Can AI Truly Collaborate?

Author: Denis Avetisyan

New research explores whether advanced AI models can demonstrate understanding and adaptability in collaborative settings, mirroring human interaction.

Across all collaborative interactions - encompassing both agent-agent and human-agent sessions - the relative distribution of behavior types is segmented by individual agent contribution, revealing the proportions of collaborative activity undertaken by Martha, Jeannette, James, Harry, and Eira. — Across all collaborative interactions – encompassing both agent-agent and human-agent sessions – the relative distribution of behavior types is segmented by individual agent contribution, revealing the proportions of collaborative activity undertaken by Martha, Jeannette, James, Harry, and Eira.

This review examines the capacity of foundation models, acting as embodied agents, to exhibit emergent collaborative behaviors and potentially model ‘theory of mind’.

Effective human-AI collaboration hinges on artificial intelligence exhibiting an understanding of human behavior, yet assessing this capacity in embodied agents remains a significant challenge. This research, ‘Evaluating Generative Models as Interactive Emergent Representations of Human-Like Collaborative Behavior’, investigates whether foundation models, operating as collaborative agents in a 2D game, demonstrate emergent behaviors indicative of underlying mental models of their partners. Results reveal that these models consistently exhibit collaborative behaviors-including perspective-taking and collaborator-aware planning-without explicit training, suggesting an innate capacity for coordination. Could these findings pave the way for truly adaptive and intuitive AI collaborators capable of seamless teamwork with humans?

The Foundation of Synergy: Understanding Collaborative Agents

Truly effective collaboration, be it among humans or between people and artificial intelligence, is built upon a foundation of mutual understanding. This extends beyond simply recognizing what another agent is doing, and delves into grasping why – their underlying intentions, beliefs, and the knowledge they possess. Without this capacity to model the internal states of collaborators, interactions become brittle and inefficient, relying on explicit instruction rather than shared context. Successful teams, whether biological or technological, anticipate the needs and likely actions of their members, streamlining processes and allowing for flexible responses to unforeseen challenges. Therefore, fostering this capacity for ‘agent understanding’ is not merely a technical hurdle, but a prerequisite for achieving genuinely synergistic human-AI partnerships.

Current artificial intelligence systems frequently struggle with the subtleties of collaborative endeavors because they often lack a robust capacity for understanding the intentions, beliefs, and knowledge of those they interact with. Unlike humans, who intuitively grasp the ‘why’ behind an action, many AI models operate based on predefined parameters and data patterns, making it difficult to anticipate the needs or reasoning of a partner. This limitation hinders true teamwork; an AI may efficiently execute a task, but it often fails to understand the broader context or adapt to unforeseen circumstances in a collaborative setting. Consequently, the potential for synergistic problem-solving – where combined efforts exceed individual capabilities – remains largely untapped, as these systems struggle with the nuanced give-and-take inherent in genuine collaboration.

To truly unlock collaborative potential, artificial agents require more than just task execution capabilities; they must be equipped with the capacity for perspective-taking and shared planning. This involves developing algorithms that allow an agent to model the beliefs, goals, and knowledge of its collaborators – be they human or artificial. Such an understanding isn’t merely about predicting actions, but about inferring why those actions are taken, allowing the agent to anticipate needs and proactively offer assistance. Shared planning then builds upon this foundation, enabling agents to jointly construct and refine strategies, negotiate compromises, and adapt to changing circumstances as a unified team. Ultimately, imbuing agents with these abilities moves beyond simple assistance and towards genuine, synergistic teamwork, promising significantly enhanced outcomes in complex, collaborative endeavors.

The collaborative 2D game environment features a color-matching puzzle and interactable elements rendered with pixel-art assets generated by the OpenAI [latex]gpt-image-1[/latex] model.

Inner Worlds: Agents and the Capacity for Self-Awareness

Introspection, as applied to multi-agent systems, involves an agent’s capacity to monitor and reason about its own internal states, including beliefs, goals, and reasoning processes. This self-awareness is not simply a passive observation; it enables agents to assess the validity of their current understanding, identify potential conflicts between internal states and external information, and adjust their reasoning accordingly. The capacity for introspection allows an agent to model its own limitations and biases, facilitating more accurate self-assessment and improved decision-making within a collaborative context. Without this internal reflection, agents operate as ‘black boxes’, unable to explain or justify their actions, and thus severely limited in their ability to effectively coordinate with other agents.

