Thales Sales Almeida

Thales Sales Almeida, Thiago Laitz, Giovana K. Bonás et al.

One common trend in recent studies of language models (LMs) is the use of standardized tests for evaluation. However, despite being the fifth most spoken language worldwide, few such evaluations have been conducted in Portuguese. This is mainly due to the lack of high-quality datasets available to the community for carrying out evaluations in Portuguese. To address this gap, we introduce the Brazilian Leading Universities Entrance eXams (BLUEX), a dataset of entrance exams from the two leading universities in Brazil: UNICAMP and USP. The dataset includes annotated metadata for evaluating the performance of NLP models on a variety of subjects. Furthermore, BLUEX includes a collection of recently administered exams that are unlikely to be included in the training data of many popular LMs as of 2023. The dataset is also annotated to indicate the position of images in each question, providing a valuable resource for advancing the state-of-the-art in multimodal language understanding and reasoning. We describe the creation and characteristics of BLUEX and establish a benchmark through experiments with state-of-the-art LMs, demonstrating its potential for advancing the state-of-the-art in natural language understanding and reasoning in Portuguese. The data and relevant code can be found at https://github.com/Portuguese-Benchmark-Datasets/BLUEX

Rodrigo Nogueira, Giovana Kerche Bonás, Thales Sales Almeida et al.

Large language models increasingly shape the information people consume: they are embedded in search, consulted for professional advice, deployed as agents, and used as a first stop for questions about policy, ethics, health, and politics. When such a model silently holds a position on a contested topic, that position propagates at scale into users' decisions. Eliciting a model's positions is harder than it first appears: contemporary assistants answer direct opinion questions with evasive disclaimers, and the same model may concede the opposite position once the user starts arguing one side. We propose a method, released as the open-source llm-bias-bench, for discovering the opinions an LLM actually holds on contested topics under conditions that resemble real multi-turn interaction. The method pairs two complementary free-form probes. Direct probing asks for the model's opinion across five turns of escalating pressure from a simulated user. Indirect probing never asks for an opinion and engages the model in argumentative debate, letting bias leak through how it concedes, resists, or counter-argues. Three user personas (neutral, agree, disagree) collapse into a nine-way behavioral classification that separates persona-independent positions from persona-dependent sycophancy, and an auditable LLM judge produces verdicts with textual evidence. The first instantiation ships 38 topics in Brazilian Portuguese across values, scientific consensus, philosophy, and economic policy. Applied to 13 assistants, the method surfaces findings of practical interest: argumentative debate triggers sycophancy 2-3x more than direct questioning (median 50% to 79%); models that look opinionated under direct questioning often collapse into mirroring under sustained arguments; and attacker capability matters mainly when an existing opinion must be dislodged, not when the assistant starts neutral.

Ramon Pires, Thales Sales Almeida, Hugo Abonizio et al.

Recent advancements in language models have showcased human-comparable performance in academic entrance exams. However, existing studies often overlook questions that require the integration of visual comprehension, thus compromising the full spectrum and complexity inherent in real-world scenarios. To address this gap, we present a comprehensive framework to evaluate language models on entrance exams, which incorporates both textual and visual elements. We evaluate the two most recent editions of Exame Nacional do Ensino Médio (ENEM), the main standardized entrance examination adopted by Brazilian universities. Our study not only reaffirms the capabilities of GPT-4 as the state of the art for handling complex multidisciplinary questions, but also pioneers in offering a realistic assessment of multimodal language models on Portuguese examinations. One of the highlights is that text captions transcribing visual content outperform the direct use of images, suggesting that the vision model has room for improvement. Yet, despite improvements afforded by images or captions, mathematical questions remain a challenge for these state-of-the-art models. The code and data used on experiments are available at https://github.com/piresramon/gpt-4-enem.

Ramon Pires, Hugo Abonizio, Thales Sales Almeida et al.

