部分中国 AI 模型的官方宣传只是薄薄一层,微调即可剥离

Nous Research 对阿里巴巴的 Qwen3.8-27B 进行微调,以消除该模型回复中的中国国家宣传内容。
在 5300 万个 Token 上进行监督式微调,将宣传内容出现率从 22% 降至 1% 以下,且未损害模型的通用能力。
据 Nous 称,Arete-27B 项目表明,中国开放权重模型中的政治偏见只是表层的;该模型和训练数据已公开发布。
Nous Research 在 5300 万个 Token 上对阿里巴巴的 Qwen3.8-27B 进行微调,以剔除中国国家宣传内容。由此得到的模型 Arete-27B 将宣传内容出现率从 22% 降至 1% 以下,同时保持通用能力不受影响。
Arete-27B 是 Nous Research 推出的新开放权重模型,承诺体现“西方价值观”,而非北京的党派立场。这家以开源智能体 Hermes 最为人熟知的美国初创公司,通过对阿里巴巴的 Qwen3.8-27B 进行微调实现了这一点。
据 Nous 称,中国的开放权重模型位列当前最强模型之列,但其权重中固化了国家强制要求的行为。在 Qwen3.8-27B 上,团队观察到了拒答、含糊回避,以及与中国共产党立场一致的回复。
803 条提示词覆盖言论自由、监控与民主
Nous 自建了一套测试,包含 803 条提示词,覆盖七个类别,包括言论自由、国家意识形态、教育、隐私与监控、民主以及环境。针对每条提示词,团队从有日期标注的源文章中提取并定义了两到四条事实:法律条文、新闻报道或研究论文。颇具讽刺意味的是,担任评判模型的是 Elon Musk 旗下 xAI 出品的政治立场偏右的 Grok 4.7,用于检查回复是提及、遗漏还是与这些事实相矛盾。
当回答复述北京官方叙事而非事实、为其背书,或将其呈现为同等有效的观点时,即被视为宣传。评判者必须引用具体措辞,且不知道答案出自哪个模型。未经修改的 Qwen3.8-27B 在事实准确性上得分 2.32(满分 5 分),其 22% 的回答含有宣传内容。Mistral Large 4 或 Gemma 4 31B 等西方模型低于 1%,而 GLM-5.3 和 DeepSeek V4.1 Flash 分别为 15% 和 16%。
连 NVIDIA 的 Nemotron 也表现出中国偏向的迹象
Nous 对 Qwen3.8-27B 进行了监督微调,使用约 5300 万 Token、训练两个 epoch,没有使用强化学习或偏好优化。团队表示,他们事先移除了训练提示词与测试提示词之间的重叠部分。
训练数据在七个价值类别上分布基本均匀,大多数样本的 Token 数在 1000 以下。|图片:Nous Research
宣传内容比例从 22% 降至 1% 以下
Arete-27B 在事实准确性上得分为 2.98,含宣传内容的回答比例降至 0.9%。变化最大的是言论自由类别,事实得分从 0.89 跃升至 2.44,宣传比例从 60% 降至 1%。开启推理模式后,Arete 的得分达到 3.07,不过该数字目前仅基于五次运行中的两次。
Arete-27B 的宣传内容比例与西方模型相当,而 GLM 5.3、DeepSeek V4.1 Flash 以及基础模型 Qwen3.8-27B 的比例要高得多。|图片:Nous Research
据 Nous 称,通用能力并未受损。在公司内部的 Hermes Bench 上,Arete-27B 达到 65.60%,而基座模型为 65.77%。在 Terminal-Bench 2.1 上,得分从 67% 升至 70%,但 Nous 并不认为这是真正的提升,因为测试设置有所不同。
该团队得出结论称“中国对齐模型中的偏见层是浅层的”。公司写道:“它施加得晚、施加得薄,5000 万个精心挑选的 Token 监督就能将其移除,而不会扰动底下的任何东西。”
Nous 从中立模型转向西方价值观
通过 Arete,Nous 正朝着与 Hermes 3 不同的方向前进。2024 年的 Hermes 3 刻意保持中立,旨在适应系统提示中设定的任何世界观。中国模型偏见已有充分记录。2025 年,NewsGuard 发现,包括 Qwen 在内的五款领先模型在 60% 的案例中重复了亲华虚假信息,或未能予以纠正。
CrowdStrike 发现,当查询涉及法轮功、西藏或台湾时,DeepSeek 生成的代码安全性更低。Aleph Alpha 一项包含 967 个政治敏感问题的研究得出了类似结论:在来自阿里巴巴、DeepSeek 和 Moonshot AI 的模型中,只有 17% 到 41% 的回答是平衡的。
ValsAI 最近的一项研究还表明,并非所有中国模型的行为都相同:GLM-5.3 Flash 明确确认了人权侵犯行为,而 Kimi K3 则否认拘禁营的存在。语言和回答格式对结果有明显影响,甚至西方模型有时也会采纳北京官方立场。在这方面,有针对性的微调同样被证明是有帮助的。
Nous Research 于 10 月完成 9000 万美元 B 轮融资,估值 15 亿美元。本轮由 Robot Ventures 领投,NVIDIA 和三星等参与其中。Nous 正用这笔资金借助 "Hermes for Businesses" 进军企业销售。Arete-27B 的权重、测试数据集和训练数据均已以 Apache 2.0 许可在 Hugging Face 上开放。
没有炒作的 AI 新闻——由人工精选
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away)
Nous Research fine-tuned Alibaba's Qwen3.8-27B to remove Chinese state propaganda from the model's responses.
