Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces the Roundtable Context Window Test (RCWT), a novel protocol designed to quantify the impact of coordination content on task-budget displacement in multi-agent and memory-augmented LLM systems. By systematically varying coordination tokens while controlling for total budget and task context, the study reveals that commercial models maintain performance until a critical threshold of reference evidence is reached, after which performance sharply declines. Notably, when the task evidence is preserved while increasing coordination content, models like GPT-4.1-mini and Claude Haiku 4.5 achieve high accuracy, suggesting that the observed performance drop is primarily due to task-budget displacement rather than inherent semantic interference from coordination.
Task-budget displacement can dramatically affect LLM performance, revealing a critical threshold beyond which models struggle to maintain accuracy.
Multi-agent and memory-augmented LLM systems often place coordination content, shared state, prior discussion, tool outputs, summaries, and role instructions, inside the same finite prompt used for the current task. This creates a practical allocation problem: every token spent on coordination is unavailable to task instructions or evidence when a call is assembled under a fixed context budget. We introduce the Roundtable Context Window Test (RCWT), a controlled protocol for measuring this task-budget displacement effect. RCWT varies coordination content while controlling total budget, position order, task family, and scoring. In the main context-dependent recall task at $W=4096$, three commercial models remain near baseline through moderate overhead and then degrade sharply once residual reference evidence falls to a few hundred tokens. Window-scaling summaries are consistent with a task-specific residual-budget interpretation rather than a fixed percentage threshold, but we treat this as descriptive evidence rather than a universal law. To test whether the fixed-budget cliff persists when task evidence remains intact, we add an intact-task ablation: the full task/reference block is kept present while coordination tokens increase by expanding total prompt length. In that setting, all tested calls return every scored field correctly across GPT-4.1-mini, Claude Haiku 4.5, and Gemini 2.5 Flash up to a 95\% coordination ratio. This ablation narrows the claim: the main RCWT cliff is best read as task-budget displacement, not as proof that coordination volume alone causes semantic interference in the original open-ended task. RCWT is therefore a measurement primitive for context-allocation budgeting, not a complete theory of multi-agent benefit or session-level coordination.