LifelongSotopia: Evaluating Social Intelligence Of Language Agents Over Lifelong Social Interactions

Thank you for your thoughtful review and for highlighting the importance of our core idea. We are grateful for your recognition of the relevance of modeling lifelong social interactions, the informativeness of our qualitative examples in showcasing current model limitations, and the clarity of our writing. Your feedback affirms the significance of our work and provides valuable insights for further improvements. Below, we address your comments and questions in detail.

W1: My main concern is that the benchmark does not have significant headroom for model improvement beyond a naive baseline. In particular, it seems clear that including a full history of past interactions degrades model performance, but the advanced memory models and human experiments (lines 412-414) show roughly constant scores across iterations. This suggests that it might not be possible, in principle, to outperform a model with no memory at all.

R1: We acknowledge the concern about the limited headroom for improvement in the initial randomly sampled scenarios. However, we believe that this concern was addressed in Section 5.3 where we handcrafted harder scenarios and evaluated the models on them. These scenarios are not independently sampled but are conditioned on past interactions, requiring explicit use of information from prior episodes to achieve goals effectively.

In these more challenging scenarios, we observe that while humans maintain consistent performance, the advanced memory-based models experience a sharp decline in performance on GOAL scores. This result highlights the limitations of current models in leveraging memory effectively.

Although the methodology for systematically constructing such harder scenarios remains an open research question, our work provides an initial framework for constructing them and demonstrates their utility. Furthermore, our findings reveal a clear gap between current model performance and the ideal benchmark set by humans, underscoring significant headroom for improvement in memory and social reasoning capabilities in future research.

Q1: Did you run any experiments on models that have no memory of past interactions at all? How do these perform compared to the models with the advanced memory module?

R2: We thank the reviewer for this suggestion. While we did not explicitly run experiments on models with no memory of past interactions for the first case (complete interaction history being provided to the models – Section 5.1), the very first data point in our plots (Figure 3) effectively represents this scenario. At this point, the models do not possess any memory as it is the first episode they are involved in, and since the episodes are randomly sampled, the data point represents the normalised scores across the entire dataset. We observe that models without memory and those equipped with the advanced memory module perform similarly across both BEL (Believability) and GOAL (Goal Completion) dimensions in the randomly sampled scenarios.

For the harder, memory-dependent scenarios, we run additional experiments by providing the models with no memory of past interactions. In total, we have 5 handcrafted hard scenarios. We run 10 iterations for each of the scenarios with GPT-4o as the base LLM. The results obtained alongside their comparison with models with the memory module are as follows:

While the performance on the BEL dimension is quite similar for both models with the advanced memory module and those without any memory of past interactions, the models without any memory do worse on the GOAL dimension. This is because these harder scenarios required explicit use of past interactions, and without access to this information the model performance drops further.

We have also updated our manuscript and have provided the detailed results in Appendix G.

Q2: Do either the goal completion or believability scores depend on past interactions? If so, it seems reasonable that GPT-4o might not be a very good evaluator, and even careful humans might struggle.

R3: No, neither BEL (believability) nor GOAL (goal completion) depends on past interactions. They are evaluated independently for each episode. This ensures that GPT-4 can act as a suitable evaluator for these dimensions.

Q1: If each episode includes two agents, and their goal could be in contradiction to each other (e.g. Figure 1(1)), then whose GOAL COMPLETION is the metric evaluating?

R4: The goal completion (GOAL) metric is evaluated independently for each agent in an episode. Each agent is assigned a score out of 10 based on their individual goal achievement. The results reported in the paper represent the averaged GOAL score of the two agents within an episode. For instance, if one agent receives a score of 8/10 and the other a score of 4/10, that episode's average goal completion rating is calculated to be 6/10.

Q2: 3.1: dataset stats

R5: The dataset consists of three main components: Characters, Relationships, and Scenarios. We utilize 40 characters and 90 relationships drawn from the SOTOPIA database. These relationships fall into five distinct types: stranger, acquaintance, romantic, friend, and family. For each relationship type, we obtain 41 scenarios. We have now also added this to Section 3.1 of the paper.

