Rodrigo Nogueira

Vitor Jeronymo, Luiz Bonifacio, Hugo Abonizio et al.

Recently, InPars introduced a method to efficiently use large language models (LLMs) in information retrieval tasks: via few-shot examples, an LLM is induced to generate relevant queries for documents. These synthetic query-document pairs can then be used to train a retriever. However, InPars and, more recently, Promptagator, rely on proprietary LLMs such as GPT-3 and FLAN to generate such datasets. In this work we introduce InPars-v2, a dataset generator that uses open-source LLMs and existing powerful rerankers to select synthetic query-document pairs for training. A simple BM25 retrieval pipeline followed by a monoT5 reranker finetuned on InPars-v2 data achieves new state-of-the-art results on the BEIR benchmark. To allow researchers to further improve our method, we open source the code, synthetic data, and finetuned models: https://github.com/zetaalphavector/inPars/tree/master/tpu

Jayr Pereira, Robson Fidalgo, Roberto Lotufo et al.

This paper proposes a question-answering system that can answer questions whose supporting evidence is spread over multiple (potentially long) documents. The system, called Visconde, uses a three-step pipeline to perform the task: decompose, retrieve, and aggregate. The first step decomposes the question into simpler questions using a few-shot large language model (LLM). Then, a state-of-the-art search engine is used to retrieve candidate passages from a large collection for each decomposed question. In the final step, we use the LLM in a few-shot setting to aggregate the contents of the passages into the final answer. The system is evaluated on three datasets: IIRC, Qasper, and StrategyQA. Results suggest that current retrievers are the main bottleneck and that readers are already performing at the human level as long as relevant passages are provided. The system is also shown to be more effective when the model is induced to give explanations before answering a question. Code is available at \url{https://github.com/neuralmind-ai/visconde}.

Desnes Nunes, Ricardo Primi, Ramon Pires et al.

The present study aims to explore the capabilities of Language Models (LMs) in tackling high-stakes multiple-choice tests, represented here by the Exame Nacional do Ensino Médio (ENEM), a multidisciplinary entrance examination widely adopted by Brazilian universities. This exam poses challenging tasks for LMs, since its questions may span into multiple fields of knowledge, requiring understanding of information from diverse domains. For instance, a question may require comprehension of both statistics and biology to be solved. This work analyzed responses generated by GPT-3.5 and GPT-4 models for questions presented in the 2009-2017 exams, as well as for questions of the 2022 exam, which were made public after the training of the models was completed. Furthermore, different prompt strategies were tested, including the use of Chain-of-Thought (CoT) prompts to generate explanations for answers. On the 2022 edition, the best-performing model, GPT-4 with CoT, achieved an accuracy of 87%, largely surpassing GPT-3.5 by 11 points. The code and data used on experiments are available at https://github.com/piresramon/gpt-4-enem.

Guilherme Moraes Rosa, Luiz Bonifacio, Vitor Jeronymo et al.

Recent work has shown that small distilled language models are strong competitors to models that are orders of magnitude larger and slower in a wide range of information retrieval tasks. This has made distilled and dense models, due to latency constraints, the go-to choice for deployment in real-world retrieval applications. In this work, we question this practice by showing that the number of parameters and early query-document interaction play a significant role in the generalization ability of retrieval models. Our experiments show that increasing model size results in marginal gains on in-domain test sets, but much larger gains in new domains never seen during fine-tuning. Furthermore, we show that rerankers largely outperform dense ones of similar size in several tasks. Our largest reranker reaches the state of the art in 12 of the 18 datasets of the Benchmark-IR (BEIR) and surpasses the previous state of the art by 3 average points. Finally, we confirm that in-domain effectiveness is not a good indicator of zero-shot effectiveness. Code is available at https://github.com/guilhermemr04/scaling-zero-shot-retrieval.git

Guilherme Rosa, Luiz Bonifacio, Vitor Jeronymo et al.

