Cause of verification-wait underestimation

Determine the cause of the systematic underestimation of verification-wait inflation by trace replay, particularly for the Gemma target and Qwen3 TaskBench configurations.

Background

The latency framework predicts the verification-wait factor β by replaying sequential timing traces. For every Gemma configuration, and for Qwen3 TaskBench, the replay underestimates the measured verification wait by notable margins. The paper reports that this discrepancy affects speedup predictions and configuration ranking, but does not establish whether it arises from concurrency effects, request-dependent generation lengths, or another serving-system phenomenon.

References

The cause of this gap is not established.

— LEAP: Learning Efficient Action Proposals For LLM Agents  (2610.02670 - Xu et al., 2 Oct 2026) in Section 4.4, “Checking the latency framework”