Training and Evaluation of Large Language Models for Cognitive State Modeling

Adrián Maldonado Robles. (2026). Training and Evaluation of Large Language Models for Cognitive State Modeling. Final Career Project. Universidad Politécnica de Madrid.

Abstract:
This work investigates multilingual emotion analysis through two complementary tasks: multi-label emotion classification and emotion-conditioned text generation. The study focuses on English and Spanish, with Portuguese included in the transfer experiments, to examine how cross-lingual training strategies can improve performance in low-resource set- tings. We compare monolingual training, mixed-language training, and sequential transfer, and show that even limited target-language supervision can substantially improve classification results. In particular, multilingual models and parameter-efficient adaptation with Llama-3.2-3B achieve strong cross-lingual generalization, especially for Spanish. The second part of the study extends the analysis to emotion-conditioned generation. Using a prompt-based approach, we evaluate whether a language model can be steered to produce text aligned with a target emotion. To measure this alignment at scale, we reuse the trained classifier as an automatic evaluator. The results show that few-shot prompt- ing slightly improves emotional control, while zero-shot generation produces more diverse outputs. Overall, the findings suggest that cross-lingual transfer can reduce annotation requirements in multilingual emotion analysis and that classifier-based evaluation provides a practical way to assess emotion-conditioned generation.