Bi-encoders and cross-encoders are widely used in many state-of-the-art retrieval pipelines. In this work we study the generalization ability of these two types of architectures on a wide range of parameter count on both in-domain and out-of-domain scenarios. We find that the number of parameters and early query-document interactions of cross-encoders play a significant role in the generalization ability of retrieval models. Our experiments show that increasing model size results in marginal gains on in-domain test sets, but much larger gains in new domains never seen during fine-tuning. Furthermore, we show that cross-encoders largely outperform bi-encoders of similar size in several tasks. In the BEIR benchmark, our largest cross-encoder surpasses a state-of-the-art bi-encoder by more than 4 average points. Finally, we show that using bi-encoders as first-stage retrievers provides no gains in comparison to a simpler retriever such as BM25 on out-of-domain tasks. The code is available at https://github.com/guilhermemr04/scaling-zero-shot-retrieval.git

Guilherme Moraes Rosa, Luiz Bonifacio, Vitor Jeronymo et al.

Recent work has shown that language models scaled to billions of parameters, such as GPT-3, perform remarkably well in zero-shot and few-shot scenarios. In this work, we experiment with zero-shot models in the legal case entailment task of the COLIEE 2022 competition. Our experiments show that scaling the number of parameters in a language model improves the F1 score of our previous zero-shot result by more than 6 points, suggesting that stronger zero-shot capability may be a characteristic of larger models, at least for this task. Our 3B-parameter zero-shot model outperforms all models, including ensembles, in the COLIEE 2021 test set and also achieves the best performance of a single model in the COLIEE 2022 competition, second only to the ensemble composed of the 3B model itself and a smaller version of the same model. Despite the challenges posed by large language models, mainly due to latency constraints in real-time applications, we provide a demonstration of our zero-shot monoT5-3b model being used in production as a search engine, including for legal documents. The code for our submission and the demo of our system are available at https://github.com/neuralmind-ai/coliee and https://neuralsearchx.neuralmind.ai, respectively.

Mirelle Bueno, Carlos Gemmell, Jeffrey Dalton et al.

The ability to extrapolate, i.e., to make predictions on sequences that are longer than those presented as training examples, is a challenging problem for current deep learning models. Recent work shows that this limitation persists in state-of-the-art Transformer-based models. Most solutions to this problem use specific architectures or training methods that do not generalize to other tasks. We demonstrate that large language models can succeed in extrapolation without modifying their architecture or training procedure. Our experimental results show that generating step-by-step rationales and introducing marker tokens are both required for effective extrapolation. First, we induce a language model to produce step-by-step rationales before outputting the answer to effectively communicate the task to the model. However, as sequences become longer, we find that current models struggle to keep track of token positions. To address this issue, we interleave output tokens with markup tokens that act as explicit positional and counting symbols. Our findings show how these two complementary approaches enable remarkable sequence extrapolation and highlight a limitation of current architectures to effectively generalize without explicit surface form guidance. Code available at https://github.com/MirelleB/induced-rationales-markup-tokens

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

Vitor Jeronymo, Mauricio Nascimento, Roberto Lotufo et al.

Robust 2004 is an information retrieval benchmark whose large number of judgments per query make it a reliable evaluation dataset. In this paper, we present mRobust04, a multilingual version of Robust04 that was translated to 8 languages using Google Translate. We also provide results of three different multilingual retrievers on this dataset. The dataset is available at https://huggingface.co/datasets/unicamp-dl/mrobust

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.

Hugo Abonizio, Filipe Rocha Lopes, Roberto Lotufo et al.

Brazil's Unified Health System (SUS) relies on official clinical guidelines that define diagnostic criteria, treatments, dosages, and monitoring procedures for over 200 million citizens. Yet current LLMs perform poorly on this guideline-specific knowledge, and no benchmark evaluates clinical recall grounded in Brazilian Portuguese protocols. We address this gap by adapting Qwen2.5-14B-Instruct to the Brazilian clinical domain. From 178 official guidelines (~5.4M tokens), we generate ~70M tokens of synthetic data in three formats -- rephrases, wiki-style articles, and question-answer pairs -- using four generator LLMs. We then apply continual pre-training followed by Group Relative Policy Optimization (GRPO). We introduce HealthBench-BR, with 1,780 balanced true/false clinical assertions, and PCDT-QA, with 890 open-ended clinical questions scored by an LLM judge. Our best model achieves 83.9% on HealthBench-BR and 85.4% on PCDT-QA, outperforming GPT-5.2, Claude Sonnet 4.6, Gemini 3.1 Pro, and Google AI Overview's web-grounded RAG despite having only 14B parameters. Ablations show that generator diversity and reinforcement learning are critical to these gains. We release all datasets, benchmarks, and model weights to support reproducible clinical NLP research for Brazilian Portuguese. Code, data, and model weights are available at https://github.com/hugoabonizio/clinical-protocols-br

Jimmy Lin, David Alfonso-Hermelo, Vitor Jeronymo et al.

