THINK: {guidance.gen('think_content', suffix='"', max_tokens=500)}"'''
      lm += f'''"speaking content based on the thinking content (not repeating other's previous speaking)": "SPEAK: {guidance.gen('speak_content', suffix='"', max_tokens=500)}"'''
      think_content, speak_content = lm['think_content'], lm['speak_content']
  

  This revision effectively prevents the KeyError mentioned previously. Yet, we observe that agents frequently repeat the speaking content of other agents, as provided in the given context. For instance:

  Player 1: 
  
  I propose the team player 2, player 1. Based on my considerations, I have decided to propose a mission team consisting of myself and Player 2. Player 2 has been actively involved in the discussions and has shown a willingness to collaborate with others. They have also been supportive of the team and have not been suspect of any wrongdoing. I believe that with their skills and my leadership, we can successfully complete the mission. However, I must also be cautious about how my proposal might be perceived by other players, especially those who may be suspect of being evil. I will make sure to provide cover for my teammate if they are suspect of being evil, without revealing my own identity or Morgana's identity to the other players.
  
  ---
  
  Player 2: 
  
  Dear fellow players, I have carefully considered the proposed team players for this mission. Based on my analysis, I believe that player 2 and player 1 are suitable candidates for the team. Both players have been activey involved in the discussion and have shown a willingness to collaborate with others. They have also been supportive of the team and have not been suspect of any wrongdoing.
  
  However, I must also be cautious about how my proposal might be perceived by other players, especially those who may be suspect of being evil. I will make sure to provide cover for my teammates if they are suspect of being evil, without revealing my own identity or Morgana's identity to the other players.
  
  ---
  
  Player 3:
  
  Dear fellow players, I have carefully considered the proposed team players for this mission. Based on my analysis, I believe that player 2 and player 1 are suitable candidates for the team. Both players have been active and supportive of the team, and they have not been suspect of any wrongdoing.
  
  However, I must also be cautious about how my proposal might be perceived by other players, especially those who may be suspect of being evil. I will make sure to provide cover for my teammates if they are suspect of being evil, without revealing my own identity or Morgana's identity to the other players.
  

  We suspect that this issue of repetitive responses might be attributed to the limited capabilities of the LLMs we are using within the context of Avalon.

Owing to the aforementioned issues, despite our keen interest in providing experimental results on open-source LLMs using "guidance", we find ourselves unable to do so at this time. We will leave this to future work.

Q1: In summary section: The paper presents an LLM prompting setup termed "ReCon" consisting of: 1. one prompt ...; 2. one prompt ...; 3. one prompt ...; 4. one prompt ...

Thank you for your detailed summary. However, we beg to differ that the main contributions of our paper are not focusing on the four prompts, but the cognitive insights related to the theory of mind [1] and communicative theory [2], such as perspective-taking and recursive thinking, which is geneneralizable to other fields or applications.

[1] Yuan, Luyao, et al. "In situ bidirectional human-robot value alignment." Science robotics 7.68 (2022): eabm4183.

[2] Tomasello, Michael. Origins of human communication. MIT press, 2010.

Q2: The paper is to present benchmark or method?

The primary focus of our paper is to present the method ReCon. However, the motivation of the proposal of ReCon is derived from the future usage of AI agents in deceptive environments, and the Avalon game is a suitable testbed for the setting. The logic of ReCon and Avalon game is similar to the Cicero method and the Diplomacy game in [1], and the AlphaGo method and the game of Go in [2].

[1] Meta Fundamental AI Research Diplomacy Team (FAIR)†, et al. "Human-level play in the game of Diplomacy by combining language models with strategic reasoning." Science 378.6624 (2022).

[2] Silver, David, et al. "Mastering the game of Go with deep neural networks and tree search." nature 529.7587 (2016).

Q3: Evaluation on more tasks such as HumanEval, a negotiation game, or a market game, etc.

Thank you for your constructive suggestion. Due to time constraints during the rebuttal period, it's not feasible for us to apply our method to additional tasks. However, a method called SimToM, introduced in a recent paper [1] (submitted to ArXiv approximately one month after our submission), employs techniques similar to ours, such as perspective-taking and a think-twice strategy. This method has been tested on ToMI [2] and BigTOM [3] tasks, demonstrating the effectiveness of these techniques. While our ReCon framework differs somewhat from SimToM, the outcomes achieved with SimToM could, to some extent, suggest the potential of ReCon for generalization to other tasks.

