SearchJev: A Fast and Calibrated System-1 Model for Search Agents

SearchJev. A Fast and Calibrated System-1 Model for Search Agents. By Congfeng Cao and eight others, from Huawei Technologies and Leiden University. Published on the fourth of October. Its code repository is public and real: seventeen source files.

A search agent makes the same small decisions over and over. Is this result relevant. Is there enough evidence yet. What search action comes next. Each is short, and there are many of them in a single search. When a generative language model makes each of those calls, two costs follow. It introduces latency, because every decision is produced by generating output. And its confidence is unreliable, so the agent cannot tell a sure call from a guess.

SearchJev is a fast, calibrated System-1 model that separates those short search decisions from System-2 reasoning and generation. Given a search state and a decision schema, it scores the legal options directly, with no autoregressive output generation. A decision becomes a choice among options rather than text to be written.

Its training method is Soft-Label Learning for Calibrated Decisions, or SLCD, which learns decision probabilities from uncertain supervision and calibrates their confidence. That calibration is what the dual-system agent relies on. SearchJev handles the short decisions and delegates the uncertain judgments to System 2, which keeps planning, query generation, and composing the answer.

The authors also build SearchDecision-Bench, a benchmark that unifies six types of search decision for training and evaluation.

On SearchDecision-Bench, SearchJev beats autoregressive Qwen3.5 models of the same size on decision quality. Its decisions are five point two to five point three times faster, and average expected calibration error falls by between forty-one and seventy-four percent.

On BrowseComp-Plus, the dual-system agents speed up active search time by three point seven to four point seven times. Answer accuracy rises at the same time, from forty-five percent to as much as fifty-four percent.

Here is the division of labor in one line. Quick, confident choices go to a scorer that never generates text, and only the uncertain ones reach the slower reasoner. Speed comes from the first half, and the calibration is what makes the handoff trustworthy.