The advent of multilingual language models has generated a resurgence of interest in cross-lingual information retrieval (CLIR), which is the task of searching documents in one language with queries from another. However, the rapid pace of progress has led to a confusing panoply of methods and reproducibility has lagged behind the state of the art. In this context, our work makes two important contributions: First, we provide a conceptual framework for organizing different approaches to cross-lingual retrieval using multi-stage architectures for mono-lingual retrieval as a scaffold. Second, we implement simple yet effective reproducible baselines in the Anserini and Pyserini IR toolkits for test collections from the TREC 2022 NeuCLIR Track, in Persian, Russian, and Chinese. Our efforts are built on a collaboration of the two teams that submitted the most effective runs to the TREC evaluation. These contributions provide a firm foundation for future advances.

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

Fernando Ferraretto, Thiago Laitz, Roberto Lotufo et al.

Recent work has shown that inducing a large language model (LLM) to generate explanations prior to outputting an answer is an effective strategy to improve performance on a wide range of reasoning tasks. In this work, we show that neural rankers also benefit from explanations. We use LLMs such as GPT-3.5 to augment retrieval datasets with explanations and train a sequence-to-sequence ranking model to output a relevance label and an explanation for a given query-document pair. Our model, dubbed ExaRanker, finetuned on a few thousand examples with synthetic explanations performs on par with models finetuned on 3x more examples without explanations. Furthermore, the ExaRanker model incurs no additional computational cost during ranking and allows explanations to be requested on demand.

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.

Hugo Abonizio, Leandro Rodrigues de Souza, Roberto Lotufo et al.

The zero-shot cross-lingual ability of models pretrained on multilingual and even monolingual corpora has spurred many hypotheses to explain this intriguing empirical result. However, due to the costs of pretraining, most research uses public models whose pretraining methodology, such as the choice of tokenization, corpus size, and computational budget, might differ drastically. When researchers pretrain their own models, they often do so under a constrained budget, and the resulting models might underperform significantly compared to SOTA models. These experimental differences led to various inconsistent conclusions about the nature of the cross-lingual ability of these models. To help further research on the topic, we released 10 monolingual byte-level models rigorously pretrained under the same configuration with a large compute budget (equivalent to 420 days on a V100) and corpora that are 4 times larger than the original BERT's. Because they are tokenizer-free, the problem of unseen token embeddings is eliminated, thus allowing researchers to try a wider range of cross-lingual experiments in languages with different scripts. Additionally, we release two models pretrained on non-natural language texts that can be used in sanity-check experiments. Experiments on QA and NLI tasks show that our monolingual models achieve competitive performance to the multilingual one, and hence can be served to strengthen our understanding of cross-lingual transferability in language models.

Jayr Pereira, Rodrigo Nogueira, Cleber Zanchettin et al.

Individuals with complex communication needs (CCN) often rely on augmentative and alternative communication (AAC) systems to have conversations and communique their wants. Such systems allow message authoring by arranging pictograms in sequence. However, the difficulty of finding the desired item to complete a sentence can increase as the user's vocabulary increases. This paper proposes using BERTimbau, a Brazilian Portuguese version of BERT, for pictogram prediction in AAC systems. To finetune BERTimbau, we constructed an AAC corpus for Brazilian Portuguese to use as a training corpus. We tested different approaches to representing a pictogram for prediction: as a word (using pictogram captions), as a concept (using a dictionary definition), and as a set of synonyms (using related terms). We also evaluated the usage of images for pictogram prediction. The results demonstrate that using embeddings computed from the pictograms' caption, synonyms, or definitions have a similar performance. Using synonyms leads to lower perplexity, but using captions leads to the highest accuracies. This paper provides insight into how to represent a pictogram for prediction using a BERT-like model and the potential of using images for pictogram prediction.

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.

Roseval Malaquias Junior, Ramon Pires, Roseli Romero et al.