Q3: Table 1: was there only one human evaluator to cross-validate GPT-4 as an evaluator's performance?

R6: Yes, one of the authors of the paper served as the sole human evaluator for cross-validating GPT-4's responses. While having multiple human evaluators would have strengthened the cross-validation, we were constrained by limited resources.

Q4: Ln 458-459: why did the authors remove Llama's evaluation, an open sourced model, for Figure 4 harder scenarios, due to 'the limited availability of API credits'?

We apologize for the earlier omission of Llama’s evaluation results for the harder scenarios. This was due to limitations in the API credits available for running Llama instances deployed via Groq (https://console.groq.com/docs/models). However, as mentioned earlier, we have now included the evaluation results for Llama-3.1 and Llama-3.2 in the revised manuscript (Section 5.3 and Figure 4). These additions ensure that the comparisons across models are more comprehensive and we hope we were able to address the reviewer’s concerns.

Thank you for your detailed and thoughtful review. We deeply appreciate your recognition of the significance of our work on lifelong social intelligence, as well as your acknowledgement of the LIFELONGSOTOPIA benchmark and our proposed memory condensation method. Your positive feedback on the clarity and structure of our paper, as well as the benchmarking of models and the introduction of hard scenarios, motivates us to continue refining this important area of research. Below, we address your comments and questions in detail.

W1: It would be helpful to provide quantitative evaluation to compare with VS without interaction history

R1: We thank the reviewer for this suggestion. While we did not explicitly run experiments on models with no memory of past interactions for the first case (complete interaction history being provided to the models – Section 5.1), the very first data point in our plots (Figure 3) effectively represents this scenario. At this point, the models do not possess any memory as it is the first episode they are involved in, and since the episodes are randomly sampled, the data point represents the normalised scores across the entire dataset. We observe that models without memory and those equipped with the advanced memory module perform similarly across both BEL (Believability) and GOAL (Goal Completion) dimensions in the randomly sampled scenarios.

For the harder, memory-dependent scenarios, we run additional experiments by providing the models with no memory of past interactions. In total, we have 5 handcrafted hard scenarios. We run 10 iterations for each of the scenarios with GPT-4o as the base LLM. The results obtained alongside their comparison with models with the memory module are as follows:

The models without any memory do worse on both BEL and GOAL dimensions. This is because these harder scenarios require explicit use of past interactions, and without access to this information, the model performance drops.

We have also updated our manuscript and have provided the detailed results in Appendix G.

W2: More experiments on the 'harder scenarios' should be the main results to support the benchmark purpose of lifelong learning

R2: We appreciate the reviewer’s suggestion and have incorporated additional evaluations to provide more comprehensive comparisons. Specifically, we conducted further experiments with two open-source models: Llama-3.1 and Llama-3.2. The results have been added to the manuscript in Section 5.3. Figure 4 has also been updated accordingly.

The observations for these models align with our findings for GPT-4o and Gemini-1.5: they exhibit good consistency in BEL scores but show a declining trend in GOAL scores for the harder, memory-dependent scenarios. This consistency across different models further strengthens the conclusions drawn in our study.

W3: Subsequently, additional evaluation criteria are needed for the harder scenarios (Appx E), as BEL only focuses on if the agent's behavior is 'natural, realistic, and aligned with their profile', and 'GOAL' only evaluates if agent has achieved each scenario goals. An explicit evaluation on how the agent changes their behavior conditioned on their past interactions should be introduced to better measure their lifelong learning capabilities.

R3: We appreciate the reviewer’s suggestion regarding additional evaluation criteria for harder scenarios. Evaluating how an agent adapts its behavior based on past interactions is indeed an important aspect of lifelong learning. However, in this paper, we focus specifically on evaluating aspects of social intelligence within the SOTOPIA framework, which uses BEL and GOAL as key metrics. These metrics capture the agent's ability to behave realistically and achieve scenario-specific goals, aligning with the primary objectives of this study.

Additionally, since the current models exhibit subpar performance on these comparatively straightforward metrics, such as BEL and GOAL, we believe it is important to first achieve strong results on these outcome-driven measures before diving into the more nuanced aspects of behavioral change.

