AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 09:18:25 UTC.

Briefings

SearchJev is a fast System-1 model for search agents that separates short decision-making from System-2 reasoning. It uses Soft-Label Learning to calibrate confidence scores from uncertain supervision, achieving 5.2-5.3 times faster decisions and reducing expected calibration error by 41-74% on SearchDecision-Bench. The dual-system architecture delegates uncertain judgments to System 2 for planning and composition. Performance on BrowseComp-Plus shows a 3.7-4.7 times speedup in active search time with improved answer accuracy from 45% to 54%.

Key points

• SearchJev uses Soft-Label Learning to calibrate confidence scores from uncertain supervision, achieving 5.2-5.3 times faster decisions than same-size Qwen3.5 autoregressive models.
• The dual-system architecture delegates uncertain judgments to System 2 for planning, query generation, and answer composition while SearchJev handles short decisions.

Context

Search agents make short decisions about relevance and search actions, which generative language models introduce latency and unreliable confidence. SearchJev addresses this by separating decision-making from reasoning and generation, enabling faster and more calibrated search actions.

Limitations

Evidence is limited to SearchDecision-Bench and BrowseComp-Plus benchmarks; no additional evaluation metrics or real-world deployment data are provided. Results are based on preprint status and may vary across different search states or decision schemas.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 19/10/2026, 15:02:42 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigzq66y4xq5hgioxyyqkyjjfigxk22bfn2iamajjnwfqkpfozycdi