Simulating Human-like Daily Activities with Desire-driven Autonomy

The authors have addressed my concerns appropriately, and I have updated my score accordingly. I appreciate the authors' efforts in revising the manuscript and responding to the feedback thoroughly.

W5: Limited qualitative evaluation of generated action sequences.

To address the fifth issue, we conducted some statistical analysis to evaluate the diversity and balance of actions generated by different methods. Specifically, we first visualized the activity sequences obtained from the indoor and outdoor multi-agent scenarios using point cloud plots. Next, we defined several statistical labels to categorize and count the actions generated by each agent in the outdoor scenario.

The word cloud figures (see Appendix D.1) emphasize that our framework is capable of exploring a broader range of actions in environments with diverse action selection spaces. For example, in outdoor settings, the word cloud reveals that our model considers behaviors spanning various dimensions, whereas other models tend to focus on a narrower set of actions within these dimensions. Specifically, our model engages in activities such as painting, walking, taking photos, and meditating to fulfill specific desires. In contrast, other models rely heavily on motivation-driven action selection, often repeatedly choosing behaviors that strictly align with their profiles, thereby limiting the diversity of their actions.

The label-counting results showed that D2A-generated actions were more evenly distributed across categories, whereas the baselines displayed more extreme distributions—some with a heavy concentration in certain categories and others completely lacking actions in specific categories. This analysis highlights that our method can generate more diverse and balanced action sequences, making it comparatively more reasonable and adaptable than methods with extreme or unbalanced distributions.

Additionally, we have included a case study in Appendix C that analyzes how desires guide the generation of actions, providing a detailed explanation of the underlying decision-making process. Meanwhile, Appendix W and X presents more examples of action sequences generated by D2A across various scenarios, offering a more intuitive understanding of the results and showcasing the diversity and quality of the generated activities.

W3: Narrow focus on physiological desires in a single-agent setting.

R3: To address this concern, we have expanded our experiments to include scenarios that explore higher-level motivations and multi-agent interactions, showcasing the versatility and broader applicability of our framework.

To better evaluate higher-level motivations, we introduced an outdoor park party scenario (see Appendix A.2), where agents interact in a rich, multi-agent social environment. In this scenario, we updated some of the desire dimensions to focus on higher-level motivations. The updated desire dimensions include hunger, joyfulness, passion, social connectivity, health, spiritual satisfaction, comfort, cleanliness, sleepiness, safety, thirst, recognition, sense of control, and sense of superiority. These refined dimensions allowed us to evaluate the agent’s ability to generate activities that align with complex and socially rich motivations, demonstrating the flexibility of our framework in adapting to different settings and needs.

Additionally, to address the single-agent limitation, the outdoor park party scenario features multi-agent interactions with NPCs (see Appendix A.3), enabling the agent to engage in conversations and collaborative activities. This multi-agent environment provided valuable insights into the framework’s ability to simulate social behaviors and interactions, further broadening the range of human-like activities D2A can generate. Due to time constraints, we were unable to generate a large amount of experimental data in the new environment for a reliable comparison of human likeness. However, we conducted a statistical analysis on some of the action sequences obtained from the indoor and newly added outdoor scenarios. We first visualized the activity sequences obtained from the indoor and outdoor multi-agent scenarios using point cloud plots. Next, we defined several statistical labels to categorize and count the actions generated by each agent in the outdoor scenario. The results of these two statistical experiments demonstrate that our D2A can produce diverse and balanced actions in both indoor and outdoor scenarios. (More details in AppendixD.) The consistent statistical conclusions obtained here, along with the reasonable activity sequences in the multi-agent environment, demonstrate the scalability of our approach and its potential for simulating diverse, dynamic scenarios beyond single-agent settings.

W4: Opaque GPT-4o evaluation methodology.

R4: It was an oversight not to include the prompt used for GPT-4o evaluations in the initial version of the paper; it has now been added to appendix Y. Specifically, GPT-4o evaluates each sequence by assigning scores (out of five) for the three dimensions of naturalness, coherence, and plausibility, and then aggregates these scores to determine which sequence appears more human-like.