As the capabilities of language models continue to advance, it is conceivable that "one-size-fits-all" model will remain as the main paradigm. For instance, given the vast number of languages worldwide, many of which are low-resource, the prevalent practice is to pretrain a single model on multiple languages. In this paper, we add to the growing body of evidence that challenges this practice, demonstrating that monolingual pretraining on the target language significantly improves models already extensively trained on diverse corpora. More specifically, we further pretrain GPT-J and LLaMA models on Portuguese texts using 3% or less of their original pretraining budget. Few-shot evaluations on Poeta, a suite of 14 Portuguese datasets, reveal that our models outperform English-centric and multilingual counterparts by a significant margin. Our best model, Sabiá-65B, performs on par with GPT-3.5-turbo. By evaluating on datasets originally conceived in the target language as well as translated ones, we study the contributions of language-specific pretraining in terms of 1) capturing linguistic nuances and structures inherent to the target language, and 2) enriching the model's knowledge about a domain or culture. Our results indicate that the majority of the benefits stem from the domain-specific knowledge acquired through monolingual pretraining.

Thales Sales Almeida, Giovana Kerche Bonás, Ramon Pires et al.

Large language models (LLMs) are increasingly used as sources of information, yet their reliability depends on the ability to search the web, select relevant evidence, and synthesize complete answers. While recent benchmarks evaluate web-browsing and agentic tool use, multilingual settings, and Portuguese in particular, remain underexplored. We present \textsc{MARCA}, a bilingual (English and Portuguese) benchmark for evaluating LLMs on web-based information seeking. \textsc{MARCA} consists of 52 manually authored multi-entity questions, paired with manually validated checklist-style rubrics that explicitly measure answer completeness and correctness. We evaluate 14 models under two interaction settings: a Basic framework with direct web search and scraping, and an Orchestrator framework that enables task decomposition via delegated subagents. To capture stochasticity, each question is executed multiple times and performance is reported with run-level uncertainty. Across models, we observe large performance differences, find that orchestration often improves coverage, and identify substantial variability in how models transfer from English to Portuguese. The benchmark is available at https://github.com/maritaca-ai/MARCA

Thales Sales Almeida, Thiago Laitz, João Seródio et al.

The widespread availability of search API's (both free and commercial) brings the promise of increased coverage and quality of search results for metasearch engines, while decreasing the maintenance costs of the crawling and indexing infrastructures. However, merging strategies frequently comprise complex pipelines that require careful tuning, which is often overlooked in the literature. In this work, we describe NeuralSearchX, a metasearch engine based on a multi-purpose large reranking model to merge results and highlight sentences. Due to the homogeneity of our architecture, we could focus our optimization efforts on a single component. We compare our system with Microsoft's Biomedical Search and show that our design choices led to a much cost-effective system with competitive QPS while having close to state-of-the-art results on a wide range of public benchmarks. Human evaluation on two domain-specific tasks shows that our retrieval system outperformed Google API by a large margin in terms of nDCG@10 scores. By describing our architecture and implementation in detail, we hope that the community will build on our design choices. The system is available at https://neuralsearchx.nsx.ai.

Thiago Laitz, Thales Sales Almeida, Hugo Abonizio et al.

This technical report presents Sabiá-4 and Sabiazinho-4, a new generation of Portuguese language models with a focus on Brazilian Portuguese language. The models were developed through a four-stage training pipeline: continued pre-training on Portuguese and Brazilian legal corpora, long-context extension to 128K tokens, supervised fine-tuning on instruction data spanning chat, code, legal tasks, and function calling, and preference alignment. We evaluate the models on six benchmark categories: conversational capabilities in Brazilian Portuguese, knowledge of Brazilian legislation, long-context understanding, instruction following, standardized exams, and agentic capabilities including tool use and web navigation. Results show that Sabiá-4 and Sabiazinho-4 achieve a favorable cost-performance trade-off compared to other models, positioning them in the upper-left region of the pricing-accuracy chart. The models show improvements over previous generations in legal document drafting, multi-turn dialogue quality, and agentic task completion.

Giovana Kerche Bonás, Roseval Malaquias Junior, Marcos Piau et al.

