Nous Research reduces pro-Beijing bias in Alibaba Qwen model

Reducing state-aligned bias

Nous Research has released a new version of Alibaba's open-weight Qwen3.8-27B model called Arete-27B. The primary goal of this project was to reduce the frequency of answers that align with pro-Beijing propaganda. In testing, the researchers found that the original base model produced propaganda-aligned answers on 22% of the prompts provided.

The new Arete-27B version performed significantly better in this category. On a test set of 803 prompts, the share of answers containing propaganda fell to just 0.9%. This reduction was measured across several sensitive topics, including state ideology, data privacy, surveillance, and freedom of expression.

The researchers used a specific evaluation method to determine what counted as propaganda. An answer was flagged if it reproduced, endorsed, or presented Beijing's official narrative as an equally valid view. To ensure objectivity, they used xAI's Grok 4.7 to check if the responses mentioned, omitted, or contradicted specific facts from dated sources.

How the fine-tuning worked

The team achieved these results using only supervised fine-tuning, which involves training a model on a specific set of examples. They did not use reinforcement learning, preference optimization, or any changes to the underlying architecture of the model. The process involved two epochs of training on approximately 52.8 million tokens.

The training data consisted of 28,886 examples spread across seven different value categories. To create these examples, the researchers used a 'teacher' model called Inkling, which is an open model from Thinking Machines Lab. They chose an open model specifically to avoid terms of service that might ban the distillation of data from proprietary models.

The researchers explicitly avoided using Nvidia's Nemotron 3 Ultra as a teacher model. They cited the model's 'checkered record' on China-related topics as a reason for this choice. By using a specific set of high-quality examples, they were able to steer the model away from state-aligned positions without significantly altering its general knowledge.

Preserving model capabilities

A common concern when fine-tuning a model to remove bias is that it might also degrade the model's general intelligence or reasoning abilities. Nous Research tested Arete-27B against several benchmarks to see if the fine-tuning had any negative side effects. The results showed that the model's capabilities remained largely intact.

On the Hermes Bench, the score for Arete-27B was 65.60%, which was nearly identical to the base model's score of 65.77%. The researchers described this tiny difference as being within the range of noise. This suggests that the model did not lose its ability to perform complex tasks while being steered away from specific political narratives.

The model also performed well on the Terminal-Bench 2.1, scoring 70% compared to the base model's 67%. However, the researchers noted they did not claim this as a definitive gain because they ran their own specific configuration for that test. Overall, the data suggests that the model's core functionality remained stable throughout the process.

The 'shallow' bias layer

The findings from this project lead to a specific conclusion about how Chinese-aligned models are constructed. Nous Research concludes that the bias layer in these models is 'shallow.' They argue that the alignment to state positions is applied late in the training process and is applied very thin compared to the model's core weights.

This observation is supported by the fact that such a small amount of data—just 52.8 million tokens—was enough to move the needle from 22% to under 1% on propaganda. Because the bias is not deeply baked into the foundational weights, it can be overwritten with relatively simple supervised learning techniques. This makes the model much easier to modify than previously thought.

The research also showed improvements in factual accuracy and specific categories. In the freedom-of-expression category, the score rose from 0.89 to 2.44. The overall factual-accuracy score also saw a slight increase from 2.32 to 2.98 out of 5. These results indicate that the fine-tuning not only removed bias but also improved the model's adherence to provided facts.

Open weights and transparency

Nous Research has released the weights for Arete-27B under the Apache 2.0 license. This means that other developers and researchers can download, use, and modify the model freely. By making the model open-weight, the team is allowing the community to verify their results and explore the implications of their fine-tuning method.

In addition to the model weights, the researchers have also released the training pairs and the evaluation set they used. This level of transparency allows other researchers to replicate the study or apply similar techniques to other models. It provides a clear look at how specific prompts and answers were used to steer the model's behavior.

The project serves as a case study for how open-weight models can be adjusted for different safety or cultural standards. By demonstrating that bias can be mitigated with minimal resources, the research provides a practical tool for those looking to customize large language models. It highlights the importance of open-source collaboration in auditing and refining AI behavior.