We implement Large Language Model Alignment algorithms to formal logic reasoning tasks involving natural-language (NL) to first-order logic (FOL) translation, formal logic inference, and premise retranslation. These methodologies were implemented using task-specific preference datasets created based on the FOLIO datasets and LLM generations. Alignment was based on DPO, this algorithm was implemented and tested on off-the-shelf and pre-aligned models, showing promising results for higher quality NL-FOL parsing, as well as general alignment strategies. In addition, we introduce a new similarity metric (LogicSim) between LLM-generated responses and gold standard values, that measures logic-relevant information such as premise count and overlap between answers and expands evaluation of NL-FOL translation pipelines. Our results show that LLMs still struggle with logical inference, however alignment benefits semantic parsing and retranslation of results from formal logic to natural language.

We present MeSSI, a multi-module system applied to SemEval 2025’s task 3: Mu-SHROOM. Our system tags questions in order to obtain semantic relevant terms that are used as information retrieval characteristics. Said characteristics serve as extraction terms for Wikipedia pages that are in turn processed to generate gold standard texts used in a hallucination evaluation system. A PoST-based entity comparison was implemented to contrast the test dataset sentences with the corresponding generated gold standards, wich in turn was the main criteria to tag hallucinations, partitioned in soft labels and hard labels. This method was tested in Spanish and English, finishing 18th and 19th respectively on the IoU based ranking.

This paper describes Gradient Ascent and Task Vectors as LLM unlearning methodologies applied to SemEval 2025’s task 4. This task focuses on LLM unlearning on specific information under the constraints of preserving the model’s advanced text generation capabilities; meaning that our implementations of these algorithms were constrained both in the information datasets as well as the overall effect of each algorithm in the model’s general performance. Our implementation produced modified language models that ranked 7th out of 14 valid participants in the 7B parameter model, and 6th out of 24 in the 1B parameter model.

The STR shared task aims at detecting the degree of semantic relatedness between sentence pairs in multiple languages. Semantic relatedness relies on elements such as topic similarity, point of view agreement, entailment, and even human intuition, making it a broader field than sentence similarity. The GIL-IIMAS UNAM team proposes a model based in the SAND characteristics composition (Sentence Transformers, AnglE Embeddings, N-grams, Sentence Length Difference coefficient) and classical regression algorithms. This model achieves a 0.83 Spearman Correlation score in the English test, and a 0.73 in the Spanish counterpart, finishing just above the SemEval baseline in English, and second place in Spanish.

This paper explores whether it is possible to train a machine learning model using Wikipedia data to detect subjectivity in sentences and generalize effectively to other domains. To achieve this, we performed experiments with the WikiBias corpus, the BABE corpus, and the CheckThat! Dataset. Various classical models for ML were tested, including Logistic Regression, SVC, and SVR, including characteristics such as Sentence Transformers similarity, probabilistic sentiment measures, and biased lexicons. Pre-trained models like DistilRoBERTa, as well as large language models like Gemma and GPT-4, were also tested for the same classification task.