We describe in this paper the participation of team CIMAT-GTO in Task 1 (hard label setting) of EXIST 2025, which focuses on identifying sexism in tweets, determining the source's intention, and categorizing the types of sexism expressed. We propose a hybrid methodology that combines generative large language models with fine-tuned transformer-based classifiers through knowledge expansion. Our approach utilizes a generative model to highlight contextually relevant elements for the task as well as provide classification answers, and subsequently extracts justification texts that support the given predictions. We then conduct a justification-guided knowledge expansion when fine-tuning a smaller transformer-based model for classification, aiming for the model to learn from the reasoning encoded in the generated texts. We evaluate both monotask and multitask fine-tuning strategies and implement ensemble methods to improve robustness. Our results demonstrate that knowledge expansion using justifications obtained from generative models enhances performance over baseline few-shot classification and ifne-tuned models. The proposed systems prove to be competitive and achieve second place in all three textual tasks.
Social media platforms have become a widely used medium for communication and information
consumption in recent years. These platforms, despite allowing for quick and easy information exchange
among users, lead to problems such as exposure to misinformation and biased content [
However, current detection systems struggle with contextual understanding and nuanced categorization. Understanding the presence and prevalence of this phenomenon, and its many-sided manifestations, sheds light on the necessity of developing systems that are capable of identifying and characterizing this type of content while also managing large volumes of information.
This paper describes CIMAT-GTO’s participation in Task 1 of EXIST (sEXism Identification in Social
neTworks) 2025 [
In previous editions of EXIST, transformer-based architectures, such as BERT or RoBERTa, were the most
widely used and efective models in textual tasks. These models were typically fine-tuned with or without
prior pre-training, combined with techniques such as data augmentation, hierarchical classification,
annotator information injection, model cascades, and ensembles [
sEXism Identification in Social neTworks (EXIST) is a shared task that aims to foster sexism identification
and characterization in social media [
Six annotators labeled each tweet, and hard labels were computed by organizers following probabilistic thresholds. As we focus on predicting hard labels, we only considered tweets where such categorization was retained. In the case of task 1.1, binary identification of sexism-related content, this reduced the set to approximately 87% of the original dataset.
The three textual tasks addressed, in the hard label setting, in this work are • Task 1.1: Sexism Identification: Binary classification to detect sexism-related content. • Task 1.2: Source intention. Classification of intention if sexism-related content was identified.
Three possible categories: direct, reported, and judgemental. • Task 1.3: Sexism Categorization. Multi-label classification of sexism-related content was identified.
Possible categories include ideological and inequality, stereotyping and dominance, objectification, sexual violence, and misogyny and non-sexual violence.
Tasks 1.2 and 1.3 derive from the classification of Task 1.1 and provide more insights into the characterization of sexism-related content.
We propose hybrid systems that utilize generative LLMs to obtain answers to classification tasks (FS-Ctx), and then extract justification texts that explain the given answers. We then use these texts as knowledge expansion to fine-tune transformer-based classifiers ( FT-KE-Mono and FT-KE-Multi). For each of these ifne-tuned variants, we train three models as an ensemble to improve robustness and performance (Ens-FT-KE-Mono and Ens-FT-KE-Multi). This two-stage process is detailed in the following subsections. An overall description of the individual systems is provided in Figure 1, while Table 1 summarizes the models studied and their corresponding run numbers.
The first step in our methodology involves prompting an LLM to provide contextual cues about the
tweet and to answer the three tasks within a single query. Items in the set of contextual elements are
obtained by identifying recurrent patterns in the automatically discovered relevant information for
the task obtained in the generation stage of the proposal of other studies [
FS-Ctx (Few-Shot-Context) This is the generation-based method following the previously described setting. Although the generation following this methodology outperforms other prompt-based methods and some simple classifiers, it falls short compared to the best-performing classifier models.
We explore in the next stage how to achieve stronger models leveraging knowledge encoded by generative models that is indirectly present in the answers. In an intermediate step, we further prompt our generative LLM to justify the answers produced in the previous step based on the classification. We expect these texts to be informative and to enhance model performance, as they are used to enrich the input for classifier models.
We fine-tune two diferent types of models, using our proposed knowledge expansion strategy that concatenates justifications and the original tweet, and feeds these as input to the model, leveraging not only the classification preferences of the particular generative LLM from the previous step, but also reasoning in justifications. We evaluate two diferent types of fine-tuning techniques. We choose an XLM-RoBERTa trained on multilingual tweets [19] as a base model, to leverage information from both languages in the corresponding domain.
RoBERta-based model for each subtask of task 1. Hence, we rely on three diferent models to provide a complete response to all text tasks. A post-processing correction process is applied to labels predicted in tasks 1.2 and 1.3 to align with the negative output of task 1.1 and avoid contradiction.
labels for the three tasks at the same time, using a single multilingual RoBERTa-based model. We train by optimizing a new loss function obtained from the losses for each task.