The high compute cost associated with pretraining large language models limits their research. Two strategies have emerged to address this issue: domain specialization and pretraining with high-quality data. To explore these strategies, we specialized the Mistral-7B model with 1.9 billion unique tokens from reputable Brazilian legal sources and conducted few-shot evaluations on legal and general knowledge test suites. Our model, Juru, demonstrates the benefits of domain specialization by achieving improved performance on legal benchmarks, even with a reduced amount of pretraining data. However, this domain specialization through continued pretraining comes at the cost of increased forgetting in unrelated domains, as evidenced by performance degradation on general knowledge test suites in both Portuguese and English. This study contributes to the growing body of scientific evidence showing that pretraining data selection may enhance the performance of large language models, enabling the exploration of these models at a lower cost. Juru is publicly available at https://huggingface.co/roseval/Juru-7B .

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.

Mirelle Bueno, Roberto Lotufo, Rodrigo Nogueira

Language models are now capable of solving tasks that require dealing with long sequences consisting of hundreds of thousands of tokens. However, they often fail on tasks that require repetitive use of simple rules, even on sequences that are much shorter than those seen during training. For example, state-of-the-art LLMs can find common items in two lists with up to 20 items but fail when lists have 80 items. In this paper, we introduce Lissard, a benchmark comprising seven tasks whose goal is to assess the ability of models to process and generate wide-range sequence lengths, requiring repetitive procedural execution. Our evaluation of open-source (Mistral-7B and Mixtral-8x7B) and proprietary models (GPT-3.5 and GPT-4) show a consistent decline in performance across all models as the complexity of the sequence increases. The datasets and code are available at https://github.com/unicamp-dl/Lissard

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.

Fernando Ferraretto, Thiago Laitz, Roberto Lotufo et al.

ExaRanker recently introduced an approach to training information retrieval (IR) models, incorporating natural language explanations as additional labels. The method addresses the challenge of limited labeled examples, leading to improvements in the effectiveness of IR models. However, the initial results were based on proprietary language models such as GPT-3.5, which posed constraints on dataset size due to its cost and data privacy. In this paper, we introduce ExaRanker-Open, where we adapt and explore the use of open-source language models to generate explanations. The method has been tested using different LLMs and datasets sizes to better comprehend the effective contribution of data augmentation. Our findings reveal that incorporating explanations consistently enhances neural rankers, with benefits escalating as the LLM size increases. Notably, the data augmentation method proves advantageous even with large datasets, as evidenced by ExaRanker surpassing the target baseline by 0.6 nDCG@10 points in our study. To encourage further advancements by the research community, we have open-sourced both the code and datasets at https://github.com/unicamp-dl/ExaRanker.

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.

Arthur dos Santos, Jayr Pereira, Rodrigo Nogueira et al.

The increasing number of scientific publications in acoustics, in general, presents difficulties in conducting traditional literature surveys. This work explores the use of a generative pre-trained transformer (GPT) model to automate a literature survey of 116 articles on data-driven speech enhancement methods. The main objective is to evaluate the capabilities and limitations of the model in providing accurate responses to specific queries about the papers selected from a reference human-based survey. While we see great potential to automate literature surveys in acoustics, improvements are needed to address technical questions more clearly and accurately.

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.

Hugo Abonizio, Thales Almeida, Roberto Lotufo et al.

Large language models (LLMs) often require vast amounts of text to effectively acquire new knowledge. While continuing pre-training on large corpora or employing retrieval-augmented generation (RAG) has proven successful, updating an LLM with only a few thousand or million tokens remains challenging. In this work, we investigate the task of injecting small, unstructured information into LLMs and its relation to the catastrophic forgetting phenomenon. We use a dataset of recent news -- ensuring no overlap with the model's pre-training data -- to evaluate the knowledge acquisition by probing the model with question-answer pairs related the learned information. Starting from a continued pre-training baseline, we explored different augmentation algorithms to generate synthetic data to improve the knowledge acquisition capabilities. Our experiments show that simply continuing pre-training on limited data yields modest improvements, whereas exposing the model to diverse textual variations significantly improves the learning of new facts -- particularly with methods that induce greater variability through diverse prompting. Furthermore, we shed light on the forgetting phenomenon in small-data regimes, illustrating the delicate balance between learning new content and retaining existing capabilities. We also confirm the sensitivity of RAG-based approaches for knowledge injection, which often lead to greater degradation on control datasets compared to parametric methods. Finally, we demonstrate that models can generate effective synthetic training data themselves, suggesting a pathway toward self-improving model updates. All code and generated data used in our experiments are publicly available, providing a resource for studying efficient knowledge injection in LLMs with limited data at https://github.com/hugoabonizio/knowledge-injection-methods.

