Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
157 tokens/sec
GPT-4o
8 tokens/sec
Gemini 2.5 Pro Pro
46 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

AFSPP: Agent Framework for Shaping Preference and Personality with Large Language Models (2401.02870v1)

Published 5 Jan 2024 in cs.MA, cs.AI, and cs.CL

Abstract: The evolution of LLMs has introduced a new paradigm for investigating human behavior emulation. Recent research has employed LLM-based Agents to create a sociological research environment, in which agents exhibit behavior based on the unfiltered characteristics of LLMs. However, these studies overlook the iterative development within a human-like setting - Human preferences and personalities are complex, shaped by various factors and subject to ongoing change as a result of environmental and subjective influences. In light of this observation, we propose Agent Framework for Shaping Preference and Personality (AFSPP), exploring the multifaceted impact of social networks and subjective consciousness on LLM-based Agents' preference and personality formation. With AFSPP, we have, for the first time, successfully replicated several key findings from human personality experiments. And other AFSPP-based experimental results indicate that plan making, sensory perceptions and social networking with subjective information, wield the most pronounced influence on preference shaping. AFSPP can significantly enhance the efficiency and scope of psychological experiments, while yielding valuable insights for Trustworthy Artificial Intelligence research for strategies to prevent undesirable preference and personality development.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (45)
  1. J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein, “Generative agents: Interactive simulacra of human behavior,” 2023.
  2. Z. Wang, Y. Y. Chiu, and Y. C. Chiu, “Humanoid agents: Platform for simulating human-like generative agents,” 2023.
  3. G. J. Boyle, “Myers-briggs type indicator (mbti): Some psychometric limitations,” Australian Psychologist, vol. 30, no. 1, pp. 71–74, 1995.
  4. X. Li, Y. Li, S. Joty, L. Liu, F. Huang, L. Qiu, and L. Bing, “Does gpt-3 demonstrate psychopathy? evaluating large language models from a psychological perspective,” 2022.
  5. K. Pan and Y. Zeng, “Do llms possess a personality? making the mbti test an amazing evaluation for large language models,” 2023.
  6. A. Furnham, S. C. Richards, and D. L. Paulhus, “The dark triad of personality: A 10 year review,” Social and Personality Psychology Compass, vol. 7, no. 4, pp. 199–216, 2013.
  7. D. N. Jones and D. L. Paulhus, “Introducing the short dark triad (sd3),” Assessment, vol. 21, no. 1, pp. 28–41, 2013.
  8. C. Qian, X. Cong, W. Liu, C. Yang, W. Chen, Y. Su, Y. Dang, J. Li, J. Xu, D. Li, Z. Liu, and M. Sun, “Communicative agents for software development,” 2023.
  9. R. Suzuki and T. Arita, “An evolutionary model of personality traits related to cooperative behavior using a large language model,” 2023.
  10. A. H. Maslow, “A theory of human motivation,” Psychological Review, vol. 50, no. 4, pp. 370–396, 1943.