To further enhance the performance of our models, we generate three distinct sets of justifications and train models for each set. Ensembles are produced using the output scores for each task to develop more robust labels in the final submission of the run, yielding the other systems submitted for evaluation (Ens-FT-KE-Mono and Ens-FT-KE-Multi).
The following section presents the results obtained during the development phase and the final results achieved on the oficial leaderboard.
We conduct a preliminary evaluation on the development set to validate our proposal and estimate the impact of each component within the proposed methods. Table 2 shows the performance of the generation-based classification and individual fine-tuned models. We observe that the only generative approach does not match the overall performance of the fine-tuned baseline, but still reaches a similar performance in task 1.2. The baseline is outperformed by systems with knowledge expansion, indicating that these systems are indeed learning from the outputs of the generative models. Multitask learning yields similar results to training individual systems, but has the advantage of requiring fewer computational resources, and shared representation benefits task 1.3 slightly.
ICM-Norm Task 1.1 ICM-Norm Task 1.2 ICM-Norm Task 1.3
System results of the selected systems are presented in Tables 3 and 4. As expected, fine-tuned submitted models outperform their generative counterparts, with the most significant improvement observed in task 1.3. Ens-FT-KE-Multi outperforms Ens-FT-KE-Mono in tasks 1.1 and 1.3, suggesting that training all tasks might provide insights across tasks that benefit the individual scores, and has the advantage of requiring fewer computational resources. Task 1.2 achieves its best performance with Ens-FT-KE-Mono, which may indicate that shared representation is not adding new information to the model.
Results segmented by language are presented in Table 4. Ranking in our system’s performance is consistent across languages, but shows diferences in metrics, with stronger results in Spanish. This suggests that further experimentation should be conducted in monolingual settings, possibly using techniques such as translation to leverage more data and information.
Our best-performing method achieves second place in all three text-based tasks, showing that combining generative reasoning with fine-tuned classifiers is a promising direction. While the metric performance is encouraging, the results indicate room for improvement and further experimentation.
This work presents a hybrid approach for sexism identification and characterization in social media, which combines the reasoning capabilities of generative LLMs with fine-tuned transformer-based classification models in a knowledge expansion process.
Key findings include: (1) generative models alone, while competitive, do not surpass fine-tuned classifiers; (2) fine-tuning with knowledge expansion using justifications from generative models of transformer-based models improves performance across all tasks; (3) multitask learning ofers computational eficiency while maintaining competitive results and benefiting task 1.3. Our systems achieve competitive results and attain high rankings on the EXIST 2025 leaderboard. The system’s performance and internal biases remain to be explored, and future research includes examining other LLMs as components used in the method.
It is essential to acknowledge that the models developed focus on predicting hard labels, overlooking the granularity of sociodemographic groups’ perspectives. In particular, when dealing with violent expressions, prioritizing the voices of victims and vulnerable groups rather than giving equal weight to the views of those who cause harm, helps to surface silenced experiences and attain responsible representation [20]. We also note that LLMs can reproduce various internal biases that might be misleading if not carefully monitored.
Task 1.1 1.2 1.3 1.1 1.2 1.3
EN ES
FS-Ctx Ens-FT-KE-Mono Ens-FT-KE-Multi FS-Ctx Ens-FT-KE-Mono Ens-FT-KE-Multi FS-Ctx Ens-FT-KE-Mono Ens-FT-KE-Multi FS-Ctx Ens-FT-KE-Mono Ens-FT-KE-Multi FS-Ctx Ens-FT-KE-Mono Ens-FT-KE-Multi FS-Ctx Ens-FT-KE-Mono Ens-FT-KE-Multi
Villarreal-Haro acknowledges Secretaría de Ciencia, Humanidades, Tecnología e Innovaciones (SECIHTI) for its support provided by the program Becas Nacionales Para Estudios de Posgrados (CVU 1309535). Sanchez-Vega acknowledges SECIHTI for its support through the program “Investigadoras e Investigadores por México” (Project ID.11989, No.1311). The authors thank SECIHTI for the computer resources provided through CIMAT Bajio Super-computing Laboratory (#300832) and Google Cloud Platform for the free cloud credits under its trial program used to access the Gemini API for experimentation in this project.
During the preparation of this work, the authors used Grammarly to perform a grammar and spelling check. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content. The prompt report: A systematic survey of prompting techniques, arXiv preprint arXiv:2406.06608 5 (2024). [19] F. Barbieri, L. Espinosa Anke, J. Camacho-Collados, XLM-T: Multilingual language models in Twitter for sentiment analysis and beyond, in: Proceedings of the 13th Language Resources and Evaluation Conference (LREC 2022), European Language Resources Association, Marseille, France, 2022, pp. 258–266. [20] J. Butler, Precarious life: The powers of mourning and violence, verso, 2004.