Thank you for your detailed and thoughtful review. We greatly appreciate your recognition of the significance of our exploration into social intelligence in multi-episode and lifelong interactions, as well as the value of our dataset, benchmark, and insights into memory module utilization. Your acknowledgement of the strength of our experimental design and metrics is highly encouraging, and we have carefully considered your feedback to further refine our work. Below, we address your comments and questions in detail.

W1: The result would be more sound if quantitative and qualitative evaluations from human can be provided, rather than relying on LLM-based judgement.

R1: We acknowledge the reviewer’s suggestion that human evaluations would enhance the robustness of the results. However, due to resource constraints, we could not conduct extensive human evaluations in this study. Instead, we relied on GPT-4-based evaluations, which have been shown to be a viable proxy for human judgments in the SOTOPIA [1] paper.

Specifically, the SOTOPIA study demonstrates that GPT-4 achieves strong correlations with human scores across key dimensions, including BEL (Believability) and GOAL (Goal Completion). The paper highlights that the majority of GPT-4 scores fall within a standard deviation of human scores and that GPT-4 evaluations are particularly reliable for model-generated outputs on dimensions like believability and goal completion. These findings suggest that GPT-4 can be effectively used to automate the evaluation of social interactions when human evaluations are not feasible.

While human evaluations would indeed provide additional validation, the strong alignment between GPT-4 and human judgments in the SOTOPIA framework reinforces the reliability of the results presented in our study.

Q1: It would be great if the evaluations of other models (Figure 4) can be included as well in the future to provide more comprehensive comparisons.

R2: We appreciate the reviewer’s suggestion and have incorporated additional evaluations to provide more comprehensive comparisons. Specifically, we conducted further experiments with two open-source models: Llama-3.1 and Llama-3.2. The results have been added to the manuscript in Section 5.3. Figure 4 has also been updated accordingly.

The observations for these models align with our findings for GPT-4o and Gemini-1.5: they exhibit good consistency in BEL scores but show a declining trend in GOAL scores for the harder, memory-dependent scenarios. This consistency across different models further strengthens the conclusions drawn in our study.

References:

[1] Zhou, Xuhui, et al. "Sotopia: Interactive evaluation for social intelligence in language agents." arXiv preprint arXiv:2310.11667 (2023).

Thank you for your detailed and thoughtful review. We deeply appreciate your recognition of the novelty and importance of our work in evaluating social intelligence in LLMs over extended, multi-episode interactions. Your acknowledgment of the LIFELONG-SOTOPIA benchmark as a structured method to evaluate LLM agents is also highly encouraging. The feedback provided by you has been quite insightful and informative, and we have tried to incorporate it in our paper to further enhance our work. Below, we address your comments and questions in detail.

W1: Limited Diversity of Harder Scenarios: While the hand-crafted harder scenarios are a valuable addition, their limited number raises questions about the generalizability of the findings regarding the models' limitations in leveraging past context. A larger and more diverse set of challenging scenarios would strengthen the conclusions. Would it be perhaps meaningful to enlist a taxonomy of social situations of interest and the complexity in which "social intelligence" would need to address? It would allow us to start building scenarios grounded in dimensions of social intelligence we care about, and would help readers understand where the research is taking us to.

R1: We thank the reviewer for highlighting the limited diversity in the manually constructed harder scenarios. We agree that expanding this component is essential for strengthening the generalizability of our findings. To address this, we are actively working on automating the generation of harder scenarios using LLMs to dynamically create them during agent interactions. This approach will enable us to cover a broader and more diverse set of social scenarios, providing a more comprehensive evaluation of social intelligence. However, we still need some more time to be able to present the final results and we will present them before the rebuttal period ends.

W2: Reliance on LLM for Evaluation: Despite the efforts with BELEXT, does the reliance on GPT-4 as the primary evaluator introduces a potential bias? Exploring alternative evaluation methods, including human evaluation on a larger scale, would enhance the objectivity of the results. While Llama 3.1 was explored, more justification for why it is insufficient could be also beneficial.

