400 LoRA steps on Gemma 4 31B beat dedicated decision model Jev
vanstriendaniel · x · 2026-10-24
'System One' decision models like Jev return calibrated decisions at a fraction of LLM cost. The team shows you don't need specialized architectures: attach a decision head (softmax over LM-head rows) to an open LLM's last hidden state and train only LoRA adapters plus head correction, one forward pass per decision.
On OpenDecisions, closed-API Jev 1.13 scores 80.1 and Gemma 4 31B out-of-the-box 79.6; after just 400 LoRA steps (5 GPU-hours) it reaches 84.4, beating the dedicated model. Paper: arxiv.org/abs/2610.17205.
More from Models
- MiniMax M3.1 Flash tops 1K tok/s decode on SGLang, ~100K tok/s per GPU — ying11231 · 2026-10-24
- 1,626-question test: free e5-large embeddings beat OpenAI's paid tier in German and Spanish — julsimon · 2026-10-24
- Muse voice assistant insists the fake "iPhone Duo" doesn't exist — jesselyu · 2026-10-24
- SWE-sweep Leaderboard: Top Models Double Bug Finds at Half the Cost, Still Under 10% — stanfordnlp · 2026-10-24
- AI Finds Software Flaws Faster Than Fix: 45,207 CVEs Logged by Late July — rohanpaul_ai · 2026-10-24
- Kimi K3.1 Pro launches on mainland China site while overseas users still capped at K3 — teortaxesTex · 2026-10-24