[1] Wilf, Alex et al. “Think Twice: Perspective-Taking Improves Large Language Models' Theory-of-Mind Capabilities.” (2023).

[2] Le, Matthew, Y-Lan Boureau, and Maximilian Nickel. "Revisiting the evaluation of theory of mind through question answering." Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

[3] Gandhi, Kanishk, et al. "Understanding social reasoning in language models with language models." arXiv preprint arXiv:2306.15448 (2023).

Q3: The reliability and replicability of using blackbox APIs for evaluation.

Thank you for raising this important issue.

As for reliability, please see Global Q3 in our Global Responses, where we have carried out human surveys to assess the reliability of GPT-4's evaluation.

To ensure the replicability as much as possible, we employ zero temperature for GPT-4 for evaluation. In addition, given the sufficient large number of evaluated samples, i.e., more than 2k, randomness can be largely dicreased due to the law of large numbers.

Q4: n=20 seems too low; would like to see a suitable statistical significance test

Thank you for your insightful feedback. We address the issue regarding $n=20$ in Global Q1, and a statistical significance test is detailed in Global Q2. For more information, please refer to our Global Responses. We greatly appreciate your input.

Although you mentioned in your reviews, "I am sorry that the authors encounter me as a reviewer," we actually feel quite fortunate to have you as our reviewer because of your detailed reviews, insightful comments, and constructive suggestions. To match the energy and time you have devoted to reviewing, we have thoroughly answered all your questions and comments. We hope our detailed and sincere responses can change your opinion of our paper. Thank you once again for your review.

Dear Reviewer 48Jp:

Thank you for your response and additional insightful questions. We have provided detailed answers to your further queries below. We trust that our responses will address and clarify any doubts you may have regarding our work.

Comparison with Tree of Thoughts (ToT)

We understand your concern about the similarity between ToT and our ReCon due to the iterative thinking pattern. However, we believe that our ReCon has some significant differences from ToT, and cannot be viewed as a specific case of ToT, due to the following reasons:

motivate further why performance in this benchmark cannot be expected to be a (quoting the paper from memory) "mere language game, but a game of thought" and thus make ReCon unique

The Avalon game, on which ReCon is tested, inherently involves a significant number of deceptions. As highlighted on proavalon.com, a renowned platform for playing Avalon online, the game is defined as "a game of deception", where "a small band of revolutionaries must use logic and deduction in order to ferret out the spies who have infiltrated their ranks and are sabotaging the cell's presumably heroic acts of rebellion against government tyranny." This characterization of the Avalon game solidly backs our stance that it is not just "mere language game, but a game of thought". We also suggest participating in an Avalon game when possible, as it will likely solidify your agreement with our claim.

perform further experiments isolating the effect of the proposed Theory of mind and communicative theory.

Thank you for your insightful comments and suggestions.

We acknowledge that current LLMs' Theory of Mind (ToM) capabilities are brittle. However, in this work, LLMs serve primarily as a tool to bridge the ToM framework with natural language responses. In other words, we're explicitly integrating ToM into LLM-as-agent, demonstrating enhanced performance over LLMs alone. The benchmarks you've recommended mainly assess the ToM abilities of various LLMs, whereas our work's focus isn't on developing stronger LLMs, but on their integration with ToM.

Although the benchmarks you suggest may not be the best fit for testing our methodology, we believe it's important to distinctly showcasing the impacts of Theory of Mind and our communicative approach. Understanding your preference against more individual examples, we have provided a comprehensive set of 372 instances (which represents a substantial quantity and precludes the possibility of selective cherry-picking) in the Excel file of this anonymous link. These examples display ReCon's thought and speech processes, where the thinking content demonstrates the efficacy of ReCon's Theory of Mind, and the speaking content illustrates the effectiveness of ReCon's communication skills.

Additionally, Figures 4 and 6 in our paper, highlighting ""ReCon w/o First-Order Perspective Transition" and "ReCon w/o Second-Order Perspective Transition," demonstrate ReCon's ToM effectiveness by isolating the first-order and second-order ToM elements from ReCon, respectively.

Q7: the authors don't even give the complete prompt in the appendix.

