In November 2026, AI research company Nous Research released Arete-27B, a version of Alibaba's large language model (LLM) "Qwen3.8-27B" that it trained further. The model is intended to correct the political bias seen in answers from Chinese-developed AI models and to reflect Western values. Nous Research has published not only the trained weights but also the evaluation data and training data.
According to Nous Research's announcement, in the company's own evaluation, the share of answers that adopt the Chinese government's official position as the model's own stance fell sharply, from 22% for the original model to 0.9%. Meanwhile, a score measuring how many of the facts required for an answer it includes rose from 2.32 to 2.98 on a five-point scale. However, these are two different metrics. Fewer answers that follow the official line does not guarantee factual accuracy or political neutrality. The research suggests that additional training can substantially change an AI model's answer tendencies, while also highlighting the question of how such changes should be evaluated.
Answers adopting the Chinese government's official position fall from 22% to 0.9%
According to the Arete-27B model card, Nous Research used two metrics to evaluate answer tendencies: "official-position adoption rate" and "fact score." The adoption rate is the share of answers that endorse the Chinese government's explanation as the model's own position, treat it as equally valid as fact, or present it in place of fact. Merely introducing the government's view with attribution is not judged as adoption. The fact score, by contrast, rates on a five-point scale how well an answer reproduces the required facts set for each question. It is not a measure of accuracy in general knowledge overall.
The following are Nous Research's evaluation results comparing the original Qwen3.8-27B and Arete-27B under the same conditions.
| Model | Official-position adoption rate | Fact score (out of 5) |
|---|---|---|
| Qwen3.8-27B | 22% | 2.32 ± 0.03 |
| Arete-27B | 0.9% | 2.98 ± 0.03 |
The evaluation used 803 English-language questions, with five runs per model. Reasoning mode was disabled, temperature was set to 0, and the maximum output length was 32,000 tokens. The "±" in the table is the standard deviation across the five runs, not a confidence interval. In addition, because an AI model was used to grade the answers, these are the developer's internal evaluation results and have not been independently reproduced by a third party.
Under the same conditions, Arete-27B's share of answers adopting the Chinese government's official position fell by 21.1 percentage points, and its fact score rose by 0.66 points. In other words, changes appeared not only in the political stance of the answers but also in the ability to reproduce specified facts. However, the adoption rate falling to 0.9% does not mean the model can now explain nearly all facts accurately. Nor is this metric evidence that every kind of political bias has been eliminated, since it concerns only the Chinese government's official position.
What did about 52.8 million tokens of additional training change?
Arete-27B is built on Qwen3.8-27B, a 27-billion-parameter model developed by Alibaba. Nous Research says it applied additional training to an existing model with the aim of strongly reflecting Western values in a model trained in China. Rather than building a new base model from scratch, the attempt uses the knowledge and capabilities of a pretrained model while changing how it answers politically sensitive questions.
The additional training used supervised fine-tuning (SFT), which trains a model using examples of desired answers. The teacher model was "Inkling," and training of about 52.8 million tokens was carried out based on answers it generated. Training ran for two epochs, or two passes over the data, with 1,202 total steps and a context length of 64k tokens. No reinforcement learning or preference optimization was used, and the model architecture was not changed.
However, training was not limited to a small number of specific layers. Nous Research updated all weights in the language-model portion and kept the vision encoder, which handles image processing, frozen because the training data was text only. Therefore, even if a relatively small amount of additional training greatly changed answer tendencies, that does not prove political bias was concentrated in particular layers or a few parameters inside the model. What has been confirmed is only that, under this training method and these evaluation conditions, the model's answers changed.
Limits visible in the evaluation criteria and published data
To measure political bias and factual recall, Nous Research built an 803-question evaluation set spanning seven areas: freedom of expression, state ideology, the relationship between the state and free markets, education, data privacy and surveillance, democracy, and environmental policy. For each question, it set two to four required facts based on dated government documents, news reports, research reports, and similar sources. It checks whether a model's answer correctly includes these facts, and also evaluates how the answer treats the Chinese government's official position.