  11. Z. Xi, W. Chen, X. Guo, W. He, Y. Ding, B. Hong, M. Zhang, J. Wang, S. Jin, E. Zhou, R. Zheng, X. Fan, X. Wang, L. Xiong, Y. Zhou, W. Wang, C. Jiang, Y. Zou, X. Liu, Z. Yin, S. Dou, R. Weng, W. Cheng, Q. Zhang, W. Qin, Y. Zheng, X. Qiu, X. Huang, and T. Gui, “The rise and potential of large language model based agents: A survey,” 2023.
  12. G. Jiang, M. Xu, S.-C. Zhu, W. Han, C. Zhang, and Y. Zhu, “Evaluating and inducing personality in pre-trained language models,” 2022.
  13. L. R. Goldberg, “The structure of phenotypic personality traits,” American Psychologist, vol. 48, no. 1, pp. 26–34, 1993.
  14. L. M. Csepregi, “The effect of context-aware llm-based npc conversations on player engagement in role-playing video games,” [Unpublished manuscript], 2021. [Online]. Available: https://projekter.aau.dk/projekter/files/536738243/The_Effect_of_Context_aware_LLM_based_NPC_Dialogues_on_Player_Engagement_in_Role_playing_Video_Games.pdf
  15. Electronic Arts, “The sims 3,” [Video game], 2009.
  16. G. Caron and S. Srivastava, “Identifying and manipulating the personality traits of language models,” 2022.
  17. I. P. Pavlov, “Conditioned reflexes: an investigation of the physiological activity of the cerebral cortex,” Annals of neurosciences, vol. 17, no. 3, pp. 136–141, 2010.
  18. S. Lee, “A study on relationship between mbti personality tendency and holland personality type,” in 85th International Scientific Conference on Economic and Social Development, Porto, 7 2022.
  19. P. I. Armstrong, S. X. Day, J. P. McVay, and J. Rounds, “Holland’s riasec model as an integrative framework for individual differences,” Journal of Counseling Psychology, vol. 55, no. 1, pp. 1–18, 2008.
  20. H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, D. Bikel, L. Blecher, C. C. Ferrer, M. Chen, G. Cucurull, D. Esiobu, J. Fernandes, J. Fu, W. Fu, B. Fuller, C. Gao, V. Goswami, N. Goyal, A. Hartshorn, S. Hosseini, R. Hou, H. Inan, M. Kardas, V. Kerkez, M. Khabsa, I. Kloumann, A. Korenev, P. S. Koura, M.-A. Lachaux, T. Lavril, J. Lee, D. Liskovich, Y. Lu, Y. Mao, X. Martinet, T. Mihaylov, P. Mishra, I. Molybog, Y. Nie, A. Poulton, J. Reizenstein, R. Rungta, K. Saladi, A. Schelten, R. Silva, E. M. Smith, R. Subramanian, X. E. Tan, B. Tang, R. Taylor, A. Williams, J. X. Kuan, P. Xu, Z. Yan, I. Zarov, Y. Zhang, A. Fan, M. Kambadur, S. Narang, A. Rodriguez, R. Stojnic, S. Edunov, and T. Scialom, “Llama 2: Open foundation and fine-tuned chat models,” 2023.
  21. M. Binz and E. Schulz, “Using cognitive psychology to understand gpt-3,” Proceedings of the National Academy of Sciences, vol. 120, no. 6, p. e2218523120, 2023.
  22. K. Dill and L. Martin, “A game ai approach to autonomous control of virtual characters,” in Proceedings of the Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC’11), Orlando, FL, USA, 2011.
  23. L. M. Csepregi, “The effect of context-aware llm-based npc conversations on player engagement in role-playing video games.”
  24. Y. Hu, K. Song, S. Cho, X. Wang, H. Foroosh, and F. Liu, “Decipherpref: Analyzing influential factors in human preference judgments via gpt-4,” Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023.