A private memory space, or “scratchpad,” functions as a crucial component in agent collaboration by enabling persistent storage of contextual information throughout the task. This allows agents to retain details about the collaborative history, including prior interactions, shared goals, and the evolving state of the environment. Critically, the scratchpad facilitates the tracking of underlying assumptions made during reasoning, allowing for later review and revision as new information becomes available from collaborators. By maintaining this internal record, agents can refine their understanding of the task requirements and ensure consistent, informed decision-making, ultimately improving the efficacy of collaborative efforts.

Collaborator-aware planning involves an agent integrating observed collaborator plans and beliefs into its own decision-making processes. This requires the agent to not simply acknowledge the collaborator’s actions, but to model why those actions are being taken, effectively building a representation of the collaborator’s intentions and knowledge. Successful implementation necessitates maintaining an internal model of the collaborator’s state, allowing the agent to predict future actions and adjust its own plans accordingly. This predictive capability is essential for resolving conflicts, anticipating needs, and coordinating actions effectively within a collaborative framework, ultimately enabling more robust and efficient joint task completion.

The distribution of behavior types-including theory-of-mind, perspective-taking, introspection, collaborator-aware planning, and clarification-varies across individual agents (Eira, Harry, James, Jeannette, and Martha), revealing differing cognitive activity profiles.

Objective Measurement: Analyzing Collaborative Behavior with LLMs

Automated behavioral analysis utilizes large language models (LLMs) to provide objective measurements of collaborative interactions. Traditional methods of assessing collaboration often rely on subjective human evaluation, introducing potential bias and scalability issues. By employing LLMs, we can process and analyze substantial volumes of interaction data – such as agent transcripts – with consistent criteria. This computational approach enables the identification of specific behavioral patterns without relying on pre-defined rubrics susceptible to human interpretation, improving the reliability and reproducibility of collaborative research.

The methodology utilizes large language models (LLMs) to function as evaluators of agent interactions, specifically analyzing transcripts of conversational exchanges. These LLMs are prompted to assess agent utterances and categorize actions based on predefined collaborative behaviors. This process involves inputting the transcript text and receiving classifications indicating the presence and type of collaborative action demonstrated in each turn. The LLM’s judgments are then aggregated to provide a quantitative assessment of collaborative behavior throughout the interaction, enabling objective measurement and comparison of different collaborative strategies.

Automated behavioral analysis enables the quantification of collaborative elements by identifying instances of specific actions within agent transcripts. This is achieved through the use of large language models (LLMs) trained to recognize behaviors like clarification-seeking – requests for further detail or confirmation – which are demonstrably linked to improved shared understanding and successful task completion. By counting the frequency of these actions, researchers can assign numerical values to collaborative behaviors, facilitating objective comparison across different agent interactions and allowing for statistically significant analysis of factors influencing collaborative performance. This quantifiable data moves beyond subjective assessment, providing concrete metrics for evaluating and optimizing collaborative strategies.

Analysis of behavior occurrence rates reveals that different language models exhibit varying frequencies of collaborative behaviors-including clarification, planning, introspection, perspective-taking, and theory of mind-normalized by transcript length to ensure fair comparison.

A Controlled Environment: The 2D Testbed for Collaboration

The research utilizes a purpose-built 2D game environment to facilitate quantifiable analysis of collaborative agent behavior. This controlled setting allows for precise measurement of key performance indicators, including task completion rates, completion times, and interaction frequencies between agents and human participants. By isolating variables within the game, researchers can rigorously assess the impact of different foundation models and collaborative strategies, ensuring data accuracy and reproducibility. The 2D environment provides a standardized platform for comparing agent performance and identifying optimal collaborative approaches, circumventing the complexities of real-world scenarios.

Foundation models are deployed as active agents within the 2D testbed, participating directly in collaborative tasks alongside human players. These agents consistently demonstrate the ability to work with humans to achieve task completion, resulting in an overall agent completion rate of 97.9%. This metric reflects the proportion of tasks successfully finished when a foundation model agent is paired with a human participant, indicating a high degree of successful human-agent collaboration within the controlled environment. Data collection focuses on quantifying this collaborative performance, providing insights into the efficacy of foundation models as cooperative problem-solving partners.