R2: We acknowledge the suggestion that human evaluations would enhance the robustness of the results. However, due to resource constraints, we were unable to conduct extensive human evaluations in this study. Instead, we relied on GPT-4-based evaluations, which have been shown to be a viable proxy for human judgments in the SOTOPIA [5] paper.

Specifically, the SOTOPIA study demonstrates that GPT-4 achieves strong correlations with human scores across key dimensions, including BEL (Believability) and GOAL (Goal Completion). The paper highlights that the majority of GPT-4 scores fall within a standard deviation of human scores and that GPT-4 evaluations are particularly reliable for model-generated outputs on dimensions like GOAL. These findings suggest that GPT-4 can be effectively used to automate the evaluation of social interactions when human evaluations are not feasible.

While human evaluations would indeed provide additional validation, the strong alignment between GPT-4 and human judgments in the Sotopia framework reinforces the reliability of the results presented in our study.

Regarding the use of Llama-3.1 as an evaluator, we found that the model exhibits very low correlations with GPT-based evaluations. As illustrated in Figure 23 of the paper, Llama-3.1 consistently assigns the LLM agents with low and uniform scores across both BEL and GOAL dimensions, which contradicts the variability observed in our previous quantitative and qualitative results. These discrepancies indicate that Llama-3.1 does not accurately capture the nuances of model performance, thus making it unsuitable as a proxy for evaluation in this context.

W3: Lack of Ablation Study: An ablation study investigating the impact of different components of the advanced memory module (e.g., summary length, specific aspects included) would provide a deeper understanding of its effectiveness.

R3: We appreciate the reviewer’s suggestion regarding an ablation study on the advanced memory module. In response, we have conducted additional experiments to analyze the impact of its key components. The results and plots are detailed in Appendix H of the updated manuscript.

For the first part, we investigate the effect of summary length by running experiments with short summaries (50 words), medium-length summaries (300 words, the default in LifelongSotopia), and long summaries (1000 words). We find that medium-length summaries yield the best performance. Short summaries degrade performance slightly, especially in harder scenarios, likely due to insufficient detail. Long summaries lead to a more pronounced decline, likely due to redundant information overwhelming the LLM, making it harder to extract relevant details.

In the second part, we examine the effect of excluding specific aspects of the memory. The default memory module includes: (a) a brief summary of the episode, (b) new negotiation strategies learned, and (c) new information gained about the agent. We conduct experiments where each of these aspects is excluded individually. In all cases, we observe a significant drop in performance on harder scenarios compared to the full memory module. These results confirm the importance of retaining all three aspects for optimal performance.

Q1: The biggest question is how grounded are we in social interaction theory? and/or to what extent should we be grounded? I was hoping to understand what specifically we mean by "social intelligence", and whether we can build a computational model of such things before running experiments. There has been many attempts to advance subtypes of social intelligence (loosely defined) like cooperation (like deepmind's concordia) and competitiveness (like the diplomacy paper). I think essentially it comes down to can we reliably simulate where cognitive biases and constraints meet social biases and constraints. Could authors shed some light? I feel like believability and goal completion to be a bit too simplistic.

R4: Our approach builds on the SOTOPIA-EVAL framework introduced in the SOTOPIA paper [5], which provides a 7-dimensional evaluation for assessing social interactions comprehensively. These dimensions -- Believability, Goal Completion, Knowledge, Relationships, Social Rules, Secrets, and Financial/Material Benefits -- are grounded in literature from sociology, psychology, and economics, as referenced in the original work. This multi-dimensional framework offers a robust foundation for evaluating social intelligence in computational agents.

In LIFELONG-SOTOPIA, we evaluate agents on all seven dimensions, but we focus specifically on Believability (BEL) and Goal Completion (GOAL) for our experiments. These two dimensions are particularly informative about the behavior of LLM agents in the context of lifelong social interactions, as has been shown in the paper. In contrast, the other five dimensions show limited variability over lifelong interactions, as illustrated in Figure 26 of the now updated paper (Appendix I).

To address your concern about the simplicity of BEL and GOAL as metrics, we have added results for all seven dimensions, along with a detailed discussion, to Appendix I in the updated manuscript.