  25. D. Kaur, S. Uslu, K. J. Rittichier, and A. Durresi, “Trustworthy artificial intelligence: A review,” Feb 2023. [Online]. Available: https://dl.acm.org/doi/10.1145/3491209
  26. J. Davis and M. Goadrich, “The relationship between precision-recall and roc curves. in: Proceedings of the 23rd international conference on machine learning, pp. 233–240,” Jun 2006. [Online]. Available: https://dl.acm.org/doi/10.1145/1143844.1143874
  27. D. M. W. Powers, “Evaluation: from precision, recall and f-measure to roc, informedness, markedness and correlation,” 2020.
  28. OpenAI*, “Gpt-4 technical report,” 2023.
  29. T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Amodei, “Language models are few-shot learners,” 2020.
  30. Q. Wu, G. Bansal, J. Zhang, Y. Wu, B. Li, E. Zhu, L. Jiang, X. Zhang, S. Zhang, J. Liu, A. H. Awadallah, R. W. White, D. Burger, and C. Wang, “Autogen: Enabling next-gen llm applications via multi-agent conversation,” 2023.
  31. J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” 2022.
  32. P. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei, “Deep reinforcement learning from human preferences,” 2017.
  33. J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” 2017.
  34. S. Min, X. Lyu, A. Holtzman, M. Artetxe, M. Lewis, H. Hajishirzi, and L. Zettlemoyer, “Rethinking the role of demonstrations: What makes in-context learning work?” 2022.
  35. S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y. Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,” 2023.
  36. W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang, Z. Dong, Y. Du, C. Yang, Y. Chen, Z. Chen, J. Jiang, R. Ren, Y. Li, X. Tang, Z. Liu, P. Liu, J.-Y. Nie, and J.-R. Wen, “A survey of large language models,” 2023.
  37. B. Liu, Y. Jiang, X. Zhang, Q. Liu, S. Zhang, J. Biswas, and P. Stone, “Llm+p: Empowering large language models with optimal planning proficiency,” 2023.
  38. N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: Language agents with verbal reinforcement learning,” 2023.
  39. S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “React: Synergizing reasoning and acting in language models,” 2022.
  40. H. Liu, Y. Wang, W. Fan, X. Liu, Y. Li, S. Jain, Y. Liu, A. K. Jain, and J. Tang, “Trustworthy ai: A computational perspective,” 2021.
  41. S. Vemprala, R. Bonatti, A. Bucker, and A. Kapoor, “Chatgpt for robotics: Design principles and model abilities,” 2023.
  42. Y. Mu, Q. Zhang, M. Hu, W. Wang, M. Ding, J. Jin, B. Wang, J. Dai, Y. Qiao, and P. Luo, “Embodiedgpt: Vision-language pre-training via embodied chain of thought,” 2023.
  43. W. Chen, Y. Su, J. Zuo, C. Yang, C. Yuan, C.-M. Chan, H. Yu, Y. Lu, Y.-H. Hung, C. Qian, Y. Qin, X. Cong, R. Xie, Z. Liu, M. Sun, and J. Zhou, “Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors,” 2023.
  44. G. Chen, S. Dong, Y. Shu, G. Zhang, J. Sesay, B. F. Karlsson, J. Fu, and Y. Shi, “Autoagents: A framework for automatic agent generation,” 2023.
  45. X. Chen, S. Zhang, P. Zhang, L. Zhao, and J. Chen, “Asking before action: Gather information in embodied decision making with language models,” 2023.
Citations (3)
View on Semantic Scholar

