Selected work

Representative Papers

Probing Latent Colombian Identity Inferences in Qwen2.5-7B with Natural Language Autoencoders

Jul 23, 2026

This study investigates whether large language models implicitly infer Colombian nationality, socioeconomic status, or associated stereotypes through linguistic cues, even when such attributes are not explicitly mentioned. We introduce the first application of natural language autoencoders (NLAs) to bias detection in Latin American Spanish varieties, analyzing residual stream activations from layer 20 of the Qwen2.5-7B-Instruct model using multilingual prompt pairs and hierarchical positional probing. Our findings reveal that the model encodes nationality- and stereotype-related information about Colombia within its internal representations prior to output generation, particularly when inputs contain explicit or implicit Colombian cues. This work bridges activation-level interpretability with fairness evaluation for underrepresented linguistic groups, establishing a novel paradigm for detecting implicit bias in multilingual language models.

0 citationsRead paper

ProfileXAI: User-Adaptive Explainable AI

Oct 27, 2025

To address heterogeneous user requirements for model interpretability, this paper proposes a model- and domain-agnostic explainable AI (XAI) framework. Methodologically, it integrates post-hoc explanation techniques (SHAP, LIME, Anchor) with retrieval-augmented large language models (LLMs), implementing a user-profile-conditioned explanation mechanism that dynamically selects the optimal explainer. Explanations are generated via multimodal knowledge base indexing and conversational prompt engineering, yielding natural-language outputs with low redundancy and high fidelity. The key contributions are personalized explanation strategy adaptation and cross-user consistency preservation. Experimental evaluation on heart disease and thyroid cancer datasets demonstrates complementary strengths among explainers: average user satisfaction reaches 4.1/5, expert-assessed explanation quality scores 3.77/5, and token consumption remains stable (σ ≤ 13%).

0 citationsRead paper

Recent publications

Latest Papers

Probing Latent Colombian Identity Inferences in Qwen2.5-7B with Natural Language Autoencoders

Jul 23, 2026

This study investigates whether large language models implicitly infer Colombian nationality, socioeconomic status, or associated stereotypes through linguistic cues, even when such attributes are not explicitly mentioned. We introduce the first application of natural language autoencoders (NLAs) to bias detection in Latin American Spanish varieties, analyzing residual stream activations from layer 20 of the Qwen2.5-7B-Instruct model using multilingual prompt pairs and hierarchical positional probing. Our findings reveal that the model encodes nationality- and stereotype-related information about Colombia within its internal representations prior to output generation, particularly when inputs contain explicit or implicit Colombian cues. This work bridges activation-level interpretability with fairness evaluation for underrepresented linguistic groups, establishing a novel paradigm for detecting implicit bias in multilingual language models.

0 citationsRead paper

ProfileXAI: User-Adaptive Explainable AI

Oct 27, 2025

To address heterogeneous user requirements for model interpretability, this paper proposes a model- and domain-agnostic explainable AI (XAI) framework. Methodologically, it integrates post-hoc explanation techniques (SHAP, LIME, Anchor) with retrieval-augmented large language models (LLMs), implementing a user-profile-conditioned explanation mechanism that dynamically selects the optimal explainer. Explanations are generated via multimodal knowledge base indexing and conversational prompt engineering, yielding natural-language outputs with low redundancy and high fidelity. The key contributions are personalized explanation strategy adaptation and cross-user consistency preservation. Experimental evaluation on heart disease and thyroid cancer datasets demonstrates complementary strengths among explainers: average user satisfaction reaches 4.1/5, expert-assessed explanation quality scores 3.77/5, and token consumption remains stable (σ ≤ 13%).

0 citationsRead paper