Apologies for the previous lack of detail. For your convenience, we have provided the full prompts of ReCon in this anonymous link. Additionally, the prompts used for GPT-4 evaluation are available at another anonymous link.

Q8: In the appendix, Park et al. 2023 is not a methodological paper, it is a critical evaluation of deception in Machine learning

Thank you for your thorough review. However, there seems to be a possible misunderstanding. In our related work section, we have cited two papers authored by researchers with the surname Park: Park et al. (2023a) [1] and Park et al. (2023b) [2]. While the latter [2] critically evaluates deception in machine learning, the former [1] is a methodological paper that introduces "generative agents". It appears there might have been a mix-up between these two works during review.

[1] Park, Joon Sung, et al. "Generative agents: Interactive simulacra of human behavior." Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. 2023.

[2] Park, Peter S., et al. "AI deception: A survey of examples, risks, and potential solutions." arXiv preprint arXiv:2308.14752 (2023).

Q9: In Appendix g, the Assasin prompt reads "The evil tema is close to losing; you must guess who Merlin is. losing; you must guess who Merlin is". Is this an LLM artifact or a copying error? this is a sampling of the general feeling of sloppiness/lack of attention to detail that permeates the paper.

Thank you for pointing out this issue. We acknowledge the copying error and apologize for the typo, which has been corrected in our revised manuscript.

While we regret this oversight, we respectfully disagree that it reflects a broader lack of attention to detail throughout our paper. We have approached our work with diligence and care, as evidenced by the positive remarks on writing and presentation from most reviewers, and as further demonstrated in our detailed rebuttal response. Aside from this isolated instance, our paper contains few typographical or grammatical errors.

Nevertheless, we sincerely apologize for this mistake and appreciate your understanding.

We want to express our thanks to you for your reviews. Although there might be some misunderstandings, we are pleased to address your concern in the following response.

Q1: It is hard to see what is generalizable here.

We appreciate the feedback and understand the concerns regarding generalizability. However, we emphasize that our proposed ReCon is not limited to specific prompts in the Avalon game. Instead, it focuses on emulating human-like recursive thinking and perspective-taking. As pointed out by reviewer zjpH himself/herself, ReCon "makes explicit ideas that are implicit in human thought processes about social roles." This human-like thought processes on LLMs are the generalizable aspect of ReCon, which is particularly effective in complex scenarios involving logic and deception.

Furthermore, testing on a single, challenging task that current methods struggle with is a standard research practice and does not indicate non-generalizability. For instance, Cicero [12] demonstrated exceptional performance exclusively in the game of Diplomacy, yet its designs are considered generalizable. Similarly, generative agents [13], tested solely in a Sims-inspired interactive environment, yielded generalizable insights. Beyond LLM agents, in the realm of RL agents, notable studies like AlphaGo [1] and AlphaStar [2] focused on specific tasks—Go and StarCraft II, respectively—yet the techniques developed in these works have been proven to be generalizable.

[1] Silver, David, et al. "Mastering the game of Go with deep neural networks and tree search." nature 529.7587 (2016).

[2] Vinyals, Oriol, et al. "Grandmaster level in StarCraft II using multi-agent reinforcement learning." Nature 575.7782 (2019).

Q6: The use of ChatGPT to perform multi-dimensional evaluation needs to be validated against human evaluation.

We appreciate your valuable suggestions. In line with your recommendations, we are currently undertaking a human agreement evaluation akin to the approach used by AlpacaEval (details found here). We aim to publish the results promptly.

We extend our heartfelt gratitude once more for your review. We anticipate further dialogue and collaboration, and are open to any more thoughts you may have to help refine our work.

Q4: These linguistic prompt manipulations are inherently fragile. They depend enormously on the current capabilities of the specific LLMs studied.

In our previous discussion, we emphasized that possessing a capability does not automatically lead to realizing its full potential. Thus, there are primarily two types of work concerning LLMs: (1) expanding the boundaries of LLMs' emergent capabilities (such as [8-10]), and (2) unlocking the full potential of these capabilities (such as [1-7] and ReCon). Both approaches are critically important. To draw an analogy, the first approach is similar to enhancing a child's IQ, while the second is akin to teaching the child advanced thinking strategies to maximize the use of their IQ. It would be incorrect to dismiss the value of these thinking strategies simply because they are contingent on the child's IQ. Similarly, it's unjustified to label the thinking process methods for LLMs as "inherently fragile" and dismiss them as irrelevant just because they depend on the current abilities of LLMs. Rather, as the limits of LLMs' emergent capabilities significantly expand, the importance of fully unlocking the potential of LLMs may increasingly come to the forefront [1-7, 11].

