Communication and Verification in LLM Agents towards Collaboration under Information Asymmetry

TL;DR

This work investigates how LLM agents can effectively collaborate under information asymmetry by adapting Einstein’s Puzzle to a two-player tabletop setting. It introduces a fine-tuning–plus–environment-based verifier framework and systematically studies communication strategies (information seeking and providing) across configurations, demonstrating that aligned bidirectional communication substantially improves task success and interpretability. An environment-based verifier further enhances rule understanding without additional training, while a human study reveals a preference for proactive information sharing despite potential inefficiencies. The results suggest a general, training-free verifier approach that can extend to other simulated environments to promote safer and more interpretable AI collaboration.

Abstract

While Large Language Model (LLM) agents are often approached from the angle of action planning/generation to accomplish a goal (e.g., given by language descriptions), their abilities to collaborate with each other to achieve a joint goal are not well explored. To address this limitation, this paper studies LLM agents in task collaboration, particularly under the condition of information asymmetry, where agents have disparities in their knowledge and skills and need to work together to complete a shared task. We extend Einstein Puzzles, a classical symbolic puzzle, to a table-top game. In this game, two LLM agents must reason, communicate, and act to satisfy spatial and relational constraints required to solve the puzzle. We apply a fine-tuning-plus-verifier framework in which LLM agents are equipped with various communication strategies and verification signals from the environment. Empirical results highlight the critical importance of aligned communication, especially when agents possess both information-seeking and -providing capabilities. Interestingly, agents without communication can still achieve high task performance; however, further analysis reveals a lack of true rule understanding and lower trust from human evaluators. Instead, by integrating an environment-based verifier, we enhance agents' ability to comprehend task rules and complete tasks, promoting both safer and more interpretable collaboration in AI systems. https://github.com/Roihn/EinsteinPuzzles

Figures (11)

• Figure 1: Illustration of our collaborative game. Each game features a final goal (top-left), where objects are assigned to specific goal bins (e.g. toy bear to top-left bin). Placement constraints are generated based on this final goal and are distributed to the two players. At the start, objects are randomly positioned in front of the players. Players must collaborate and communicate effectively to reason the final goal and accordingly complete the placement.
• Figure 2: Distributions of answers with 108 data points from 12 human participants for each question. Each participant plays 9 games with one model. The white dots represent the average scores.
• Figure 3: Full game playthrough with actions and rationales. The game begins with Player 1, and the two players take turns performing physical moves, sharing information, or asking questions until all objects are correctly placed.
• Figure 4: Illustration of three types of verifications we consider. Following the same setup as we showed in Figure 1, the game environment supports providing feedbacks related to action affordance, communication and strategies, which can be directly used as verifiers for agents' decisions.
• Figure 5: A visual overview of the collaborative game setting. Two players (you and the AI agent) work together to place objects into goal bins based on a set of relational constraints.