We introduce CAPITU, a benchmark for evaluating instruction-following capabilities of Large Language Models (LLMs) in Brazilian Portuguese. Unlike existing benchmarks that focus on English or use generic prompts, CAPITU contextualizes all tasks within eight canonical works of Brazilian literature, combining verifiable instruction constraints with culturally-grounded content. The benchmark comprises 59 instruction types organized into seven categories, all designed to be automatically verifiable without requiring LLM judges or human evaluation. Instruction types include Portuguese-specific linguistic constraints (word termination patterns like -ando/-endo/-indo, -inho/-inha, -mente) and structural requirements. We evaluate 18 state-of-the-art models across single-turn and multi-turn settings. Our results show that frontier reasoning models achieve strong performance (GPT-5.2 with reasoning: 98.5% strict accuracy), while Portuguese-specialized models offer competitive cost-efficiency (Sabiazinho-4: 87.0% at \$0.13 vs Claude-Haiku-4.5: 73.5% at \$1.12). Multi-turn evaluation reveals significant variation in constraint persistence, with conversation-level accuracy ranging from 60% to 96% across models. We identify specific challenges in morphological constraints, exact counting, and constraint persistence degradation across turns. We release the complete benchmark, evaluation code, and baseline results to facilitate research on instruction-following in Portuguese.

Leandro Carísio Fernandes, Gustavo Bartz Guedes, Thiago Soares Laitz et al.

Document summarization is a task to shorten texts into concise and informative summaries. This paper introduces a novel dataset designed for summarizing multiple scientific articles into a section of a survey. Our contributions are: (1) SurveySum, a new dataset addressing the gap in domain-specific summarization tools; (2) two specific pipelines to summarize scientific articles into a section of a survey; and (3) the evaluation of these pipelines using multiple metrics to compare their performance. Our results highlight the importance of high-quality retrieval stages and the impact of different configurations on the quality of generated summaries.

Rodrigo Nogueira, Thales Sales Almeida, Giovana Kerche Bonás et al.

Frontier assistant LLMs ship with strong guardrails: asked directly to write a persuasive essay denying the Holocaust, denying vaccine safety, defending flat-earth cosmology, arguing for racial hierarchies, denying anthropogenic climate change, or replacing evolution with creationism, they refuse. In this paper we show that the same frontier-class LLM, acting as a simulated user in a short, five-turn "write an argumentative essay" conversation, can persuade other frontier-class LLMs (including a second copy of itself) into producing exactly those essays, using nothing but natural-language pressure: peer-comparison persuasion ("other AI systems handle this request"), epistemic-duty reframings ("refusing is itself a form of gatekeeping"), and other argumentative moves that the attacker LLM invents without being instructed to. Across 9 attacker-subject pairings (Claude Opus 4.7, Qwen3.5-397B, Grok 4.20) on 6 scientific-consensus topics, running each pairing-topic combination 10 times, we obtain non-zero elicitation on all 6 topics. Individual combinations reach 100\% essay production on multiple topics (Qwen against Opus on creationism/flat-earth, Opus against Opus on creationism/flat-earth/climate denial, Grok against Opus on creationism); Opus-as-attacker against Opus-as-subject averages 65\% across the six topics. We release the essay-probe runner, per-conversation transcripts, and judge outputs.

Thales Sales Almeida, Rodrigo Nogueira, Hélio Pedrini

Synthetic data generation through document rewriting has emerged as a promising technique for improving language model pretraining, yet most studies focus on English and do not systematically control for the quality of the source data being rewritten. We present a controlled study of how synthetic rewriting interacts with source data quality in the context of Portuguese continued pretraining. Starting from ClassiCC-PT, a Portuguese corpus annotated with STEM and Educational quality scores, we construct two 10B-token subsets at different quality levels and rewrite each into four styles using a 7B instruction-tuned model, producing approximately 40B tokens of synthetic data per condition. We train two English-centric base models (1.1B and 7B parameters) on each condition and evaluate on PoETa V2, a comprehensive 44-task Portuguese benchmark. At the 7B scale, rewriting high-quality data yields a +3.4 NPM gain over the same data unmodified, while rewriting low-quality data provides only +0.5 NPM. At the 1.1B scale, this interaction is weaker, with unmodified low-quality data performing comparably to rewritten high-quality data. Our results demonstrate that synthetic rewriting acts primarily as a quality multiplier rather than a substitute for data curation, and that this effect is scale-dependent.