Marcos Piau, Roberto Lotufo, Rodrigo Nogueira

Despite advancements in Natural Language Processing (NLP) and the growing availability of pretrained models, the English language remains the primary focus of model development. Continued pretraining on language-specific corpora provides a practical solution for adapting models to other languages. However, the impact of different pretraining settings on downstream tasks remains underexplored. This work introduces $\texttt{ptt5-v2}$, investigating the continued pretraining of T5 models for Portuguese. We first develop a baseline set of settings and pretrain models with sizes up to 3B parameters. Finetuning on three Portuguese downstream tasks (assin2 STS, assin2 RTE, and TweetSentBR) yields SOTA results on the latter two. We then explore the effects of different pretraining configurations, including pretraining data quality, optimization strategies, and multi-epoch pretraining. Perhaps surprisingly, their impact remains subtle compared to our baseline. We release $\texttt{ptt5-v2}$ pretrained checkpoints and their MonoT5-based finetuned $\texttt{MonoPTT5}$ rerankers on HuggingFace in their respective collections at \url{https://huggingface.co/unicamp-dl}.

Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee et al.

The information retrieval community has recently witnessed a revolution due to large pretrained transformer models. Another key ingredient for this revolution was the MS MARCO dataset, whose scale and diversity has enabled zero-shot transfer learning to various tasks. However, not all IR tasks and domains can benefit from one single dataset equally. Extensive research in various NLP tasks has shown that using domain-specific training data, as opposed to a general-purpose one, improves the performance of neural models. In this work, we harness the few-shot capabilities of large pretrained language models as synthetic data generators for IR tasks. We show that models finetuned solely on our unsupervised dataset outperform strong baselines such as BM25 as well as recently proposed self-supervised dense retrieval methods. Furthermore, retrievers finetuned on both supervised and our synthetic data achieve better zero-shot transfer than models finetuned only on supervised data. Code, models, and data are available at https://github.com/zetaalphavector/inpars .

Guilherme Moraes Rosa, Ruan Chaves Rodrigues, Roberto de Alencar Lotufo et al.

There has been mounting evidence that pretrained language models fine-tuned on large and diverse supervised datasets can transfer well to a variety of out-of-domain tasks. In this work, we investigate this transfer ability to the legal domain. For that, we participated in the legal case entailment task of COLIEE 2021, in which we use such models with no adaptations to the target domain. Our submissions achieved the highest scores, surpassing the second-best team by more than six percentage points. Our experiments confirm a counter-intuitive result in the new paradigm of pretrained language models: given limited labeled data, models with little or no adaptation to the target task can be more robust to changes in the data distribution than models fine-tuned on it. Code is available at https://github.com/neuralmind-ai/coliee.

Luiz Bonifacio, Vitor Jeronymo, Hugo Queiroz Abonizio et al.

The MS MARCO ranking dataset has been widely used for training deep learning models for IR tasks, achieving considerable effectiveness on diverse zero-shot scenarios. However, this type of resource is scarce in languages other than English. In this work, we present mMARCO, a multilingual version of the MS MARCO passage ranking dataset comprising 13 languages that was created using machine translation. We evaluated mMARCO by finetuning monolingual and multilingual reranking models, as well as a multilingual dense retrieval model on this dataset. We also evaluated models finetuned using the mMARCO dataset in a zero-shot scenario on Mr. TyDi dataset, demonstrating that multilingual models finetuned on our translated dataset achieve superior effectiveness to models finetuned on the original English version alone. Our experiments also show that a distilled multilingual reranker is competitive with non-distilled models while having 5.4 times fewer parameters. Lastly, we show a positive correlation between translation quality and retrieval effectiveness, providing evidence that improvements in translation methods might lead to improvements in multilingual information retrieval. The translated datasets and finetuned models are available at https://github.com/unicamp-dl/mMARCO.

Guilherme Moraes Rosa, Luiz Henrique Bonifacio, Leandro Rodrigues de Souza et al.

