SearchJev: A Fast and Calibrated System-1 Model for Search Agents
SearchJev replaces small search-agent judgments with direct option scoring, cutting decision latency while improving calibration.
The paper says SearchJev handles short choices such as relevance, evidence sufficiency, and next search actions without autoregressive generation. Its SLCD training method is designed to learn calibrated probabilities from uncertain supervision. On SearchDecision-Bench, it beats same-size Qwen3.5 autoregressive models, runs 5.2-5.3x faster, and lowers expected calibration error by 41-74%. In dual-system agents on BrowseComp-Plus, active search time is 3.7-4.7x faster while answer accuracy rises from 45% to as high as 54%. HF Daily Papers' note
The paper says SearchJev handles short choices such as relevance, evidence sufficiency, and next search actions without autoregressive generation. Its SLCD training method is designed to learn calibrated probabilities from uncertain supervision. On SearchDecision-Bench, it beats same-size Qwen3.5 autoregressive models, runs 5.2-5.3x faster, and lowers expected calibration error by 41-74%. In dual-system agents on BrowseComp-Plus, active search time is 3.7-4.7x faster while answer accuracy rises from 45% to as high as 54%. HF Daily Papers' note
score 5