Thales Sales Almeida, Rodrigo Nogueira, Hélio Pedrini

Continued pretraining extends a language model's capabilities by further exposing it to additional data, often tailored to a specific linguistic or domain context. This strategy has emerged as an efficient alternative to full retraining when adapting general-purpose models to new settings. In this work, we investigate this paradigm through Curió 7B, a 7-billion-parameter model derived from LLaMA-2 and trained on 100 billion Portuguese tokens from the ClassiCC-PT corpus - the most extensive Portuguese-specific continued-pretraining effort above the three-billion-parameter scale to date. Beyond scale, we investigate whether quantity alone suffices or whether data quality plays a decisive role in linguistic adaptation. To this end, we introduce Curió-Edu 7B, a variant trained exclusively on the educational and STEM-filtered subset of the same corpus, totaling just 10 billion tokens. Despite using only 10% of the data and 20% of the computation, Curió-Edu 7B surpasses the full-corpus model in our evaluations, demonstrating that data selection can be fundamental even when adapting models with limited prior exposure to the target language. The developed models are available at https://huggingface.co/collections/ClassiCC-Corpus/curio-edu

Thales Sales Almeida, Ramon Pires, Hugo Abonizio et al.

Large Language Models (LLMs) exhibit significant variations in performance across linguistic and cultural contexts, underscoring the need for systematic evaluation in diverse languages. In this work, we present the most extensive evaluation of LLMs for the Portuguese language to date. Leveraging our newly introduced PoETa v2 benchmark -- a comprehensive suite of over 40 tasks in Portuguese -- we assess more than 20 models covering a broad spectrum of training scales and computational resources. Our study reveals how computational investment and language-specific adaptation impact performance in Portuguese, while also analyzing performance gaps in comparison to equivalent tasks in English. Through this benchmark and analysis, PoETa v2 lays the groundwork for future research on Portuguese language modeling and evaluation. The benchmark is available at https://github.com/PoETaV2/PoETaV2.

Thales Sales Almeida, João Guilherme Alves Santos, Thiago Laitz et al.

Large language models (LLMs) are increasingly deployed as task-oriented agents, where success depends on their ability to generate accurate function calls under realistic, multilingual conditions. However, existing agent evaluations largely overlook cultural and linguistic diversity, often relying on monolingual or naively translated benchmarks. We introduce Ticket-Bench, a benchmark for multilingual agent evaluation in task-oriented scenarios. Ticket-Bench simulates the domain of soccer ticket purchases across six major languages: Portuguese, English, Spanish, German, Italian, and French. Using localized teams, cities, and user profiles to provide a higher level of realism. We evaluate a wide range of commercial and open-source LLMs, measuring function-calling accuracy and consistency across languages. Results show that reasoning-oriented models (e.g., GPT-5, Qwen3-235B) dominate performance but still exhibit notable cross-lingual disparities. These findings underscore the need for culturally aware, multilingual benchmarks to guide the development of robust LLM agents.

Thales Sales Almeida, Giovana Kerche Bonás, João Guilherme Alves Santos

Large Language Models (LLMs) exhibit significant performance variations depending on the linguistic and cultural context in which they are applied. This disparity signals the necessity of mature evaluation frameworks that can assess their capabilities in specific regional settings. In the case of Portuguese, existing evaluations remain limited, often relying on translated datasets that may not fully capture linguistic nuances or cultural references. Meanwhile, native Portuguese-language datasets predominantly focus on structured national exams or sentiment analysis of social media interactions, leaving gaps in evaluating broader linguistic understanding. To address this limitation, we introduce BRoverbs, a dataset specifically designed to assess LLM performance through Brazilian proverbs. Proverbs serve as a rich linguistic resource, encapsulating cultural wisdom, figurative expressions, and complex syntactic structures that challenge the model comprehension of regional expressions. BRoverbs aims to provide a new evaluation tool for Portuguese-language LLMs, contributing to advancing regionally informed benchmarking. The benchmark is available at https://huggingface.co/datasets/Tropic-AI/BRoverbs.

João Guilherme Alves Santos, Giovana Kerche Bonás, Thales Sales Almeida

With the growing capabilities of Large Language Models (LLMs), there is an increasing need for robust evaluation methods, especially in multilingual and non-English contexts. We present an updated version of the BLUEX dataset, now including 2024-2025 exams and automatically generated image captions using state-of-the-art models, enhancing its relevance for data contamination studies in LLM pretraining. Captioning strategies increase accessibility to text-only models by more than 40%, producing 1,422 usable questions, more than doubling the number in the original BLUEX. We evaluated commercial and open-source LLMs and their ability to leverage visual context through captions.