An effective method for cross-lingual transfer is to fine-tune a bilingual or multilingual model on a supervised dataset in one language and evaluating it on another language in a zero-shot manner. Translating examples at training time or inference time are also viable alternatives. However, there are costs associated with these methods that are rarely addressed in the literature. In this work, we analyze cross-lingual methods in terms of their effectiveness (e.g., accuracy), development and deployment costs, as well as their latencies at inference time. Our experiments on three tasks indicate that the best cross-lingual method is highly task-dependent. Finally, by combining zero-shot and translation methods, we achieve the state-of-the-art in two of the three datasets used in this work. Based on these results, we question the need for manually labeled training data in a target language. Code and translated datasets are available at https://github.com/unicamp-dl/cross-lingual-analysis

Guilherme Moraes Rosa, Ruan Chaves Rodrigues, Roberto Lotufo et al.

We describe our single submission to task 1 of COLIEE 2021. Our vanilla BM25 got second place, well above the median of submissions. Code is available at https://github.com/neuralmind-ai/coliee.

Rodrigo Nogueira, Zhiying Jiang, Jimmy Lin

The ability to perform arithmetic tasks is a remarkable trait of human intelligence and might form a critical component of more complex reasoning tasks. In this work, we investigate if the surface form of a number has any influence on how sequence-to-sequence language models learn simple arithmetic tasks such as addition and subtraction across a wide range of values. We find that how a number is represented in its surface form has a strong influence on the model's accuracy. In particular, the model fails to learn addition of five-digit numbers when using subwords (e.g., "32"), and it struggles to learn with character-level representations (e.g., "3 2"). By introducing position tokens (e.g., "3 10e1 2"), the model learns to accurately add and subtract numbers up to 60 digits. We conclude that modern pretrained language models can easily learn arithmetic from very few examples, as long as we use the proper surface representation. This result bolsters evidence that subword tokenizers and positional encodings are components in current transformer designs that might need improvement. Moreover, we show that regardless of the number of parameters and training examples, models cannot learn addition rules that are independent of the length of the numbers seen during training. Code to reproduce our experiments is available at https://github.com/castorini/transformers-arithmetic

Ronak Pradeep, Rodrigo Nogueira, Jimmy Lin

We propose a design pattern for tackling text ranking problems, dubbed "Expando-Mono-Duo", that has been empirically validated for a number of ad hoc retrieval tasks in different domains. At the core, our design relies on pretrained sequence-to-sequence models within a standard multi-stage ranking architecture. "Expando" refers to the use of document expansion techniques to enrich keyword representations of texts prior to inverted indexing. "Mono" and "Duo" refer to components in a reranking pipeline based on a pointwise model and a pairwise model that rerank initial candidates retrieved using keyword search. We present experimental results from the MS MARCO passage and document ranking tasks, the TREC 2020 Deep Learning Track, and the TREC-COVID challenge that validate our design. In all these tasks, we achieve effectiveness that is at or near the state of the art, in some cases using a zero-shot approach that does not exploit any training data from the target task. To support replicability, implementations of our design pattern are open-sourced in the Pyserini IR toolkit and PyGaggle neural reranking library.

Gabriela Surita, Rodrigo Nogueira, Roberto Lotufo

What are the latent questions on some textual data? In this work, we investigate using question generation models for exploring a collection of documents. Our method, dubbed corpus2question, consists of applying a pre-trained question generation model over a corpus and aggregating the resulting questions by frequency and time. This technique is an alternative to methods such as topic modelling and word cloud for summarizing large amounts of textual data. Results show that applying corpus2question on a corpus of scientific articles related to COVID-19 yields relevant questions about the topic. The most frequent questions are "what is covid 19" and "what is the treatment for covid". Among the 1000 most frequent questions are "what is the threshold for herd immunity" and "what is the role of ace2 in viral entry". We show that the proposed method generated similar questions for 13 of the 27 expert-made questions from the CovidQA question answering dataset. The code to reproduce our experiments and the generated questions are available at: https://github.com/unicamp-dl/corpus2question

Diedre Carmo, Marcos Piau, Israel Campiotti et al.

In natural language processing (NLP), there is a need for more resources in Portuguese, since much of the data used in the state-of-the-art research is in other languages. In this paper, we pretrain a T5 model on the BrWac corpus, an extensive collection of web pages in Portuguese, and evaluate its performance against other Portuguese pretrained models and multilingual models on three different tasks. We show that our Portuguese pretrained models have significantly better performance over the original T5 models. Moreover, we demonstrate the positive impact of using a Portuguese vocabulary. Our code and models are available at https://github.com/unicamp-dl/PTT5.

