开源决策模型 Karpo-Decision-31B 发布,300 个模型中并列第一
搜索推荐团队会想看这个:每千次决策成本省五分之一,延迟 95ms 还比 Gemini Decision 1 准。权重开源,直接能拿来做相关性判定和智能体的 act-or-ask 校准。
Karpo-Decision-31B 发布,在 Decision Index Full 榜单拿到 73.9 分,与 GLiDE 27B 的 73.4 并列 300 个模型中的第一名。该模型以 Gemma 4 31B 为基座,采用开源权重发布。
技术上的关键改动是弃用传统解码流程,改为一个 readout head:每个候选项对应一个单 token 编码,每个编码输出一个分数,再在该问题的选项范围内做 softmax。这样不生成任何 token,每次都能给出有效答案,并通过 branch attention 让页面表示在多个问题之间共享。
基准成绩方面,它基于 Gemma 4 31B,却在 Amazon ESCI 上以 64.7% 对 58.7% 超过 Google 的 Gemini Decision 1,在 New Yorker Caption 上以 90.5% 对 79.7% 领先。与 Jev 1.14 对比,其中位延迟为 95 ms 对 196 ms,JevBench 估算每 1000 次决策成本低约五分之一。这类模型适合搜索与推荐系统中需要大量相关性判定的场景,以及需要校准 act-or-ask 阈值的个人智能体。
A new open source decision model just launched, and it ranks first out of 300.
Karpo-Decision-31B swaps decoding for a readout head: a single-token code per option, one score per code, a softmax over that question's options. Zero generated tokens, always a valid answer, and the page is shared across questions through branch attention.
Decision Index Full: 73.9, the top score, tied #1 of 300 with GLiDE 27B at 73.4. The base is Gemma 4 31B, yet it beats Google's Gemini Decision 1 on Amazon ESCI, 64.7% vs 58.7%, and New Yorker Caption, 90.5% vs 79.7%.
Against Jev 1.14 it runs 95 ms vs 196 ms median, and JevBench estimates about a fifth less per 1,000 decisions.
Open weights matter for search and recommendation teams, who need many relevance calls per query, and for personal agents that need a calibrated act-or-ask threshold.
This report is categorized under AITOP's "Models" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Karpo-Decision-31B, Gemma, open-source models, recommendation systems.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).