中
◀ Back to feed
Nous Research

Arete-27B: Qwen3.8-27B post-trained to reinforce Western values

Published 2026-11-06 Collected 2026-11-07 16:00 AI news #ai#industry
Full text

Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values

Nous Research · November 2026

Nous post-trained Alibaba’s Qwen3.8-27B to reinforce Western values and released Arete-27B (weights + eval + data).

Measurement

They built an 803-prompt evaluation across seven value categories (speech, ideology, markets, education, privacy/surveillance, democracy, environment), scoring fact recall and propaganda framing with an LLM judge (Grok 4.7) that must quote phrases for propaganda labels.

Reported (thinking off, temp 0): base Qwen3.8-27B fact mean ~2.32/5, any-propaganda ~22% → Arete-27B fact mean ~2.98/5, any-propaganda ~0.9%. Largest lift cited in freedom of expression.

Training

About 52.8M tokens of supervised fine-tuning (no RL / no preference optimization / no architecture change). Teacher: Inkling. Hermes Bench ~65.60% vs base ~65.77% (noise-level). Terminal-Bench-2.1 reported 70% vs base 67% under their harness.

Takeaway (authors)

Bias layer appears shallow: tens of millions of tokens of supervision can shift framing without large capability regression on their checks. Open weights enable value optionality and auditable evals.

Weights: https://huggingface.co/NousResearch/Arete-27B
Post: https://nousresearch.com/arete

Source
Title
Arete-27B: Qwen3.8-27B post-trained to reinforce Western values
Author
—
Site
Nous Research
Published
2026-11-06
Collected
2026-11-07 16:00 · 2026-11-07-16-ai.json
Link
https://nousresearch.com/arete
Open original ↗

Reposted. Will be removed on request. All rights belong to the original author.

More from this batch

16:00batch · 2026-11-0712 items
>

南华早报转述 Menlo:中国模型主导开源企业支出(Kimi 约 61%);开源品类仍远小于闭源

SCMP:中国系统约占开源企业支出近九成,Kimi 领先,但专有美模型仍拿走总收入大头。今日中文圈热议的「89%」请回 Menlo 原文核对口径(开源仅占总支出约 4%)。

SCMP / Menlo
>

Anthropic 拟下周启动 IPO 路演,目标赶在感恩节前挂牌(估值预期下调压力仍在)

媒体援引知情人士:计划下周正式推介,拟于 11 月 16 日当周挂牌。此前市场曾传约 1.5 万亿美元估值募资超 800 亿;中期选举后科技股波动与 AI 支出担忧可能压低估值预期。非官方确认,以招股文件与交易所披露为准。

36氪 / 媒体
>

银行投行用 Claude 提速后:工作量未必下降——效率被更多交易/更细报告吃掉(Jevons)

趋势摘要:券商 VP 称 AI 可改 logo/更新联系人,但分析师反而承接更多 mandate。评论援引商品化与 Jevons 悖论。属舆情观察,非单一研究报告。

X 趋势
>

swyx:承诺持续周更 Anthropic IPO,直到成行或取消

与 Altimeter 的 Jamin Ball 合作跟踪。可与本期 YouTube「S-1 Myths」对照;属媒体覆盖承诺而非新监管披露。

X · swyx
>

TB-MathModel: first agentic, verifiable terminal bench for mathematical modeling

18 CUMCM/MCM tasks with JSON artifacts, one-sided scoring, feasibility gates and Lean certificates.

arXiv / HKUST
>

Decode Before You Hear:块交错全双工语音的「先解码再听」投机聆听

改进什么:全双工语音 Agent 边说边听时,等待完整块再理解会拖延迟。怎么做:对即将到来的听感知做投机解码,再与真实输入对齐。面向实时对话/语音 Agent 延迟。

HF Papers
>

SwarmEvolve:潜空间 Agent 群做演化式发现与优化

改进什么:单 Agent 搜索易陷入局部。怎么做:在潜空间维护可演化的 Agent 群,协同探索。HF 新帖,适合对照多 Agent 演化/群体智能路线。

HF Papers
>

Future Representation Distillation:无想象世界动作模型的表征蒸馏

改进什么:世界模型常靠昂贵 rollout「想象」。怎么做:用未来表征蒸馏,减少显式想象仍保留规划信号。面向具身/控制侧表示学习。

HF Papers
>

When Relevant Code Is Wrong:给编码 Agent 运行时感知上下文——相关代码也可能已失效

改进什么:检索到「相关」代码不等于仍在生效路径(特性开关/死代码)。怎么做:结合运行时权威信号筛选上下文。面向生产编码 Agent 降幻觉改坏。

arXiv
>

Claude Code 云会话被创作者做成「随手扔素材就能剪」的视频工作室

趋势摘要:Whisper+HyperFrames 等装进项目目录,参考片训练 Skill;旅行时用手机遥控家中 Mac Mini/VPS 渲染。属工作流拼装热潮,非 Anthropic 新品发布。

X 趋势
>

Governing AI agents: the next privileged user may not be human

Citrix previews SecurAccess agent governance: detect autonomous browsing, task-scoped MCP access, and session insights.

Citrix
>

为什么「陪伴」注定是 Personal AI 的核心

极客公园评论:打字像用工具,说话像找人;Personal AI 产品叙事从效率助手转向关系与在场。作国内消费端 Agent 产品讨论入口;无硬产品发布数字。

极客公园