Jev-like Decision Models Spark a Wave of Papers: Papers with Code Roundup

zainhas · x · 2026-10-13

Papers with Code now hosts a method page for "Decision Model" (System One model), introduced by TypeSafe AI with its model Jev: it maps unstructured state directly to a typed decision over predefined options instead of free-form text, with schema-valid outputs and calibrated probabilities—acting as a "smart if-statement" for routing, classification, scoring and guardrails. Jev decodes all outputs in a single parallel query and is trained via Reinforcement Learning for Calibrated Decisions (RLCD).

A wave of related papers has emerged, covering System-One-controlled agentic memory (Jev-Mem), paired and self-audited evaluations of System-1 decision models in LLM agent harnesses, when selection can replace extraction in agent memory, looped "System 1.5" thinking (SanSi), ordinal-scale bias analysis, Chinese-language adaptation, zero-shot detection of alignment failures, JEV-as-a-Judge with confidence-based escalation, and graph-task benchmarking (GraphDecisionBench).

Original post →

Jev-like Decision Models Spark a Wave of Papers: Papers with Code Roundup

More from coding & agent