SearchJev makes search-agent decisions 5x faster with 41-74% lower calibration error
Congfeng Cao · hf · 2026-10-06
SearchJev is a fast, calibrated System-1 model that splits search-agent decisions (relevance, evidence sufficiency, next actions) from System-2 reasoning and generation. Instead of autoregressive generation, it directly scores legal options given a search state and decision schema, and uses Soft-Label Learning for Calibrated Decisions (SLCD) to learn probabilities from uncertain supervision.
- Ships SearchDecision-Bench, unifying six types of search decisions for training and eval
- Beats same-size autoregressive Qwen3.5 models on decision quality, runs 5.2-5.3x faster, and cuts expected calibration error by 41-74%
- On BrowseComp-Plus, dual-system agents speed up active search 3.7-4.7x while lifting accuracy from 45% to up to 54%
Related event: SearchJev Speeds Up Search Agent Decisions While Improving Accuracy(2 posts)→
More from coding & agent
- Codex spins its wheels while Claude Opus shows far deeper understanding, user reports — iruletheworldmo · 2026-10-06
- Urlbox MCP Server Lets You Capture Screenshots, PDFs and Markdown via Natural Language — modelcontextprotocol · 2026-10-06
- OpenAI Dots Thread Steps 7-8: Approval Gates and Cost-Tiered Intelligence — FinanceYF5 · 2026-10-06
- OpenAI Dots Thread Steps 5-6: Wire in Business Tools and Schedules — FinanceYF5 · 2026-10-06
- 10-Step Blueprint Turns OpenAI Dots Into a 24/7 Company Operations Layer — FinanceYF5 · 2026-10-06
- ODS offers one-click local AI: auto hardware detection, model setup, agents and plugins — Teknium · 2026-10-06