Searcharxiv⌕ Search

arXiv subjects

Gonzalo Martínez

Publications and source records attributed to Gonzalo Martínez.

At least 19 recordsLinked to original sources
arXiv2026

Decoding-Level Taboo: A Diagnostic Stress Test for LLM Robustness

Large language model evaluations typically focus on performance under nominal conditions, creating an illusion of capability where models comfortably walk a narrow, highly optimized generation corridor. In real-world deployments, however, complex system prompts, safety guardrails, and structural constraints continuously force models off this nominal path, driving a divergence between benchmark scores and deployment performance. To address this issue, we introduce Decoding-Level Taboo, a zero-prompt diagnostic stress test that intervenes directly in logit space at runtime, forcing models out of their nominal paths. By dynamically masking primary candidate tokens at word boundaries, Taboo forces machine circumlocution. Evaluating Taboo across several open-weight model families reveals that off-path robustness is heavily influenced by both parameter scale and post-training instruction alignment, with robustness generally improving with model size and alignment. Beyond the results presented in this paper, Taboo provides a novel primitive for generating diverse synthetic datasets, stress-testing runtime safety guardrails, and auditing model reliability prior to real-world deployment.

cs.CL↗
arXiv2026

Beyond the Hivemind: Escaping LLM Homogeneity via Meta-Persona Anchoring and Sequential Temperature Scaling

Recent studies have identified an ``Artificial Hivemind'' effect in Large Language Models (LLMs) causing models to converge on a narrow, homogenized consensus even for open questions. This semantic collapse limits the diversity of AI, resulting in high inter-response similarity ($\approx 0.80-0.90$) even under high-temperature sampling. In this paper, we propose a novel mitigation framework to increase diversity: Meta-Persona Anchoring combined with Filtered Temperature Scaling (FTS). Our approach utilizes a two-stage generation process: first, the model is prompted to self-select a unique, idiosyncratic persona to anchor its starting point; second, we apply a dual-stage sampling sieve, utilizing Top-$p$ filtering to preserve grammatical validity followed by extreme temperature scaling ($T \ge 4.0$) on the surviving candidates to explore the broadened probability distribution. We evaluate our method using the INFINITY-CHAT dataset on state-of-the-art open weight models under $\sim$20B parameters. Our results demonstrate a significant reduction in semantic convergence, with average pairwise cosine similarity dropping from ($\approx 0.85$) to ($\approx 0.65$). Our scheme achieves a majority of questions below the 0.7 threshold, effectively reducing the gap between artificial mode collapse and human-level typological diversity. We provide our implementation as an open-source framework to enable more diverse and creative AI deployments.

cs.AI↗
arXiv2026

Multiple Choice Questions: Reasoning Makes Large Language Models (LLMs) More Self-Confident, Especially When They are Wrong

Multiple Choice Question (MCQ) tests are among the most used methods for evaluating large language models (LLMs). Besides checking the correctness of the selected answer, evaluations often consider the model's confidence through the probability assigned to its response. In this work, we investigate how LLM confidence is influenced by the answering approach when the model answers directly or reasons before responding. Experiments on a general knowledge benchmark, covering 57 subjects and seven LLMs, show that models are systematically more confident when providing reasoning before answering, and that this confidence increase is larger when the selected answer is incorrect than when it is correct. We hypothesize that the reasoning process alters token probabilities, as the final answer prediction depends jointly on the question and the model's self-generated reasoning, leading to inflated confidence estimates. Using standard calibration metrics such as Expected Calibration Error and Brier score, we further show that Chain-of-Thought (CoT) prompting degrades calibration by increasing the proportion of high-confidence wrong answers. These findings indicate that, in MCQ evaluation settings with CoT prompting, LLM-estimated probabilities should be used with caution as a basis for evaluation and metacognitive mechanisms.

cs.CL↗
arXiv2026

To Words and Beyond: Probing Large Language Models for Sentence-Level Psycholinguistic Norms of Memorability and Reading Times

