Large language models have shown strong progress on scientific reasoning, but they still face a fundamental bottleneck: when they need external information, explicit retrieval interrupts the flow of thought. The authors of a new arXiv paper describe this as a hidden "tool tax" — the extra tokens and steps required to fetch knowledge mid-reasoning.

To address this, they propose Eigen-Agent, a multi-agent system built around monitor-based retrieval. A monitor component decides when external knowledge is actually needed, allowing the model to reason continuously instead of stopping to search. This design is intended to make retrieval a background process rather than a disruptive interruption.

The paper, posted as arXiv:2509.21193, presents this as a way to reduce the overhead of retrieval while preserving the benefits of grounding in external sources. The abstract does not include experimental results, but the framework's emphasis on adaptive, monitor-driven retrieval points toward a more seamless integration of knowledge access and reasoning.