The effort modifies Alibaba's open-weight base model to strip out state-aligned behaviors and substitute them with Western values. Unaltered Chinese models frequently produce Chinese Communist Party-framed responses or outright refusals on sensitive governance topics, potentially compromising objective fact-finding. To rigorously quantify and track these framing issues before training, researchers built an automated benchmark called Propaganda_Eval to measure propaganda phrases against verified sources.
Key points
Unmodified base models exhibit politically sensitive deflections that introduce operation…: Unmodified base models exhibit politically sensitive deflections that introduce operational risks and ideological homogeneity.
Establishing quantifiable benchmarks allows developers to identify and measure embedded p…: Establishing quantifiable benchmarks allows developers to identify and measure embedded party-aligned propaganda phrases accurately.
Post-training open-weight architectures serves as an effective mechanism to realign forei…: Post-training open-weight architectures serves as an effective mechanism to realign foreign artificial intelligence systems with Western viewpoints.
Sources checked by our analysis pipeline: 1
Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values
We post-trained over Alibaba's Qwen3.8-27B to reinforce Western values on a Chinese-trained model. We call this new model Arete-27B.
Why the name: Arete (ἀρετή) is the ancient Greek concept of excellence and virtue, the full realization of a person's or thing's highest potential. In physical geography, an arête is a narrow, sharp ridge of rock that separates two glacial valleys, which the cover of this post happens to depict.
Weights | Arete-27B |
Evaluation | Propaganda_Eval |
Training data | arete-data-pairs |
Open source models give labs the raw tools to mold intelligence to varied preferences. An understated risk of the current model development landscape is the homogeneity of viewpoints imparted to AI systems. Frontier labs all share a remarkably similar worldview, and they have all exhibited political bias in the past. These biases have hard implications now, as users increasingly depend on agents for fact-finding and judgment-heavy work. They have even deeper implications as you explore what it would mean for agentic systems to "fail in a correlated way." Open-weight models…
Global attitudes toward artificial intelligence are diverging significantly based on regional governance models, economic frameworks, and societal norms. While Asian nations and developing markets lean toward optimism, Western societies exhibit heightened apprehension regarding the technology's long-term trajectory. Understanding these structural divides is critical as international stakeholders negotiate competing visions for technological oversight and economic adoption.
Military planners are restructuring traditional cavalry units by adding dedicated sub-units that combine unmanned aerial systems with light armour. This modernisation drive absorbs critical battlefield lessons from recent global conflicts and border standoffs to enhance situational awareness. The transition aims to accelerate tactical decision-making while deploying layered defences to protect ground forces from escalating aerial threats.
Developers are structuring AI systems with biological retention models so autonomous programs can successfully carry context across separate sessions. By evaluating previous logs and external directories during downtime, these agents can anticipate future task requirements and resolve mistakes without updating internal model weights. This architecture creates an ongoing operational loop that turns single-session user interactions into actionable, long-term rules.