Large Language Models (LLMs) have recently been shown to produce estimates of psycholinguistic norms, such as valence, arousal, or concreteness, for words and multiword expressions, that correlate with human judgments. These estimates are obtained by prompting an LLM, in zero-shot fashion, with a question similar to those used in human studies. Meanwhile, for other norms such as lexical decision time or age of acquisition, LLMs require supervised fine-tuning to obtain results that align with ground-truth values. In this paper, we extend this approach to the previously unstudied features of sentence memorability and reading times, which involve the relationship between multiple words in a sentence-level context. Our results show that via fine-tuning, models can provide estimates that correlate with human-derived norms and exceed the predictive power of interpretable baseline predictors, demonstrating that LLMs contain useful information about sentence-level features. At the same time, our results show very mixed zero-shot and few-shot performance, providing further evidence that care is needed when using LLM-prompting as a proxy for human cognitive measures.

cs.CL↗
arXiv2026

Beyond Reproducibility: Token Probabilities Expose Large Language Model Nondeterminism

The execution of Large Language Models (LLMs) has been shown to produce nondeterministic results when run on Graphics Processing Units (GPUs), even when they are configured to produce deterministic results. This is due to the finite precision effects of the arithmetic operations, which depend on the order in which they are executed. This order, in turn, depends on the processes that are running concurrently on the GPU. Previous studies have focused on the impact of nondeterminism on the text generated by the LLMs or on proposing mechanisms to achieve deterministic execution. This work takes a closer look at nondeterminism by analyzing the variations on the token probabilities, not on the generated text. Interestingly, all the models evaluated have similar results in both the trends and the actual values of the variations of the probabilities. In particular, the results show that the effects of nondeterminism are significant for token probabilities that are in the range of 0.1 to 0.9, while they are much smaller when the probabilities are close to 0 or 1. This has significant implications for our understanding of nondeterminism. The first is that nondeterminism will likely have a non-negligible impact on generated text when the temperature is not zero, as it introduces significant variations in the token probabilities except when they are close to 0 or 1. Secondly, it suggests that all models have similar non deterministic variations at the token probability level. Therefore, different variations in the performance of the generated text, for example, when measuring accuracy on a benchmark, seem to come from different token probabilities or response lengths. A third implication is that we may be able to estimate the impact of nondeterminism by running a single inference and analyzing the token level probabilities, instead of having to run the same inference many times.

cs.AI↗
arXiv2025

Adding LLMs to the psycholinguistic norming toolbox: A practical guide to getting the most out of human ratings

Word-level psycholinguistic norms lend empirical support to theories of language processing. However, obtaining such human-based measures is not always feasible or straightforward. One promising approach is to augment human norming datasets by using Large Language Models (LLMs) to predict these characteristics directly, a practice that is rapidly gaining popularity in psycholinguistics and cognitive science. However, the novelty of this approach (and the relative inscrutability of LLMs) necessitates the adoption of rigorous methodologies that guide researchers through this process, present the range of possible approaches, and clarify limitations that are not immediately apparent, but may, in some cases, render the use of LLMs impractical. In this work, we present a comprehensive methodology for estimating word characteristics with LLMs, enriched with practical advice and lessons learned from our own experience. Our approach covers both the direct use of base LLMs and the fine-tuning of models, an alternative that can yield substantial performance gains in certain scenarios. A major emphasis in the guide is the validation of LLM-generated data with human "gold standard" norms. We also present a software framework that implements our methodology and supports both commercial and open-weight models. We illustrate the proposed approach with a case study on estimating word familiarity in English. Using base models, we achieved a Spearman correlation of 0.8 with human ratings, which increased to 0.9 when employing fine-tuned models. This methodology, framework, and set of best practices aim to serve as a reference for future research on leveraging LLMs for psycholinguistic and lexical studies.

cs.CL↗
arXiv2025

Artificial Intelligence and Misinformation in Art: Can Vision Language Models Judge the Hand or the Machine Behind the Canvas?

