Searcharxiv⌕ Search

arXiv subjects

Javier Coronado-Blázquez

Publications and source records attributed to Javier Coronado-Blázquez.

12 recordsLinked to original sources
arXiv2026

Lost in Sampling: Assessing Lexical Reachability in LLMs via the Word Coverage Score (WCS)

Modern Large Language Models (LLMs) are often criticized for producing repetitive and homogeneous text, despite possessing vast latent vocabularies. While previous research has focused on model knowledge and training data, we investigate the role of decoding mechanics in suppressing linguistic diversity. We introduce the Word Coverage Score (WCS), a metric that quantifies the extent to which contextually appropriate human vocabulary is mathematically pruned by standard sampling filters (e.g., Top-$p$, Top-$k$, and Min-$p$). Rather than assessing static knowledge, the WCS measures the lexical survival rate of low-frequency, high-information human words as a function of sampling parameters. By auditing open-weight models on human-authored corpus fragments, we identify which logical lexical choices are rendered unreachable by the decoder, even when they reside within the probability space. Our results provide quantitative evidence that industry-standard sampling defaults act as unintended censorship mechanisms, smoothing the unique textures of human expression into a homogenized discourse. The WCS offers a rigorous framework for optimizing the trade-off between text coherence and lexical richness, providing a diagnostic tool for preserving the diversity of human language in generative models.

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

Large Language Models and Book Summarization: Reading or Remembering, Which Is Better?

Summarization is a core task in Natural Language Processing (NLP). Recent advances in Large Language Models (LLMs) and the introduction of large context windows reaching millions of tokens make it possible to process entire books in a single prompt. At the same time, for well-known books, LLMs can generate summaries based only on internal knowledge acquired during training. This raises several important questions: How do summaries generated from internal memory compare to those derived from the full text? Does prior knowledge influence summaries even when the model is given the book as input? In this work, we conduct an experimental evaluation of book summarization with state-of-the-art LLMs. We compare summaries of well-known books produced using (i) only the internal knowledge of the model and (ii) the full text of the book. The results show that having the full text provides more detailed summaries in general, but some books have better scores for the internal knowledge summaries. This puts into question the capabilities of models to perform summarization of long texts, as information learned during training can outperform summarization of the full text in some cases.

cs.CL↗
arXiv2025

Evaluating book summaries from internal knowledge in Large Language Models: a cross-model and semantic consistency approach

We study the ability of large language models (LLMs) to generate comprehensive and accurate book summaries solely from their internal knowledge, without recourse to the original text. Employing a diverse set of books and multiple LLM architectures, we examine whether these models can synthesize meaningful narratives that align with established human interpretations. Evaluation is performed with a LLM-as-a-judge paradigm: each AI-generated summary is compared against a high-quality, human-written summary via a cross-model assessment, where all participating LLMs evaluate not only their own outputs but also those produced by others. This methodology enables the identification of potential biases, such as the proclivity for models to favor their own summarization style over others. In addition, alignment between the human-crafted and LLM-generated summaries is quantified using ROUGE and BERTScore metrics, assessing the depth of grammatical and semantic correspondence. The results reveal nuanced variations in content representation and stylistic preferences among the models, highlighting both strengths and limitations inherent in relying on internal knowledge for summarization tasks. These findings contribute to a deeper understanding of LLM internal encodings of factual information and the dynamics of cross-model evaluation, with implications for the development of more robust natural language generative systems.

cs.CL↗
arXiv2025

Deterministic or probabilistic? The psychology of LLMs as random number generators

Large Language Models (LLMs) have transformed text generation through inherently probabilistic context-aware mechanisms, mimicking human natural language. In this paper, we systematically investigate the performance of various LLMs when generating random numbers, considering diverse configurations such as different model architectures, numerical ranges, temperature, and prompt languages. Our results reveal that, despite their stochastic transformers-based architecture, these models often exhibit deterministic responses when prompted for random numerical outputs. In particular, we find significant differences when changing the model, as well as the prompt language, attributing this phenomenon to biases deeply embedded within the training data. Models such as DeepSeek-R1 can shed some light on the internal reasoning process of LLMs, despite arriving to similar results. These biases induce predictable patterns that undermine genuine randomness, as LLMs are nothing but reproducing our own human cognitive biases.

cs.CL↗
arXiv2024

A NLP Approach to "Review Bombing" in Metacritic PC Videogames User Ratings

Many videogames suffer "review bombing" -a large volume of unusually low scores that in many cases do not reflect the real quality of the product- when rated by users. By taking Metacritic's 50,000+ user score aggregations for PC games in English language, we use a Natural Language Processing (NLP) approach to try to understand the main words and concepts appearing in such cases, reaching a 0.88 accuracy on a validation set when distinguishing between just bad ratings and review bombings. By uncovering and analyzing the patterns driving this phenomenon, these results could be used to further mitigate these situations.

cs.CL↗
arXiv2023

Redshift prediction of Fermi-LAT gamma-ray sources using CatBoost gradient boosting decision trees