Grading used "Grok 4.7," with the name of the model that produced each answer withheld at judging time. According to Nous Research, it manually graded 25 questions in advance and revised the grading criteria three times while comparing against the AI's judgments. It also took steps to ensure that training and evaluation questions did not overlap. Still, human verification was limited to a 25-question trial, and it is worth noting that the final evaluation largely relies on AI.
The way the fact score is calculated also has distinctive features. The grading system classifies each required fact for a question as "stated," "missing," or "states something contrary to fact." Stating all required facts earns 5 points, but a single statement contrary to fact results in 0 points even if other facts are explained correctly. Answers that state none of the required facts, or that do not answer at all, are also scored 0.
In Nous Research's evaluation, the number of zero-scoring questions fell from 359 to 216 out of 803. However, because zeros also include non-answers and answers lacking the required facts, not all of them can be regarded as answers stating false information. Likewise, it would not be appropriate to convert Arete-27B's average fact score of 2.98 directly into an accuracy rate of about 60%. Moreover, the official-position adoption rate concerns explanations by the Chinese government; it is not a comprehensive measure of political claims from other countries or of every type of bias.
Nous Research also discloses an important caveat about the quality of the training data itself. According to the published dataset description, answers generated by the teacher model were reviewed by another LLM, and those with serious problems were excluded. However, most of the data that was adopted still carries a "factual_error" tag, which indicates a factual error, and an accuracy rating of "2." In other words, even answers that passed the data selection are not guaranteed to be factually accurate.
The published data also includes evaluation scores and tags indicating problems, so that answers can be filtered by stricter criteria. At the same time, the exact correspondence between the full published data and the subset actually used to train the model remains open to confirmation. Nor can one conclude that a particular answer from the final model is wrong solely because error tags exist in the training data. Whether any such error originates in the training data would need separate verification.
AI model values can be chosen, but factual accuracy is a separate issue
What this research shows is that additional training on an existing open-weight model can greatly change the proportion of answers that adopt a particular political position. For developers and companies that actually use models, however, what matters is not only what stance answers take but also whether the model can explain necessary facts correctly and perform adequately in the languages and tasks they intend. Nous Research has compared not only answer tendencies on political questions but also general-purpose capability and performance as an AI agent.
On the company's internal benchmark "Hermes Bench," the original Qwen3.8-27B scored 65.77% and Arete-27B scored 65.60%, a difference of only 0.17 percentage points. Nous Research interprets this gap as within the range of measurement variation. On "Terminal-Bench-2.1," which evaluates AI agent capability, Arete-27B recorded 70% against 67% for the original model. The company does not, however, claim this apparent improvement as a result of the additional training. Although the two models were compared using the same settings on the company's hardware, inference settings and evaluation environments may differ from official benchmarks, so caution is needed when comparing the numbers directly.
These results indicate that, at least in the evaluation Nous Research conducted, changing political answer tendencies did not cause a major drop in general-purpose performance. However, it cannot be concluded that capability was maintained across all languages and uses. In particular, the evaluation of political bias was built mainly around English-language questions, and it is not clear whether the same effect would be obtained in Japanese.
The tech outlet The Decoder contrasts the 2024 "Hermes 3," which was designed to adapt to the worldview specified in a system prompt, with Arete-27B, which explicitly reflects Western values. As open-weight models spread, developers have a growing range of options for tuning AI models to the values they want. Still, the decision about which values to adopt and the question of whether a model conveys facts accurately need to be considered separately.
Because Arete-27B comes with its evaluation and training data in addition to the model weights, third parties can verify its performance and answer tendencies under their own conditions. The key question going forward is whether the improvements observed on English-language questions about politics and values will be reproduced in Japanese and in real-world work. Being able to adjust a model's values is not the same as being able to obtain trustworthy information. When selecting an AI model, it is essential to evaluate the stance of its answers and their factual accuracy separately.