The attribution of artworks in general and of paintings in particular has always been an issue in art. The advent of powerful artificial intelligence models that can generate and analyze images creates new challenges for painting attribution. On the one hand, AI models can create images that mimic the style of a painter, which can be incorrectly attributed, for example, by other AI models. On the other hand, AI models may not be able to correctly identify the artist for real paintings, inducing users to incorrectly attribute paintings. In this paper, both problems are experimentally studied using state-of-the-art AI models for image generation and analysis on a large dataset with close to 40,000 paintings from 128 artists. The results show that vision language models have limited capabilities to: 1) perform canvas attribution and 2) to identify AI generated images. As users increasingly rely on queries to AI models to get information, these results show the need to improve the capabilities of VLMs to reliably perform artist attribution and detection of AI generated images to prevent the spread of incorrect information.

cs.CY↗
arXiv2025

The Generative Energy Arena (GEA): Incorporating Energy Awareness in Large Language Model (LLM) Human Evaluations

The evaluation of large language models is a complex task, in which several approaches have been proposed. The most common is the use of automated benchmarks in which LLMs have to answer multiple-choice questions of different topics. However, this method has certain limitations, being the most concerning, the poor correlation with the humans. An alternative approach, is to have humans evaluate the LLMs. This poses scalability issues as there is a large and growing number of models to evaluate making it impractical (and costly) to run traditional studies based on recruiting a number of evaluators and having them rank the responses of the models. An alternative approach is the use of public arenas, such as the popular LM arena, on which any user can freely evaluate models on any question and rank the responses of two models. The results are then elaborated into a model ranking. An increasingly important aspect of LLMs is their energy consumption and, therefore, evaluating how energy awareness influences the decisions of humans in selecting a model is of interest. In this paper, we present GEA, the Generative Energy Arena, an arena that incorporates information on the energy consumption of the model in the evaluation process. Preliminary results obtained with GEA are also presented, showing that for most questions, when users are aware of the energy consumption, they favor smaller and more energy efficient models. This suggests that for most user interactions, the extra cost and energy incurred by the more complex and top-performing models do not provide an increase in the perceived quality of the responses that justifies their use.

cs.AI↗
arXiv2025

La Leaderboard: A Large Language Model Leaderboard for Spanish Varieties and Languages of Spain and Latin America

Leaderboards showcase the current capabilities and limitations of Large Language Models (LLMs). To motivate the development of LLMs that represent the linguistic and cultural diversity of the Spanish-speaking community, we present La Leaderboard, the first open-source leaderboard to evaluate generative LLMs in languages and language varieties of Spain and Latin America. La Leaderboard is a community-driven project that aims to establish an evaluation standard for everyone interested in developing LLMs for the Spanish-speaking community. This initial version combines 66 datasets in Basque, Catalan, Galician, and different Spanish varieties, showcasing the evaluation results of 50 models. To encourage community-driven development of leaderboards in other languages, we explain our methodology, including guidance on selecting the most suitable evaluation setup for each downstream task. In particular, we provide a rationale for using fewer few-shot examples than typically found in the literature, aiming to reduce environmental impact and facilitate access to reproducible results for a broader research community.

cs.CL↗
arXiv2025

Is There a Case for Conversation Optimized Tokenizers in Large Language Models?

The computational and energy costs of Large Language Models (LLMs) have increased exponentially driven by the growing model sizes and the massive adoption of LLMs by hundreds of millions of users. The unit cost of an LLM is the computation of a token. Therefore, the tokenizer plays an important role in the efficiency of a model, and they are carefully optimized to minimize the number of tokens for the text in their training corpus. One of the most popular applications of LLMs are chatbots that interact with users. A key observation is that, for those chatbots, what is important is the performance of the tokenizer in the user text input and the chatbot responses. Those are most likely different from the text in the training corpus. So, a question that immediately arises is whether there is a potential benefit in optimizing tokenizers for chatbot conversations. In this paper, this idea is explored for different tokenizers by using a publicly available corpus of chatbot conversations to redesign their vocabularies and evaluate their performance in this domain. The results show that conversation-optimized tokenizers consistently reduce the number of tokens in chatbot dialogues, which can lead to meaningful energy savings, in the range of 5% to 10% while having minimal or even slightly positive impact on tokenization efficiency for the original training corpus.

cs.CL↗
arXiv2025

Psycholinguistic Word Features: a New Approach for the Evaluation of LLMs Alignment with Humans

