Language models assigned to represent absent team members during live discussions face a fundamental coordination bottleneck. Rather than failing at text generation, proxy agents struggle to detect conversational openings in real time. Standard prompt-based implementations frequently remain passive during active discussions, leaving absent stakeholders unrepresented.
The Failure of Prompt-Only Proxies
When language model delegates rely solely on direct prompting over raw conversational transcripts, they miss conversational timing entirely. In benchmarks on the AMI corpus, prompt-only delegates stay silent on 51.4% of an absent participant's valid talking opportunities. This passivity occurs because naive LLMs struggle to track conversational floor control, participant stances, and whether a specific agenda item has already been adequately covered.
Prompt-only delegates stay silent on 51.4% of an absent participant's talking opportunities on the AMI corpus.
Simply expanding the context window does not solve the issue. As researchers observed, raw-context scaling fails to close the recognition gap, because longer transcripts do not inherently improve reasoning over multi-party conversational state transitions.
Modular Loop and State Decomposition
To address this limitation, researchers Muneeb Khan, Frederic Kirstein, Terry Ruas, and Bela Gipp from the University of Göttingen introduced CAPA, short for Collaborative Agent Predictive Architecture. The system operates on an online decision loop that separates state tracking from turn generation through specialized functional components.
At the core of the framework, a Perceiver updates an explicit meeting state from each observed turn. CAPA maintains this structured meeting state by tracking action-relevant fields: topic, decisions, questions, stances, coverage, and floor. A Predictor module then forecasts the immediate continuation of the conversation, while a Controller evaluates whether to intervene and selects the appropriate proposition. A Generator phrases the chosen proposition in the participant's style, and two judges score the forecast and action against the subsequent observed turn, allowing a Recalibrator to update the meeting state for upcoming turns.
Empirical Results and Baseline Reductions
Explicit state representation significantly sharpens intervention accuracy. Across 137 AMI meetings evaluated by schema-constrained LLM judges that align with human annotations at Cohen's kappa = 0.71, CAPA reduces the delegate silence rate from 51.4% down to 2.5%.
Furthermore, the architecture doubles credited recovery from 26.1 to 52.2 while maintaining a hallucination rate of 0.6%. Mechanism ablations confirm that maintaining the structured meeting state serves as the critical mechanism closing the recognition gap, shifting residual errors from complete omission to fine-grained proposition selection.
Replacing passive post-hoc transcript summarization with autonomous meeting representation requires explicit tracking of conversational state rather than naive prompt-only context scaling. While structured state updates and predictive recalibration resolve the silence gap on the AMI benchmark, deploying such architectures to unconstrained enterprise workflows will ultimately hinge on how reliably the modular loop handles rapid multi-speaker interruptions and domain-specific terminology under real-world latency constraints.