Researchers probe Qwen2.5-7B for latent Colombian identity biases

A team of Colombian researchers has published a pilot study probing how Qwen2.5-7B-Instruct internally represents Colombian identity, socioeconomic status, and stereotype-related information from linguistic cues. Using a technique called Natural Language Autoencoders (NLA), they extracted and verbalized residual-stream activations from the model’s 20th layer across four positional quartiles per prompt. The dataset comprised 30 prompts arranged as 15 matched Spanish-English pairs, spanning explicit Colombian cues, implicit Colombian cues, and neutral controls.

The goal was to determine whether the model forms latent nationality or stereotype representations before those attributes appear in the generated output. The study reports descriptive rates and qualitative evidence rather than statistically powered effects, but the findings indicate that the model can infer demographic attributes from subtle phrasing—even when those attributes are not explicitly stated. This work connects activation-level interpretability with bias evaluation for underrepresented Spanish varieties, offering a new method to detect early-stage stereotyping in LLMs. The findings have implications for fairness in multilingual AI systems.

While the paper is a preliminary pilot with a small dataset, it demonstrates a promising approach for uncovering hidden biases encoded in model layers. The researchers emphasize that such bias detection is critical as LLMs are deployed across diverse linguistic communities. The study also opens the door for larger-scale analyses using similar autoencoding techniques.