The evaluation of LLMs has so far focused primarily on how well they can perform different tasks such as reasoning, question-answering, paraphrasing, or translating. For most of these tasks, performance can be measured with objective metrics, such as the number of correct answers. However, other language features are not easily quantified. For example, arousal, concreteness, or gender associated with a given word, as well as the extent to which we experience words with senses and relate them to a specific sense. Those features have been studied for many years by psycholinguistics, conducting large-scale experiments with humans to produce ratings for thousands of words. This opens an opportunity to evaluate how well LLMs align with human ratings on these word features, taking advantage of existing studies that cover many different language features in a large number of words. In this paper, we evaluate the alignment of a representative group of LLMs with human ratings on two psycholinguistic datasets: the Glasgow and Lancaster norms. These datasets cover thirteen features over thousands of words. The results show that alignment is \textcolor{black}{generally} better in the Glasgow norms evaluated (arousal, valence, dominance, concreteness, imageability, familiarity, and gender) than on the Lancaster norms evaluated (introceptive, gustatory, olfactory, haptic, auditory, and visual). This suggests a potential limitation of current LLMs in aligning with human sensory associations for words, which may be due to their lack of embodied cognition present in humans and illustrates the usefulness of evaluating LLMs with psycholinguistic datasets.

cs.CL↗
arXiv2025

Recursive InPainting (RIP): how much information is lost under recursive inferences?

The rapid adoption of generative artificial intelligence (AI) is accelerating content creation and modification. For example, variations of a given content, be it text or images, can be created almost instantly and at a low cost. This will soon lead to the majority of text and images being created directly by AI models or by humans assisted by AI. This poses new risks; for example, AI-generated content may be used to train newer AI models and degrade their performance, or information may be lost in the transformations made by AI which could occur when the same content is processed over and over again by AI tools. An example of AI image modifications is inpainting in which an AI model completes missing fragments of an image. The incorporation of inpainting tools into photo editing programs promotes their adoption and encourages their recursive use to modify images. Inpainting can be applied recursively, starting from an image, removing some parts, applying inpainting to reconstruct the image, revising it, and then starting the inpainting process again on the reconstructed image, etc. This paper presents an empirical evaluation of recursive inpainting when using one of the most widely used image models: Stable Diffusion. The inpainting process is applied by randomly selecting a fragment of the image, reconstructing it, selecting another fragment, and repeating the process a predefined number of iterations. The images used in the experiments are taken from a publicly available art data set and correspond to different styles and historical periods. Additionally, photographs are also evaluated as a reference. The modified images are compared with the original ones by both using quantitative metrics and performing a qualitative analysis. The results show that recursive inpainting in some cases modifies the image so that it still resembles the original one while in others leads to degeneration.

cs.CV↗
arXiv2025

It's the same but not the same: Do LLMs distinguish Spanish varieties?

In recent years, large language models (LLMs) have demonstrated a high capacity for understanding and generating text in Spanish. However, with five hundred million native speakers, Spanish is not a homogeneous language but rather one rich in diatopic variations spanning both sides of the Atlantic. For this reason, in this study, we evaluate the ability of nine language models to identify and distinguish the morphosyntactic and lexical peculiarities of seven varieties of Spanish (Andean, Antillean, Continental Caribbean, Chilean, Peninsular, Mexican and Central American and Rioplatense) through a multiple-choice test. The results indicate that the Peninsular Spanish variety is the best identified by all models and that, among them, GPT-4o is the only model capable of recognizing the variability of the Spanish language. -- En los últimos años, los grandes modelos de lenguaje (LLMs, por sus siglas en inglés) han demostrado una alta capacidad para comprender y generar texto en español. Sin embargo, con quinientos millones de hablantes nativos, la española no es una lengua homogénea, sino rica en variedades diatópicas que se extienden a ambos lados del Atlántico. Por todo ello, evaluamos en este trabajo la capacidad de nueve modelos de lenguaje de identificar y discernir las peculiaridades morfosintácticas y léxicas de siete variedades de español (andino, antillano, caribeño continental, chileno, español peninsular, mexicano y centroamericano y rioplatense) mediante un test de respuesta múltiple. Los resultados obtenidos indican que la variedad de español peninsular es la mejor identificada por todos los modelos y que, de entre todos, GPT-4o es el único modelo capaz de identificar la variabilidad de la lengua española.