To validate the evaluation of GPT-4o, we randomly selected 19 pairs of activity sequences for pairwise comparison (16 pairs from the indoor scenarios and 3 pairs from the newly added outdoor scenarios) and created three questionnaires, each assigned to 8 human annotators. The human annotators were asked to judge which sequence in each pair appeared more human-like. Based on the level of agreement among the annotators, we categorized the 19 samples into three groups:

• High consensus: Over 75% agreement on one side—indicating strong human consensus.
• Moderate preference: Agreement between 50.1%–74.9%—indicating moderate human preference.
• Difficult to distinguish: Exactly 50% agreement—indicating humans found it difficult to decide between the two sequences.

We also provided the same pairs to GPT-4o, using our evaluation prompt and both sequences as context, asking it to determine which sequence appeared more human-like. The consistency rate between human evaluations and GPT-4o’s assessments is summarized in the table below:

Note that for the three pairs in the table where humans found it difficult to distinguish, GPT-4o also outputs a draw result for two of them. This verification demonstrates the robustness of our evaluation framework and confirms that GPT-4o’s assessments are reliable and align closely with human judgments.

Thank you for your detailed and thoughtful review of our work. We appreciate your recognition of the novelty of the D2A framework, particularly its proactive action mechanism based on intrinsic motivations and the human-inspired multi-dimensional desire model. Your comments highlight several areas for improvement, and we have carefully addressed each point below.

W1: Inconsistent application of Maslow’s Theory.

R1: We appreciate your thoughtful comments on the theoretical alignment and desire-action dynamics in our framework. Below, we address these concerns in detail:

• Our framework design is inspired by the theory of needs, where the construction of value dimensions is not strictly bound by a predefined hierarchy such as Maslow’s. While we referenced Maslow’s Theory of Needs as inspiration, we believe there is no universal standard for structuring these dimensions. To demonstrate the flexibility and scalability of our framework, we introduced a new outdoor social scenario in our expanded experiments, where we replaced some of the original desire dimensions with alternative ones tailored to the new setting. The consistent effectiveness of the framework in guiding action generation across these different scenarios underscores its adaptability and robustness.
• We acknowledge that certain simplifications were made in our current implementation. However, the strong performance of our simplified model underscores the validity and importance of our paradigm. It demonstrates that even with these simplifications, the intrinsic motivation framework effectively guides agent behaviors, showcasing the robustness of our approach. Exploring more complex mental hierarchies will be a key focus of our future work. We are confident that incorporating more sophisticated intrinsic motivational architectures, such as hierarchical relationships and non-linear interactions, will not only improve performance but also enhance the realism of agent behaviors.

We improved our paper writing in the revision to avoid such misunderstanding of our motivations.

W2: Overly simplistic desire-action dynamics.

R2: We appreciate your observation regarding the simplicity of the desire-action dynamics in our framework. Similar to our response to the previous point, our current implementation focuses on a simplified dynamic model that prioritizes computational efficiency. Despite its simplicity, the model has demonstrated strong performance in driving diverse and coherent activities, indicating the effectiveness of even a straightforward approach in guiding agent behavior.

At the same time, we fully recognize the potential for enhancing the realism and depth of the framework by incorporating more complex dynamics. Future work will explore advanced modeling techniques, such as non-linear decay, situational adjustments, and time-of-day cycles. We are confident that these enhancements will further improve performance and allow the framework to better capture the nuanced variations in human desires, resulting in even more authentic and human-like behaviors.

W3: Psychological support.

R3: We appreciate your thoughtful question regarding the definition and composition of the intermediate variables, such as human needs, desires, characteristics, and values, used to guide the model’s behavior. Our design is inspired by the theory of needs that needs can drive actions, but we consider that there is no universally agreed-upon standard for defining the conceptual space of needs or values. The 11 dimensions implemented in our work represent a relatively simplified design intended to strike a balance between theoretical inspiration and computational cost. In the newly added outdoor social scenario, we added some of the value dimensions in desire while maintaining their effectiveness in guiding action generation. Specifically, the added dimensions included [recognition,sense of control,sense of superiority]. This demonstrates the flexibility and adaptability of our framework in incorporating alternative designs while preserving its core functionality. In our future work, we aim to explore more generalizable and systematic models of human intrinsic motivation to further refine and enhance the realism and robustness of the framework. The simulating activities, driven by different desire models, can also benefit psychology research, verifying the limitations of existing theories/hypotheses or exploring a more accurate understanding of the desire models.