The determination of distance is fundamental in astrophysics. Gamma-ray sources are poorly characterized in this sense, as the limited angular resolution and poor photon-count statistics in gamma-ray astronomy makes it difficult to associate them to a multiwavelength object with known redshift. Taking the 1794 active galactic nuclei (AGNs) with known redshift from the Fermi-LAT latest AGN catalog, 4LAC-DR3, we employ machine learning techniques to predict the distance of the rest of AGNs based on their spectral and spatial properties. The state-of-the-art CatBoost algorithm reaches an average 0.56 R2 score with 0.46 root-mean-squared error (RMSE), predicting an average redshift value of $z_{avg}=0.63$, with a maximum $z_{max}=1.97$. We use the SHAP explainer package to gain insights into the variables influence on the outcome, and also study the extragalactic bakground light (EBL) implications. In a second part, we use this regression model to predict the redshift of the unassociated sample of the latest LAT point-source catalog, 4FGL-DR3, using the results of a previous paper to determine the possible AGNs within them.

astro-ph.HE↗
arXiv2022

Classification of Fermi-LAT unidentified gamma-ray sources using CatBoost gradient boosting decision trees

The latest $\textit{Fermi}$-LAT gamma-ray catalog, 4FGL-DR3, presents a large fraction of sources without clear association to known counterparts, i.e., unidentified sources (unIDs). In this paper, we aim to classify them using machine learning algorithms, which are trained with the spectral characteristics of associated sources to predict the class of the unID population. With the state-of-the-art $\texttt{CatBoost}$ algorithm, based on gradient boosting decision trees, we are able to reach a 67% accuracy on a 23-class dataset. Removing a single of these classes -- blazars of uncertain type -- increases the accuracy to 81%. If interested only in a binary AGN/pulsar distinction, the model accuracy is boosted up to 99%. Additionally, we perform an unsupervised search among both known and unID population, and try to predict the number of clusters of similar sources, without prior knowledge of their classes. The full code used to perform all calculations is provided as an interactive Python notebook.

astro-ph.HE↗
arXiv2021

Sensitivity of the Cherenkov Telescope Array to dark subhalos

In this work, we study the potential of the Cherenkov Telescope Array (CTA) for the detection of Galactic dark matter (DM) subhalos. We focus on low-mass subhalos that do not host any baryonic content and therefore lack any multiwavelength counterpart. If the DM is made of weakly interacting massive particles (WIMPs), these dark subhalos may thus appear in the gamma-ray sky as unidentified sources. A detailed characterization of the instrumental response of CTA to dark subhalos is performed, for which we use the {\it ctools} analysis software and simulate CTA observations under different array configurations and pointing strategies, such as the scheduled extragalactic survey. This, together with information on the subhalo population as inferred from N-body cosmological simulations, allows us to predict the CTA detectability of dark subhalos, i.e., the expected number of subhalos in each of the considered observational scenarios. In the absence of detection, for each observation strategy we set competitive limits to the annihilation cross section as a function of the DM particle mass, that are at the level of $\langleσv\rangle\sim4\times10^{-24}$ ($7\times10^{-25}$) $\mathrm{cm^3s^{-1}}$ for the $b\bar{b}$ ($τ^+τ^-$) annihilation channel in the best case scenario. Interestingly, we find the latter to be reached with no dedicated observations, as we obtain the best limits by just accumulating exposure time from all scheduled CTA programs and pointings over the first 10 years of operation. This way CTA will offer the most constraining limits from subhalo searches in the intermediate range between $\sim 1-3$ TeV, complementing previous results with \textit{Fermi}-LAT and HAWC at lower and higher energies, respectively.

astro-ph.HE↗
arXiv2020

X-ray and gamma-ray limits on the primordial black hole abundance from Hawking radiation

The non-observation of Hawking radiation from primordial black holes of $10^{16}$g sets a conservative strong bound on their cosmological abundance. We revisit this bound and show how it can be improved (both in mass reach and strength) by an adequate modeling of the combined AGN and blazar emission in the MeV range. We also estimate the sensitivity to the primordial black hole abundance of a future X-ray experiment capable of identifying a significantly larger number of astrophysical sources contributing to the diffuse background in this energy range.

astro-ph.CO↗
arXiv2019

Spectral and spatial analysis of the dark matter subhalo candidates among Fermi Large Area Telescope unidentified sources

Fermi-LAT unidentified sources (unIDs) have proven to be compelling targets for performing indirect dark matter (DM) searches. In a previous work, we found that among the 1235 unIDs in Fermi-LAT catalogs (3FGL, 2FHL and 3FHL) only 44 of those are DM subhalos candidates. We now implement a spectral analysis to test whether these remaining sources are compatible or not with DM origin. This analysis is executed using almost 10 years of Pass 8 Fermi-LAT data. None of the unIDs are found to significantly prefer DM-induced emission compared to other, more conventional, astrophysical sources. In order to discriminate between pulsar and DM sources, we developed a new method which is based on the source's spectral curvature, peak energy, and its detection significance. We also look for spatial extension, which may be a hint for a DM origin according to our N-body simulation studies of the subhalo population. In addition, we used Gaia DR2 data to search for a potential stellar counterpart to our best DM subhalo candidates and, although no firm associations could be found, one of them coincides with the Sagittarius stream. Finally, previous constraints on the DM annihilation cross section are updated with the new number of remaining DM subhalo candidates among unIDs. Our limits now rule out canonical thermal WIMPs up to masses of 10 GeV for $b\bar{b}$ and 20 GeV for $τ^+τ^-$ annihilation channels, in this way being as sensitive and complementary to those obtained from other targets and probes.

astro-ph.HE↗