cs.CL↗
arXiv2025

Playing with words: Comparing the vocabulary and lexical diversity of ChatGPT and humans

The introduction of Artificial Intelligence (AI) generative language models such as GPT (Generative Pre-trained Transformer) and tools such as ChatGPT has triggered a revolution that can transform how text is generated. This has many implications, for example, as AI-generated text becomes a significant fraction of the text, would this have an effect on the language capabilities of readers and also on the training of newer AI tools? Would it affect the evolution of languages? Focusing on one specific aspect of the language: words; will the use of tools such as ChatGPT increase or reduce the vocabulary used or the lexical richness? This has implications for words, as those not included in AI-generated content will tend to be less and less popular and may eventually be lost. In this work, we perform an initial comparison of the vocabulary and lexical richness of ChatGPT and humans when performing the same tasks. In more detail, two datasets containing the answers to different types of questions answered by ChatGPT and humans, and a third dataset in which ChatGPT paraphrases sentences and questions are used. The analysis shows that ChatGPT tends to use fewer distinct words and lower lexical richness than humans. These results are very preliminary and additional datasets and ChatGPT configurations have to be evaluated to extract more general conclusions. Therefore, further research is needed to understand how the use of ChatGPT and more broadly generative AI tools will affect the vocabulary and lexical richness in different types of text and languages.

cs.CL↗
arXiv2025

Can ChatGPT Learn to Count Letters?

Large language models (LLMs) struggle on simple tasks such as counting the number of occurrences of a letter in a word. In this paper, we investigate if ChatGPT can learn to count letters and propose an efficient solution.

cs.CL↗
arXiv2024

Beware of Words: Evaluating the Lexical Diversity of Conversational LLMs using ChatGPT as Case Study

The performance of conversational Large Language Models (LLMs) in general, and of ChatGPT in particular, is currently being evaluated on many different tasks, from logical reasoning or maths to answering questions on a myriad of topics. Instead, much less attention is being devoted to the study of the linguistic features of the texts generated by these LLMs. This is surprising since LLMs are models for language, and understanding how they use the language is important. Indeed, conversational LLMs are poised to have a significant impact on the evolution of languages as they may eventually dominate the creation of new text. This means that for example, if conversational LLMs do not use a word it may become less and less frequent and eventually stop being used altogether. Therefore, evaluating the linguistic features of the text they produce and how those depend on the model parameters is the first step toward understanding the potential impact of conversational LLMs on the evolution of languages. In this paper, we consider the evaluation of the lexical richness of the text generated by LLMs and how it depends on the model parameters. A methodology is presented and used to conduct a comprehensive evaluation of lexical richness using ChatGPT as a case study. The results show how lexical richness depends on the version of ChatGPT and some of its parameters, such as the presence penalty, or on the role assigned to the model. The dataset and tools used in our analysis are released under open licenses with the goal of drawing the much-needed attention to the evaluation of the linguistic features of LLM-generated text.

cs.CL↗
arXiv2024

Open Conversational LLMs do not know most Spanish words

The growing interest in Large Language Models (LLMs) and in particular in conversational models with which users can interact has led to the development of a large number of open-source chat LLMs. These models are evaluated on a wide range of benchmarks to assess their capabilities in answering questions or solving problems on almost any possible topic or to test their ability to reason or interpret texts. Instead, the evaluation of the knowledge that these models have of the languages has received much less attention. For example, the words that they can recognize and use in different languages. In this paper, we evaluate the knowledge that open-source chat LLMs have of Spanish words by testing a sample of words in a reference dictionary. The results show that open-source chat LLMs produce incorrect meanings for an important fraction of the words and are not able to use most of the words correctly to write sentences with context. These results show how Spanish is left behind in the open-source LLM race and highlight the need to push for linguistic fairness in conversational LLMs ensuring that they provide similar performance across languages.

cs.CL↗