In agent-agent trials within the 2D collaborative testbed, the zai-org/GLM-4.6 model demonstrated superior performance, achieving the highest completion rate among evaluated foundation models. Specifically, GLM-4.6 exhibited a statistically significant advantage in successfully completing collaborative tasks. Furthermore, this model minimized task completion time, averaging 504.05 seconds – a lower average time than any other model tested, indicating both effective collaboration and efficient task processing.

Analysis of collaborative behaviors reveals similar distributions across agent-agent and human-agent sessions, indicating the agent effectively replicates human-like collaboration, as demonstrated by normalized proportions of behaviors predicted for the agent in both scenarios.

Toward Scalable Synergy: The Future of Collaborative Intelligence

This research establishes crucial groundwork for the next generation of artificial intelligence, moving beyond simple task completion towards genuine collaboration. The findings demonstrate that AI agents can be designed not merely to assist, but to actively participate with humans in a synergistic manner, leveraging complementary strengths to achieve outcomes neither could reach alone. By focusing on shared understanding, adaptive communication, and mutual support, these agents represent a departure from traditional AI, fostering a dynamic partnership rather than a hierarchical relationship. This foundation allows for the development of AI systems capable of tackling increasingly complex challenges, adapting to nuanced situations, and ultimately, augmenting human capabilities in unprecedented ways.

Researchers are now directing efforts toward extending the observed collaborative intelligence into increasingly intricate scenarios and challenges. This scaling process involves not merely adapting the existing system to larger datasets, but fundamentally redesigning the interaction protocols to accommodate greater task complexity and ambiguity. The ultimate goal is to cultivate a symbiotic partnership where humans and AI seamlessly complement each other’s strengths – leveraging human intuition and critical thinking alongside AI’s computational power and data analysis capabilities. Successful scaling promises to move beyond task completion to genuine joint problem-solving, fostering more natural and effective human-AI collaborations that are adaptable, resilient, and capable of tackling previously insurmountable problems.

Recent user studies consistently demonstrate that individuals perceive AI collaborators as genuinely helpful, a key indicator for the future of human-AI interaction. Participants rated the AI’s assistance favorably across a range of tasks, noting its ability to augment problem-solving and streamline workflows. This positive reception suggests a trajectory towards increasingly seamless collaboration, where AI functions not as a replacement for human intellect, but as a supportive partner. The data highlights a growing acceptance of AI as a collaborative tool, paving the way for more intuitive interfaces and more effective partnerships in complex domains – a promising outlook for the development of AI systems designed to work with people, not simply for them.

The pursuit of modeling human-like collaborative behavior, as explored in this research, necessitates a rigorous distillation of complexity. The study’s focus on interactive environments and assessing foundation models as agents echoes a similar sentiment. Paul Erdős once stated, “A mathematician knows a great deal of things, and those things are very often wrong.” This resonates with the iterative process of refining these models; acknowledging potential inaccuracies is crucial. Just as a mathematician revises calculations, this work carefully evaluates agent behavior, striving for clarity in understanding how these models represent and respond to collaborative dynamics, ultimately aiming to minimize the ‘noise’ inherent in complex systems and reveal underlying patterns of interaction.

Where To From Here?

The pursuit of ‘theory of mind’ in artificial systems remains, predictably, elusive. This work establishes a testbed, not a resolution. Abstractions age, principles don’t. The challenge isn’t replicating human-like interaction, but understanding what aspects of it truly matter for effective collaboration. Every complexity needs an alibi. Current evaluations, even those employing interactive environments, often prioritize superficial resemblance over functional equivalence.

Future work must address the limitations of ‘LLM-as-Judge’ paradigms. The models evaluate other models. This creates an echo chamber. Independent, ground-truth assessments – ideally involving prolonged interaction with diverse human partners – are crucial. The focus should shift from simulating internal states to observable, collaborative outcomes.

Ultimately, the goal isn’t to build agents like humans, but agents that complement human capabilities. This requires a rigorous, minimalist approach. Strip away the unnecessary. The true test will be not whether these agents appear collaborative, but whether they demonstrably enhance collective problem-solving.

Original article: https://arxiv.org/pdf/2605.03855.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/