W4: The experiment is limited (Randomness and prompt design).

R4: We have supplemented new experiments by introducing randomness into the indoor environment. Specifically, we predefined an item list for each room, and during the environment generation process for each experimental round, a certain number of items were randomly selected from each room’s list. These items were then used as context to prompt the LLM to generate a detailed textual representation of the indoor experimental environment.

We conducted experiments in these randomized environments following the same procedures as in Sections 6.3.1 and 6.3.2, and reached consistent conclusions: (1) D2A outperformed other baselines in GPT-4o evaluations by generating more human-like activity sequences. (2) D2A’s action sequences achieved the lowest dissatisfaction value in desire. These findings demonstrate that our framework and design are robust in randomized environments, with no evidence of environmental bias. More details are introduced in Appendix Z. At the same time, we acknowledge the value of broader experimental settings. Beyond the outdoor social environment we discussed earlier, we plan to introduce various new experimental settings in future work to further benchmark agents’ capabilities across diverse scenarios.

As researchers and users of LLMs, we recognize the inherent difficulty in qualitatively assessing the exact impact of prompts in experiments. However, our observations highlight a significant advantage of the D2A framework: incorporating desire descriptions into the prompts effectively guides the agent’s action selection process. This approach stands in contrast to baseline methods that rely heavily on profiles or explicit goals, often resulting in overly concentrated or repetitive actions. By anchoring actions to intrinsic desires, D2A fosters greater diversity and generates more human-like activity sequences, underscoring its robustness and adaptability. More details are discussed in Appendix D.

Thank you for your detailed and thoughtful review. We appreciate your recognition of the originality of our framework, the quality of our writing, and the comprehensive comparative analysis we conducted. Your feedback highlights key areas for improvement, and we have carefully considered each point to enhance our work. Below, we address your concerns and questions in detail.

W1: Limited technological innovation.

R1: We sincerely appreciate your recognition of the conceptual and theoretical innovation behind our desire-driven autonomy framework. However, we respectfully hold a different perspective regarding the concern of Limited Technological Innovation. We believe that research capable of driving paradigm shifts holds far greater impact and importance than work that merely introduces incremental tricks to improve performance by a few points on specific tasks. Our work is dedicated to fostering such a paradigm shift, aiming to demonstrate that leveraging desire-based intrinsic motivations to model agents in open-world scenarios offers significant advantages over traditional modeling approaches. Building on this paradigm shift, the technological innovations you mentioned gain even broader potential for development and application. These innovations are also highly compatible with our framework. For instance, in the components of activity proposal, evaluation, and selection, we integrated the concept of single-layer Monte Carlo Tree Search (MCTS) with the human habit of pre-imagining activity outcomes, seamlessly embedding it within our framework. Looking ahead, we are committed to further exploring technologies that align with and enhance our intrinsic motivation framework. This effort will enable us to continue improving agent performance across diverse open-world tasks, further advancing and solidifying this paradigm shift in both theory and application.

W2: Generalization and scalability to other environments.

R2: To address your concerns about the scalability of our experiments, we added a multi-agent social scenario in an outdoor park party (See Appendix A.2 and A.3). Specifically, we provided a variety of interactive facilities and introduced background agents who could engage in conversations and interactions with the tested autonomous agents. Due to time constraints, we were unable to generate a large amount of experimental data in the new environment for a reliable comparison of human likeness. However, we conducted a statistical analysis on some of the action sequences obtained from the indoor and newly added outdoor scenarios. We first visualized the activity sequences obtained from the indoor and outdoor multi-agent scenarios using word cloud plots. Next, we defined several statistical labels to categorize and count the actions generated by each agent in the outdoor scenario. The results of these two statistical experiments demonstrate that our D2A can produce diverse and balanced actions in both indoor and outdoor scenarios (See more details in AppendixD). The consistent statistical conclusions obtained here, along with the reasonable activity sequences in the multi-agent environment (See Appendix X), demonstrate the generalization and adaptability of our framework across different settings.

Thank you for your time and thoughtful review. Your valuable feedback has been instrumental in helping us enhance and refine our paper. We are pleased to share that the concerns raised by other reviewers have been fully addressed, and they have all raised their scores as a result. If you have any concerns or require further clarification, we would be more than happy to provide additional explanations.

