2026-11-13 · America/Los_Angeles · 社区动态 · #12
Do decision models actually decide better? I tested them against the LLMs they're built from
When Jev came out, a bunch of similar "decision models" followed pretty quickly. I'm using Jev in a side project, so I wanted to know if the decision training actually does anything, or if it's basically an LLM answering in one token. Elephant 1 and Microsoft-Decision-1 are both built on Qwen3.5-9B, so I ran them against plain Qwen3.5-9B answering in one token, and the same Qwen thinking first. 450 questions generated by code (so the answers are known for sure), 5 runs each, all through OpenRouter. On the reasoning questions: GPT-6.1 Sol: 1.00 ($1.49 per 1k) Qwen3.5-9B thinking: 0.99 (23.5s,…
热度 46.0 / 100;排名与评分保留该期记录。