Preprint 2026

Safe to Omit, Worth Omitting?

Keeping the Answer and Cutting Latency When an LLM Agent Skips Another’s Message

Xinyang Li1, Chih-Hsuan Yang2*, Xiaolong Ma2, Anjir Ahmed Chowdhury1, Vikram Vasudevan3, Rajeev Thakur2, Feng Yan1

1 University of Houston2 Argonne National Laboratory3 Oregon State University* Corresponding author

(a) The deployed system: every query waits for the helper agent's message, text or KV cache, before the receiver agent answers. (b) Our method: a probe reads the receiver agent's uncertainty u(x); if u(x) is at most the threshold tau, the receiver agent answers alone; otherwise the deployed system runs. Tau is certified on unlabeled calibration questions so that at most 5% of skipped answers change. View full-size figure ↗
Skip the helper’s message when the receiver agent is confident, with a certified bound on how many answers change.The threshold is certified on unlabeled calibration questions: with high probability, at most 5% of skipped answers differ from the deployed system’s. If no threshold passes, the system keeps communicating.

When can an agent skip another agent’s message?

A receiver agent often waits for a helper agent’s message, sent as text or as the helper’s KV cache. We read the receiver’s confidence without the message and certify, with a finite-sample test, a threshold at which at most 5% of skipped answers change. Across five benchmarks and three model pairs, oracle headroom proves weak evidence that messages help: control inputs without answer-specific content reproduce a median 76% of the oracle’s gain. Whether a message can be skipped depends on how often and how predictably it changes the receiver’s answer. And safety does not determine efficiency: skipping text saves hundreds of milliseconds per query at batch size one, while skipping the KV cache saves far less.

Xinyang Li, Chih-Hsuan Yang, Xiaolong Ma, Anjir Ahmed Chowdhury, Vikram Vasudevan, Rajeev Thakur, and Feng Yan. Preprint, 2026.

BibTeX
@misc{li2026safetoomit,
  title  = {Safe to Omit, Worth Omitting? Keeping the Answer and Cutting Latency When an {LLM} Agent Skips Another's Message},
  author = {Li, Xinyang and Yang, Chih-Hsuan and Ma, Xiaolong and Chowdhury, Anjir Ahmed and Vasudevan, Vikram and Thakur, Rajeev and Yan, Feng},
  year   = {2026},
  note   = {Preprint}
}
Figure 1Open in new tab ↗
Figure 1 of the paper at full size.

Click the figure to switch between fit and actual size. Press Esc to close.