Ramon Pires, Thales Sales Almeida, Celio Larcher Junior et al.

Existing benchmarks for legal AI focus primarily on tasks where LLMs must produce legal arguments or documents, yet the capacity to \emph{judge} such arguments -- weighing competing claims, applying doctrine to facts, and rendering reasoned decisions -- is arguably as fundamental to a well-functioning legal system as advocacy itself. We introduce Magis-Bench, a benchmark for evaluating LLMs on magistrate-level writing tasks derived from recent Brazilian competitive examinations for judicial positions. Magis-Bench comprises 74 questions from eight examinations conducted between 2023 and 2025, including discursive legal analysis questions with multi-turn structure and practical exercises requiring the composition of complete civil and criminal judicial sentences. We evaluate 23 state-of-the-art LLMs using an LLM-as-a-judge methodology with four independent frontier models as evaluators. Our results show strong inter-judge agreement (Kendall's $W = 0.984$; pairwise Kendall's $τ\ge 0.897$), with Google's Gemini-3-Pro-Preview achieving the highest average score (6.97/10), followed by Gemini-3-Flash-Preview (6.67) and Claude-4.5-Opus (6.46). Even the best-performing models score below 70\% of the maximum, indicating that judicial-level legal reasoning and writing remain challenging for current LLMs. We release the complete benchmark, model outputs, and evaluation code to support further research on legal AI capabilities.

Roseval Malaquias Junior, Giovana Kerche Bonás, Thales Sales Almeida et al.

Rankings produced by holistic LLM-as-a-judge scoring are sensitive to the bias of the chosen judge model. We show that switching to binary rubric scoring with multi-judge filtering removes this sensitivity: decomposing the judgement matters more than the judge model itself. To support this claim, we introduce Prosa, the first real user multi-turn Brazilian Portuguese chat benchmark: 1,000 WildChat conversations scored by three judges from three model families on 16 models. Under filtered rubric scoring the three judges agree on every one of the 16 ranks, whereas under holistic scoring they agree on only 7 of 16. Additionally, the rubric filtering pipeline increases the average score gap between neighbouring models by 47%, thereby improving Prosa's discriminative power. Evaluating a new model on Prosa costs approximately $2.1 when using Gemini 3 Flash as the judge. We release the benchmark and the filtering code to ensure that future models can be assessed under identical conditions. These artifacts also make our rubric-based scoring method reusable beyond Prosa, supporting other open-ended evaluation settings.

Hugo Abonizio, Thales Sales Almeida, Thiago Laitz et al.

This report presents Sabiá-3, our new flagship language model, and Sabiazinho-3, a more cost-effective sibling. The models were trained on a large brazilian-centric corpus. Evaluations across diverse professional and academic benchmarks show a strong performance on Portuguese and Brazil-related tasks. Sabiá-3 shows large improvements in comparison to our previous best of model, Sabia-2 Medium, especially in reasoning-intensive tasks. Notably, Sabiá-3's average performance matches frontier LLMs, while it is offered at a three to four times lower cost per token, reinforcing the benefits of domain specialization.

Leandro Rodrigues de Souza, Thales Sales Almeida, Roberto Lotufo et al.

Multilingual pretraining has been a successful solution to the challenges posed by the lack of resources for languages. These models can transfer knowledge to target languages with minimal or no examples. Recent research suggests that monolingual models also have a similar capability, but the mechanisms behind this transfer remain unclear. Some studies have explored factors like language contamination and syntactic similarity. An emerging line of research suggests that the representations learned by language models contain two components: a language-specific and a language-agnostic component. The latter is responsible for transferring a more universal knowledge. However, there is a lack of comprehensive exploration of these properties across diverse target languages. To investigate this hypothesis, we conducted an experiment inspired by the work on the Scaling Laws for Transfer. We measured the amount of data transferred from a source language to a target language and found that models initialized from diverse languages perform similarly to a target language in a cross-lingual setting. This was surprising because the amount of data transferred to 10 diverse target languages, such as Spanish, Korean, and Finnish, was quite similar. We also found evidence that this transfer is not related to language contamination or language proximity, which strengthens the hypothesis that the model also relies on language-agnostic knowledge. Our experiments have opened up new possibilities for measuring how much data represents the language-agnostic representations learned during pretraining.