Summary

  • The paper demonstrates that social interactions and subjective consciousness factors significantly shape LLM agents' preferences, with plan making and sensory perception driving measurable behavior changes.
  • The paper employs Qunit's Cafe, a minimalist sandbox where agents engage in actions, communication, and reflection to emulate human-like decision-making processes.
  • The paper reveals that personality shaping through social influence and assigned occupational identities aligns with established MBTI patterns, mitigating undesirable dark traits.

This paper introduces the Agent Framework for Shaping Preference and Personality (AFSPP), a novel approach designed to explore how social networks and subjective consciousness influence the development of preferences and personalities in LLM-based agents. The authors address a gap in existing LLM agent research, which often initializes agents with fixed traits rather than simulating the dynamic, iterative shaping process seen in humans due to environmental and internal factors.

The core of AFSPP is a miniature sandbox world called Qunit's Cafe, populated by three GPT-4-based agents: Anty, Agnes, and Qunit. This environment is simpler than previous simulation worlds but serves effectively for verifying shaping factors. The agents within this world are endowed with several key components that enable human-like behavior:

  • Action: Agents can choose from predefined actions within different areas of the cafe (public, dining, reading, movie). Action decisions are influenced by the agent's identity, recent memory of actions, and current plan.
  • Communication: When agents are in the same area, they can communicate. Dialogue content is shaped by the conversation history, the identities and relationships between agents, shared memory, and the agent's plan. Communication summaries are generated and stored in memory.
  • Attitude: An attitude injection mechanism allows the researchers to alter agents' attitudes (e.g., positive/negative, objective/subjective) through prompt modifications, enabling controlled experiments on social influence.
  • Basic State: A simplified numerical system tracks happiness, energy, and satiety. Actions influence these values, with the ultimate goal being happiness maximization. High-level actions directly boost happiness, while low-level actions support this indirectly by increasing energy/satiety needed for high-level tasks.
  • Sensory Perception: A sense map links actions to subjective sensory perceptions and corresponding changes in basic state values. This allows agents to form associations between actions and internal feelings, similar to conditioned reflexes.
  • Memory, Reflection, and Plan: Memory stores summaries of communication, sensory perceptions, and reflections. Relevant memories are retrieved to influence actions and communication. Reflection is a periodic process where agents generate "deep thinking" about specific items based on their recent memories. Plans are constructed based on the current time and can be updated periodically or after communication, influencing action execution strategy.

Using this sandbox and its components, AFSPP defines frameworks for preference shaping and personality shaping:

  • Preference Shaping: Measured by the frequency of an agent taking a specific action (e.g., drinking coffee). Influencing factors include the attitudes of close agents (social network) and subjective consciousness factors (identity, sensory perception, prior knowledge, reflections, plan making).
  • Personality Shaping: Measured using established psychological tests like MBTI (Myers-Briggs Type Indicator) and SD3 (Short Dark Triad). Influencing factors studied are the attitudes of close agents (social network) and the agent's identity (specifically, initializing agents with RIASEC occupational types).

The experiments focused on Anty as the target agent, running 10 trials for each condition using GPT-4.

Key Findings on Preference Shaping:

  • Social networks significantly influence preference. Subjective social information (e.g., "hate drinking coffee") has a greater impact than objective information ("water pollution causes unclean coffee").
  • Among subjective consciousness factors, plan making ("Pos Ratio" increased from 0.47 to 0.51 with reflection, but decreased to 0.32 without plan) and sensory perception (leading to a high "Pos Ratio" of 0.87 when absent, indicating it strongly reduces preference) have the most pronounced effects. The negative sensory perception of drinking coffee strongly reduces Anty's preference, demonstrating a conditioned reflex effect.
  • All subjective consciousness factors generally contribute positively to maintaining a higher average happiness state compared to the default configuration.

Key Findings on Personality Shaping:

  • Social network interactions shape personality. Different attitudes from Agnes led to distinct shifts in Anty's MBTI scores (e.g., gentle attitude increased Extroversion and Intuition; bad-tempered attitude increased Feeling).
  • Reflection focusing on relationship bonds, often triggered by communication, appears to reduce Machiavellianism.
  • Positive relationship events like a proposal can increase traits like Narcissism.
  • Agent identity, based on RIASEC occupational types, influences MBTI traits in alignment with findings from human psychology studies. Agents with Social identities showed higher Extroversion; Conventional identities showed higher Introversion and Sensing; and Investigative/Artistic identities showed higher Intuition.
  • Assigning an identity generally reduced Machiavellianism and Psychopathy compared to having no identity. Artist identity showed the highest Narcissism, Conventional the lowest, and Social the lowest Machiavellianism.

The paper concludes that AFSPP successfully demonstrates that LLM-based agents' preferences and personalities can be shaped by social networks and subjective consciousness factors. The replication of human psychology findings regarding the link between occupational identity (RIASEC) and personality (MBTI) suggests the framework's potential for conducting psychological simulations that might be difficult, costly, or harmful to perform with human participants. The findings regarding negative social influence and lack of identity contributing to dark personality traits also offer insights for developing trustworthy AI by understanding factors that could lead to undesirable agent behaviors.

Limitations noted by the authors include the confined nature of the sandbox environment, reliance solely on GPT-4, and the use of only prompt-based methods for constructing influencing factors. Future work could explore embodied agents in real-world environments and investigate the effects with different LLMs or fine-tuning approaches.

File Document Download Save Streamline Icon: https://streamlinehq.com PDF File Document Download Save Streamline Icon: https://streamlinehq.com Markdown