I greatly appreciate the authors' efforts in addressing my concerns, and I have carefully read the revised paper. I am willing to update my score based on the overall quality of the paper and the authors' efforts in revising it. However, to be frank, I still believe that the paper may not be technologically solid beyond designing chains or prompts to query or test LLMs, and it may be challenging to generalize to real-world applications, such as using the desire-driven autonomy framework to effectively understand human needs or assist in planning. Nevertheless, this critique should be directed more towards the topic itself rather than the authors, as they have worked diligently on this topic and provided a substantial contribution.

W4: Methodological flaw in the design of the human reference agent

R4: We have to point out that the methodological flaw mentioned in your fourth weakness stems from a misunderstanding of our experimental design. As described in Section 6.3.2: “The annotators were only provided with the environment description and asked to propose activities for each time step, as if they were living in the house.” We did not provide the annotators with the agent’s current desire states for each dimension, as we intended to collect data resembling real-life actions taken after waking up in the morning. Moreover, it is also difficult to collect data about real human activities at scale in various scenarios for reference. We chose not to directly use real human activity datasets to maintain consistency in the environment.

Q2: Statistics of the generated activities.

R5: Thanks for your suggestions. To give you a more intuitive understanding of the statistical distribution of activities generated by D2A and other baselines, we first visualized the activity sequences obtained from the indoor and outdoor multi-agent scenarios using word cloud plots. Next, we defined several statistical labels to categorize and count the actions generated by each agent in the outdoor scenario. The results of these two statistical experiments demonstrate that our D2A can produce diverse and balanced actions in both indoor and outdoor scenarios. More details in Appendix D.

Q3: Predetermined action space

R6: Finally, we would like to clarify that we did not provide the agents with a specific list of predetermined activities. Instead, the agents are prompted to interact with items available in the environment (strictly monitored by the Game Master). Therefore, if there is any restriction on the action space, it lies in the text-based description of the environment we provided.

Others: I believe this paper should cite...

R7: Thank you for pointing out the articles we should reference. We have included them in the updated revision.

We appreciate your acknowledgment of the originality of our work and the quality of the presentation. We are confident that the following responses will effectively address the concerns raised in your comments.

W1 & Q1: End-goal of the paper.

R1: Thank you for raising the question regarding the importance of simulating and generating human activity sequences. In our paper, we only listed a few applications that have been widely recognized in academia, such as their use in psychology, economics, and sociology research. However, this does not mean the application scope is limited to these fields. On the contrary, the task of simulating human activity's significance and value lies in its immense potential for real-world applications. We fully agree with Reviewer gp3a’s statement: By adopting an intrinsic motivation model, this framework could be used for applications such as social robotics, interactive gaming, and assistive technology, where human-like adaptability and engagement are essential. It may appear that we are merely generating human-like activity sequences; indeed, this task can reflect how well agents behave within a specific environment——Our ultimate objective is to equip agents/robots with the intrinsic motivation we have designed so they can act more rationally and intelligently in real-world scenarios. When applied to assistive technology, this approach may have the potential to enable service agents to better align with human needs and values.

W2: Definition of agents' intelligence.

R2: The discussion above can also serve as our response to your point: “I cannot see how giving agents simulated desires will make them score higher on GSM8K for instance.” We contend that intelligent agents should not be viewed solely as math problem solvers. The manner in which their cognitive processes are modeled and their capacity to robustly select actions is critical yet often challenging to quantify through conventional benchmarks. We argue that these capabilities are often overlooked yet essential indicators of an agent’s intelligence, particularly for general-purpose AI agents.

W3: A lack of Human-judge evaluation.

R3: We appreciate your suggestion to incorporate human-judge evaluations for more convincing results. However, human evaluations are costly and challenging to scale and reproduce, which is why we use GPT-4o for evaluation. To validate the reliability of the GPT-4o evaluation, we additionally conducted a human validation study. We randomly selected 19 pairs of activity sequences for pairwise comparison (16 pairs from the indoor scenarios and 3 pairs from the newly added outdoor scenarios) and created three questionnaires, each assigned to 8 human annotators. The human annotators were asked to judge which sequence in each pair appeared more human-like. Based on the level of agreement among the annotators, we categorized the 19 samples into three groups:

• High consensus: Over 75% agreement on one side—indicating strong human consensus.
• Moderate preference: Agreement between 50.1%–74.9%—indicating moderate human preference.
• Difficult to distinguish: Exactly 50% agreement—indicating humans found it difficult to decide between the two sequences.