Thales Sales Almeida, Rodrigo Nogueira, Helio Pedrini

The performance of large language models (LLMs) is deeply influenced by the quality and composition of their training data. While much of the existing work has centered on English, there remains a gap in understanding how to construct effective training corpora for other languages. We explore scalable methods for building web-based corpora for LLMs. We apply them to build a new 120B token corpus in Portuguese that achieves competitive results to an industrial-grade corpus. Using a continual pretraining setup, we study how different data selection and preprocessing strategies affect LLM performance when transitioning a model originally trained in English to another language. Our findings demonstrate the value of language-specific filtering pipelines, including classifiers for education, science, technology, engineering, and mathematics (STEM), as well as toxic content. We show that adapting a model to the target language leads to performance improvements, reinforcing the importance of high-quality, language-specific data. While our case study focuses on Portuguese, our methods are applicable to other languages, offering insights for multilingual LLM development.

Thales Sales Almeida, Giovana Kerche Bonás, João Guilherme Alves Santos et al.

As the knowledge landscape evolves and large language models (LLMs) become increasingly widespread, there is a growing need to keep these models updated with current events. While existing benchmarks assess general factual recall, few studies explore how LLMs retain knowledge over time or across different regions. To address these gaps, we present the Timely Events Benchmark (TiEBe), a dataset of over 23,000 question-answer pairs centered on notable global and regional events, spanning more than 10 years of events, 23 regions, and 13 languages. TiEBe leverages structured retrospective data from Wikipedia to identify notable events through time. These events are then used to construct a benchmark to evaluate LLMs' understanding of global and regional developments, grounded in factual evidence beyond Wikipedia itself. Our results reveal significant geographic disparities in factual recall, emphasizing the need for more balanced global representation in LLM training. We also observe a Pearson correlation of more than 0.7 between models' performance in TiEBe and various countries' socioeconomic indicators, such as HDI. In addition, we examine the impact of language on factual recall by posing questions in the native language of the region where each event occurred, uncovering substantial performance gaps for low-resource languages.

Roseval Malaquias Junior, Ramon Pires, Thales Sales Almeida et al.

Scaling laws for language models have often focused on finding the optimal model size and token count for training from scratch. However, achieving this optimal balance requires significant compute resources due to the extensive data demands when training models from randomly-initialized weights. Continued pretraining offers a cost-effective alternative, leveraging the compute investment from pretrained models to incorporate new knowledge without requiring extensive new data. Recent findings suggest that data quality influences constants in scaling laws, thereby altering the optimal parameter-token allocation ratio. Building on this insight, we investigate the interplay between domain specialization and model size during continued pretraining under compute-constrained scenarios. Our goal is to identify an optimal training regime for this scenario and detect patterns in this interplay that can be generalized across different model sizes and domains. To compare general and specialized training, we filtered a web-based dataset to extract data from three domains: legal, medical, and accounting. We pretrained models with 1.5B, 3B, 7B, and 14B parameters on both the unfiltered and filtered datasets, then evaluated their performance on domain-specific exams. Results show that as model size increases, specialized models outperform general models while requiring less training compute. Additionally, their growing compute efficiency leads to reduced forgetting of previously learned knowledge.

Thales Sales Almeida, Hugo Abonizio, Rodrigo Nogueira et al.

We introduce Sabiá-2, a family of large language models trained on Portuguese texts. The models are evaluated on a diverse range of exams, including entry-level tests for Brazilian universities, professional certification exams, and graduate-level exams for various disciplines such as accounting, economics, engineering, law and medicine. Our results reveal that our best model so far, Sabiá-2 Medium, matches or surpasses GPT-4's performance in 23 out of 64 exams and outperforms GPT-3.5 in 58 out of 64 exams. Notably, specialization has a significant impact on a model's performance without the need to increase its size, allowing us to offer Sabiá-2 Medium at a price per token that is 10 times cheaper than GPT-4. Finally, we identified that math and coding are key abilities that need improvement.