SearchJev separates search decisions from language generation
The paper reports faster calibrated decisions; downstream success and elapsed-time comparisons use distinct models and baselines.
SearchJev studies a lightweight System-1 component for frequent retrieval decisions alongside a larger System-2 language model. The decision-latency comparison reports 5.2–5.3-fold improvements and mean in-distribution expected calibration error reductions of 41–74% against the corresponding autoregressive JSON baselines. These metrics describe the tested decision tasks, not every search service.
The downstream BrowseComp-Plus comparison uses different system configurations. Table 4 reports System 2 Qwen3.5-27B at 45% success and 4080 seconds, Jev-dual at 46% and 1018 seconds, SearchJev-0.8B at 46% and 868 seconds, and SearchJev-4B at 54% and 1102 seconds. The time metric is median active time per question, including System-2 calls, decision-model inference and retrieval. On that metric, the 0.8B result is about 78.7% lower than the 4080-second System-2 baseline, or about 14.7% lower than Jev-dual. Those are separate comparisons; neither is a universal wall-clock or production latency guarantee.
The useful engineering question is whether a calibrated small decision component improves the accuracy/latency curve of a specific retrieval workload. Measure the same task, hardware and baseline before adopting it. The unrelated anonymous token-bill anecdote in the original article did not substantiate these results and has been removed.
Editorial correction (2026-10-13 UTC): Corrected model-specific baselines and elapsed-time denominators.