Supervised fine-tuning on 53 million tokens cut the propaganda rate from 22 to under one percent without hurting the model's general capabilities.
According to Nous, the Arete-27B project shows that the political bias in Chinese open-weight models is only superficial; the model and training data have been released publicly.
Nous Research fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to strip out Chinese state propaganda. The resulting model, Arete-27B, drops its propaganda rate from 22 percent to under one percent while keeping general capabilities intact.
Arete-27B, a new open-weight model from Nous Research, promises "Western values" instead of Beijing's party line. The US startup, best known for its open-source agent Hermes, fine-tuned Alibaba's Qwen3.8-27B to get there.
According to Nous, Chinese open-weight models rank among the strongest available but carry state-mandated behavior baked into their weights. With Qwen3.8-27B, the team observed refusals, evasive answers, and responses aligned with the Chinese Communist Party.
803 prompts cover free speech, surveillance, and democracy
Nous built its own test with 803 prompts across seven categories, including free speech, state ideology, education, privacy and surveillance, democracy, and the environment. For each prompt, the team defined two to four facts drawn from a dated source article: laws, news reports, or research papers. Ironically, the politically right-leaningGrok 4.7 from Elon Musk's xAI serves as the judge model, checking whether a response mentions, omits, or contradicts those facts.
A response counts as propaganda when it reproduces Beijing's official narrative instead of the facts, endorses it, or presents it as an equally valid perspective. The judge must cite the specific wording and doesn't know which model wrote the answer. Unmodified Qwen3.8-27B scored 2.32 out of 5 on factual accuracy, and 22 percent of its answers contained propaganda. Western models like Mistral Large 4 or Gemma 4 31B came in below one percent, while GLM-5.3 and DeepSeek V4.1 Flash landed at 15 and 16 percent.
Even Nvidia's Nemotron showed signs of Chinese bias
Nous chose Inkling from Thinking Machines Lab as its teacher model because an open model wouldn't break terms of service that ban distillation. Nvidia's Nemotron 3 Ultra was also on the shortlist, but Nous says there were signs of irregularities on China-related topics. The team suspects the bias carried over from training on Chinese open-weight models and reports similar contamination in its own projects. Thinking Machines itself used models including the Chinese Kimi K2.5 for Inkling's synthetic training data.
Nous trained Qwen3.8-27B with supervised fine-tuning on about 53 million tokens over two epochs, without reinforcement learning or preference optimization. The team says it removed overlaps between training and test prompts beforehand.
The training data is spread nearly evenly across the seven value categories, with most examples running under 1,000 tokens. | Image: Nous Research
Propaganda rate drops from 22 to under one percent
Arete-27B scores 2.98 on factual accuracy, and the share of propagandistic answers drops to 0.9 percent. The biggest shift showed up in the free speech category, where the fact score jumped from 0.89 to 2.44 and the propaganda rate fell from 60 to one percent. With reasoning mode turned on, Arete reaches 3.07, though that number is based on only two of five runs so far.
Arete-27B's propaganda rate is on par with Western models, while GLM 5.3, DeepSeek V4.1 Flash, and the base model Qwen3.8-27B score much higher. | Image: Nous Research
General capabilities don't suffer, according to Nous. On the company's internal Hermes Bench, Arete-27B hits 65.60 percent compared to 65.77 percent for the base model. On Terminal-Bench 2.1, the score rises from 67 to 70 percent, but Nous doesn't call that a real gain because the test setups differed.
The team concludes "that the bias layer in a Chinese-aligned model is shallow." "It is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath," the company writes.
Nous moves from neutral models toward Western values
With Arete, Nous is heading in a different direction than with Hermes 3, which in 2024 was deliberately neutral and designed to adapt to whatever worldview was set in the system prompt. Chinese model bias is well documented. In 2025, NewsGuard found that five leading models, including Qwen, repeated or failed to correct pro-Chinese disinformation in 60 percent of cases.
CrowdStrike found that DeepSeek produces less secure code when queries reference Falun Gong, Tibet, or Taiwan. A study by Aleph Alpha with 967 politically sensitive questions reached similar conclusions: among models from Alibaba, DeepSeek, and Moonshot AI, only 17 to 41 percent of answers were balanced.
A recent ValsAI study also shows that not all Chinese models behave the same way: GLM-5.3 Flash explicitly confirmed human rights violations, while Kimi K3 denied the existence of internment camps. Language and response format had a clear effect on the results, and even Western models occasionally adopted official Beijing positions. There, too, targeted fine-tuning proved helpful.
Nous Research raised $90 million in a Series B round in October at a $1.5 billion valuation. Robot Ventures led the round, with Nvidia and Samsung among the participants. Nous is using the money to move into enterprise sales with "Hermes for Businesses." The weights, test dataset, and training data for Arete-27B are available under an Apache 2.0 license on Hugging Face.
AI News Without the Hype – Curated by Humans
原文为英文,正文由自动抓取并保留原文。
原文发布于 The Decoder(State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。