KBC-Coop: Knowledge-Grounded Role Cooperation for Single-Model LLMs in Collaborative Tasks
Abstract
Recent studies have shown that common multiagent collaboration schemes for large language models (LLMs)- such as multi-round group discussion, debate, and unconstrained multi-role prompting-do not consistently outperform carefully designed single-agent baselines and can even degrade performance due to redundant deliberation and error amplification. This paper addresses the lack of explicit knowledge grounding, clear role specialization, and task-aware control of collaboration strength by proposing KBC-Coop, a knowledge- and budgetcontrolled cooperation framework implemented entirely within a single backbone LLM. KBC-Coop defines functional roles for problem framing, method design, evidence scouting, risk analysis, and coordination; each non-coordinator role retrieves a role-specific slice of a lightweight textual knowledge bank, and a difficulty-aware budget controller adaptively selects which roles participate and whether a refinement stage is invoked, thereby avoiding unnecessary collaboration on simple tasks while enabling richer multi-perspective reasoning on complex ones. Extensive experiments on three representative tasks from MultiAgentBench-research co-authoring, database error analysis, and coding collaboration-against a strong single-agent baseline and a naive multi-agent configuration demonstrate consistent improvements across all tasks. Index Terms-Knowledge-Based Cooperation, Large Language Models, Collaborative AI Systems I. INTRODUCTION Research on computer-supported cooperative work (CSCW) and related system design has long examined how computational systems mediate coordination, articulation work, and shared artifacts in collaborative settings [1], [2], [3], [4]. Classic CSCW studies emphasize that effective cooperation hinges not only on communication, but also on how work is decomposed, how responsibilities are distributed, and how common information spaces are structured and made intelligible to participants [3]. Against this backdrop, large language models (LLMs) are increasingly deployed as autonomous teammates, planners, or mediators in collaborative workflows, blurring the line between groupware infrastructure and AI-based collaborators [5]. R