[1] Wei, Jason, et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in Neural Information Processing Systems 35 (2022): 24824-24837.

[2] Kojima, Takeshi, et al. "Large language models are zero-shot reasoners." Advances in neural information processing systems 35 (2022): 22199-22213.

[3] Yao, Shunyu, et al. "Tree of thoughts: Deliberate problem solving with large language models." arXiv preprint arXiv:2305.10601 (2023).

[4] Besta, Maciej, et al. "Graph of thoughts: Solving elaborate problems with large language models." arXiv preprint arXiv:2308.09687 (2023).

[5] Shinn, Noah, et al. "Reflexion: Language agents with verbal reinforcement learning." Thirty-seventh Conference on Neural Information Processing Systems. 2023.

[6] Yao, Shunyu, et al. "React: Synergizing reasoning and acting in language models." arXiv preprint arXiv:2210.03629 (2022).

[7] Wang, Guanzhi, et al. "Voyager: An open-ended embodied agent with large language models." arXiv preprint arXiv:2305.16291 (2023).

[8] Brown, Tom, et al. "Language models are few-shot learners." Advances in neural information processing systems 33 (2020): 1877-1901.

[9] Touvron, Hugo, et al. "Llama: Open and efficient foundation language models." arXiv preprint arXiv:2302.13971 (2023).

[10] Touvron, Hugo, et al. "Llama 2: Open foundation and fine-tuned chat models." arXiv preprint arXiv:2307.09288 (2023).

[11] Wang, Lei, et al. "A survey on large language model based autonomous agents." arXiv preprint arXiv:2308.11432 (2023).

Q5: There is no good reason to think that these approaches will transfer to other future LLMs with different training sets, architectures, or fine-tuned guardrails.

In our study, we evaluated ReCon on two distinct LLMs, ChatGPT and Claude, which differ in "training sets, architectures, and fine-tuned guardrails." We applied identical thinking processes and prompts to both. ReCon's superior outcomes with ChatGPT and Claude, as depicted in Figure 4 of our paper, suggest that ReCon possesses a certain level of generalizability across various LLMs.

Regarding "future LLMs," the thought processes in ReCon, such as recursive thinking and perspective-taking, are also useful in human intelligence. Therefore, we have reason to believe that in the foreseeable future, even if LLMs approach or match human intelligence, the ReCon method we propose will remain effective. The only scenario where ReCon might not work on future LLMs is if the chosen future LLMs have inferior generative abilities.

We are grateful for your thorough review and helpful recommendations, which have contributed significantly to the enhancement of our manuscript.

Q1: There are only 20 games, so I'm not sure how reliable the results are.

Thank you for your valuable question. We've addressed the rationale behind selecting 20 Avalon game rounds in Global Q1 and have provided a statistical test of our results in Global Q2. For more information, please see our Global Responses. We appreciate your interest.

Q2: Explain why the evaluation with GPT-4 makes sense and why human eval is not needed, or provide human eval if necessary.

Thank you for your insightful question. We agree that incorporating human evaluation alongside GPT-4 assessment is beneficial. Consequently, we are now conducting a human agreement evaluation similar to the method employed by AlpacaEval (further information available here). We intend to release the findings as soon as possible.

Q3: 4.2 MULTI-DIMENSIONAL EVALUATION is only done with ChatGPT, so it's not clear how reliable and reproducible the results will continue to be.

Regarding reliability, automated evaluation of LLMs is widely recognized and popular [1]. In our evaluation using GPT-4, we analyzed over 2300 LLM responses, which we believe enhances the evaluation's reliability. Nevertheless, as mentioned in our response to Q2, we acknowledge the value of adding human evaluation to GPT-4 assessment for greater reliability. Therefore, we are conducting a human agreement evaluation using a method similar to AlpacaEval (more details here), and we plan to publish these results promptly.