We also provided the same pairs to GPT-4o, using our evaluation prompt and both sequences as context, asking it to determine which sequence appeared more human-like. The consistency rate between human evaluations and GPT-4o’s assessments is summarized in the table below:

Note that for the three pairs in the table where humans found it difficult to distinguish, GPT-4o also outputs a draw result for two of them. This verification demonstrates the robustness and generalization of our evaluation framework and confirms that GPT-4o’s assessments are reliable and align closely with human judgments.

Thank you for your thoughtful and detailed review of our work. We appreciate your recognition of the originality and significance of the Desire-driven Autonomous Agent (D2A) framework, as well as your positive feedback on the structure and clarity of the paper. Below, we address the weaknesses and questions you raised:

W1&Q1: Unclear mathematics in Section 3.

R1: We apologize for any confusion caused by the mathematical expressions in our paper, which may not have clearly conveyed the step-by-step generation process. We have made efforts to clarify the mathematical expressions in the revised version of the paper. The formulas in our paper are not strict mathematical formulas or probability distributions but rather a way to illustrate the context or information we provide to the agent and the types of outputs that can be sampled based on that context. Specifically, in Section 3, we aim to uniformly denote the activity generation process as $a_t \sim Agent ( \cdot | a_{0:t-1}, o_{0:t-1}, I, p, e; \theta)$ , which represents generating the activity at the $t$-th step given past activities $a_{0:t-1}$, observations $o_{0:t-1}$, customized information $I$, profile $p$, and environment $e$. We have elaborated on the specific steps involved in D2A’s activity generation process in the later sections. However, due to space constraints, we could not detail the specific steps for all the baseline methods. You can refer to the prompts provided in the appendix for further understanding.

W2&Q2: Reasoning behind the use of the three evaluation dimensions: naturalness, coherence, and plausibility.

R2: Regarding your concerns about the three evaluation criteria—Naturalness, Coherence, and Plausibility—we would like to clarify their basis and rationale. These criteria were developed with reference to Thagard, P. (2002). Coherence in Thought and Action, MIT Press. According to Thagard’s framework, evaluating whether an action sequence appears human-like necessitates examining its alignment with typical human patterns. Specifically, we considered:

• Naturalness: To assess whether the actions conform to typical human abilities and behaviors, avoiding repetitive or physically implausible actions.
• Coherence: To ensure logical and sequential consistency in the flow of actions, avoiding abrupt transitions or contradictions.
• Plausibility: To confirm that actions do not deviate excessively from what would be regarded as normal human behavior.

These three dimensions collectively address critical aspects of evaluating human-like behavior pattern in action sequences. While we acknowledge that other methods may exist, we believe these dimensions provide a comprehensive and practical approach for assessing human-likeness. We have included the specific prompts provided to GPT-4o's evaluation in the Appendix Y for your review.

W3&Q3: Reliability of GPT-4o's evaluation.

R3: Regarding the reliability of GPT-4o’s evaluations, we additionally conducted a human verification experiment to validate the consistency between GPT-4o and human judges. To be specific, we randomly selected 19 pairs of activity sequences for pairwise comparison (16 pairs from the indoor scenarios and 3 pairs from the newly added outdoor scenarios) and created three questionnaires, each assigned to 8 human annotators. The human annotators were asked to judge which sequence in each pair appeared more human-like. Based on the level of agreement among the annotators, we categorized the 19 samples into three groups:

• High consensus: Over 75% agreement on one side—indicating strong human consensus.
• Moderate preference: Agreement between 50.1%–74.9%—indicating moderate human preference.
• Difficult to distinguish: Exactly 50% agreement—indicating humans found it difficult to decide between the two sequences.

We also provided the same pairs to GPT-4o, using our evaluation prompt and both sequences as context, asking it to determine which sequence appeared more human-like. The consistency rate between human evaluations and GPT-4o’s assessments is summarized in the table below:

Note that for the three pairs in the table where humans found it difficult to distinguish, GPT-4o also outputs a draw result for two of them. This verification demonstrates the robustness of our evaluation framework and confirms that GPT-4o’s assessments are reliable and align closely with human judgments.