This paper investigates deliberative large language model (LLM) agents in cooperative joint decision-making tasks characterized by partial and asymmetric observations. The authors formalize this scenario as a problem requiring information exchange through deliberation to achieve a shared reward.

  • Introduces a scalable benchmark instantiating the problem across multiple task settings and domains.
  • Establishes a reference scaffold and evaluation protocol for deliberative agents.
  • Conducts a systematic evaluation of representative LLMs on these tasks.
  • Finds that complex deliberative collaboration continues to challenge state-of-the-art language models.
  • Notes that while external mathematical tools help, models may still fail in alignment or reasoning.
  • Reveals that deliberation allows for reflection and error correction, sometimes improving performance over centralized baselines.

The work establishes a foundation for evaluating and improving LLM agents in deliberative collaboration and provides insights into the strengths and limitations of current multi-agent systems.