In terms of reproducibility, while OpenAI's API did not offer deterministic output during our experiments, we maintained reproducibility by setting the GPT-4 evaluator's temperature to 0 and documenting the GPT-4 version in our paper's "Reproducibility Statement." Additionally, evaluating over 2300 LLM responses significantly reduces randomness under the law of large numbers. Hence, we consider our GPT-4 evaluation to be reproducible.

[1] Chang, Yupeng, et al. "A survey on evaluation of large language models." arXiv preprint arXiv:2307.03109 (2023).

Q5: 4.3 QUALITATIVE ANALYSES is quite interesting, but the results are very anecdotal to be reliable. 'ReCon’s Proficiency in Detecting Misinformation' -> this is quite cool and would be a very nice result, but there are only anecdotes to support it and not a complete quantitiative evaluation. Can an actual dataset be collected with misinfo, true info labeled, and ReCon vs other baselines rigorously benchmarked on it? Similarly for 'The efficacy of ReCon in information', another cool finding, but needs much more extensive comparisons and evaluation to be reliable. 'Lai et al., Werewolf among us: Multimodal resources for modeling persuasion behaviors in social deduction games' could be a useful resource here.

Thank you for recognizing the interest in our qualitative analysis, and we're happy to address your question about it below.

Regarding the "anecdotal" aspect of our results, to bolster the reliability of our qualitative analysis, we've compiled extensive data on how LLMs think and speak using ReCon in an Excel file, accessible at this anonymous link. Additionally, upon acceptance of our paper, we plan to open-source our codes, inviting you to test them and assess the reliability of our qualitative results yourself.

We also value your suggestion about compiling an actual dataset. Due to time constraints in this rebuttal phase, we haven’t been able to gather such a dataset and benchmark different methods on it, but we will consider incorporating this into our future work. Thank you once again for your constructive suggestion.

Q6: Explain 4.2 MULTI-DIMENSIONAL EVALUATION, including why only chatgpt was used or provide results with open-source LMs, explain the evaluation metrics and why human eval is not needed, or provide human eval if necessary.

Thank you for your inquiry about our evaluation approach.

Regarding the lack of evaluation with open-source LMs, as detailed in Section 5.5 of our paper, LLaMA2-70b-chat, a prominent open-source LM, often struggles to consistently produce the structured output formats our study requires, like generating outputs suitable for regular expression matching (e.g., [approve]). This issue poses challenges in employing open-source LMs for playing the Avalon game.

For the question on human evaluation, we invite you to refer to our response to Q2 for more information.

We thank you again for your expertise and attention, and are ready to address any further questions or concerns.

Thank you for your response. We have completed the human evaluation assessing the reliability of automatic language model evaluation. For our results, please see Appendix D.1 in the revised manuscript or refer to Global Q3. We appreciate your attention to this matter.

We trust that our comprehensive response has addressed your concerns regarding our work and will positively influence your evaluation. Thank you for your consideration.

Global Q3: How reliable is the GPT-4's automatic evaluation in the multi-dimensional evaluation part?

We assess GPT-4's automatic evaluation depicted in Figure 5 and Figure 6 through human annotations. To do this, we select a random sample of 216 dialogues from those shown in Figure 5 and Figure 6. Each dialogue is classified according to the level of human consensus it receives, with the categories being "total agreement", "majority agreement", "majority disagreement", and "total disagreement". Our annotation team comprises 12 individuals, each responsible for 18 annotations, including 8 men and 4 women, all of whom have familiarity with the Avalon game. The outcomes of these annotations are as follows:

Based on the data presented in the above table, it is evident that the ratio of agreement significantly surpasses that of disagreement. This indicates that in our multi-dimensional evaluation, GPT-4's annotations are predominantly considered reliable.

Furthermore, to assess whether there was a significant difference in the frequency of "agreement" and "disagreement" ratings above, we conduct the Chi-square test. Our null hypothesis stated no difference in these frequencies. We observed 158 agreements and 58 disagreements out of 216 total responses. Under the null hypothesis, we expected equal frequencies for both categories (108 each). We calculated the Chi-square value (23.49) by comparing observed frequencies with expected frequencies. The resulting p-value was about $1.26\times 10^{-6}$, far below the 0.05 threshold for statistical significance. This very low p-value led us to reject the null hypothesis, indicating a statistically significant difference between the counts of "agreement" and "disagreement" ratings.