SearchJev paper: System-1 search agent decisions run 5x faster with better calibration
_reachsumit · x · 2026-10-06
A new arXiv paper introduces SearchJev, a fast and calibrated System-1 model for search agents.
- Core idea: separate short search decisions (relevance, evidence sufficiency, search actions) from System-2 planning/query generation/answer composition. SearchJev scores options directly from token logits without autoregressive generation, delegating only uncertain cases to a larger model.
- Method: Soft-Label Learning for Calibrated Decisions (SLCD) learns decision probabilities from uncertain supervision and calibrates confidence; the authors also release SearchDecision-Bench, unifying six types of search decisions.
- Results: outperforms same-size Qwen3.5 autoregressive models, 5.2-5.3x faster decisions, 41-74% lower expected calibration error; dual-system agents on BrowseComp-Plus get 3.7-4.7x speedup in active search time and improve accuracy from 45% to up to 54%.
More from coding & agent
- Replit CEO: someone left an AI agent running overnight and woke up to $10,000 of wasted tokens — amasad · 2026-10-06
- Replit's Shlomi Fruchter: MCP, Harness and Skills are absurd concepts doomed like prompt engineering — shlomifruchter · 2026-10-06
- One Prompt Dump, Full App: Developer Wowed by Spawn Agent's Single-Shot Build — TAbrodi · 2026-10-06
- After Five Years, MATHPETS Launches as a Language and IDE for Agent-Based Models — jessi_cata · 2026-10-06
- Reading the source of 7 LLM eval tools uncovered 13 scoring bugs, 6 fixes merged — maverick_man1111 · 2026-10-06
- 6 open-source agent harnesses built for teams, from Deep Agents to Omnigent — femke_plantinga · 2026-10-06