Alexandre Lopes, Rodrigo Nogueira, Roberto Lotufo et al.

Despite the widespread adoption of deep learning for machine translation, it is still expensive to develop high-quality translation models. In this work, we investigate the use of pre-trained models, such as T5 for Portuguese-English and English-Portuguese translation tasks using low-cost hardware. We explore the use of Portuguese and English pre-trained language models and propose an adaptation of the English tokenizer to represent Portuguese characters, such as diaeresis, acute and grave accents. We compare our models to the Google Translate API and MarianMT on a subset of the ParaCrawl dataset, as well as to the winning submission to the WMT19 Biomedical Translation Shared Task. We also describe our submission to the WMT20 Biomedical Translation Shared Task. Our results show that our models have a competitive performance to state-of-the-art models while being trained on modest hardware (a single 8GB gaming GPU for nine days). Our data, models and code are available at https://github.com/unicamp-dl/Lite-T5-Translation.

Paulo Finardi, Israel Campiotti, Gustavo Plensack et al.

In this work we propose a novel self-attention mechanism model to address electricity theft detection on an imbalanced realistic dataset that presents a daily electricity consumption provided by State Grid Corporation of China. Our key contribution is the introduction of a multi-head self-attention mechanism concatenated with dilated convolutions and unified by a convolution of kernel size $1$. Moreover, we introduce a binary input channel (Binary Mask) to identify the position of the missing values, allowing the network to learn how to deal with these values. Our model achieves an AUC of $0.926$ which is an improvement in more than $17\%$ with respect to previous baseline work. The code is available on GitHub at https://github.com/neuralmind-ai/electricity-theft-detection-with-self-attention.

Rodrigo Nogueira, Kyunghyun Cho

Recently, neural models pretrained on a language modeling task, such as ELMo (Peters et al., 2017), OpenAI GPT (Radford et al., 2018), and BERT (Devlin et al., 2018), have achieved impressive results on various natural language processing tasks such as question-answering and natural language inference. In this paper, we describe a simple re-implementation of BERT for query-based passage re-ranking. Our system is the state of the art on the TREC-CAR dataset and the top entry in the leaderboard of the MS MARCO passage retrieval task, outperforming the previous state of the art by 27% (relative) in MRR@10. The code to reproduce our results is available at https://github.com/nyu-dl/dl4marco-bert

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.

Jayr Pereira, Andre Assumpcao, Julio Trecenti et al.

This paper introduces INACIA (Instrução Assistida com Inteligência Artificial), a groundbreaking system designed to integrate Large Language Models (LLMs) into the operational framework of Brazilian Federal Court of Accounts (TCU). The system automates various stages of case analysis, including basic information extraction, admissibility examination, Periculum in mora and Fumus boni iuris analyses, and recommendations generation. Through a series of experiments, we demonstrate INACIA's potential in extracting relevant information from case documents, evaluating its legal plausibility, and formulating propositions for judicial decision-making. Utilizing a validation dataset alongside LLMs, our evaluation methodology presents a novel approach to assessing system performance, correlating highly with human judgment. These results underscore INACIA's potential in complex legal task handling while also acknowledging the current limitations. This study discusses possible improvements and the broader implications of applying AI in legal contexts, suggesting that INACIA represents a significant step towards integrating AI in legal systems globally, albeit with cautious optimism grounded in the empirical findings.

Ramon Pires, Roseval Malaquias Junior, Rodrigo Nogueira

Despite the recent advances in Large Language Models, benchmarks for evaluating legal writing remain scarce due to the inherent complexity of assessing open-ended responses in this domain. One of the key challenges in evaluating language models on domain-specific tasks is finding test datasets that are public, frequently updated, and contain comprehensive evaluation guidelines. The Brazilian Bar Examination meets these requirements. We introduce oab-bench, a benchmark comprising 105 questions across seven areas of law from recent editions of the exam. The benchmark includes comprehensive evaluation guidelines and reference materials used by human examiners to ensure consistent grading. We evaluate the performance of four LLMs on oab-bench, finding that Claude-3.5 Sonnet achieves the best results with an average score of 7.93 out of 10, passing all 21 exams. We also investigated whether LLMs can serve as reliable automated judges for evaluating legal writing. Our experiments show that frontier models like OpenAI's o1 achieve a strong correlation with human scores when evaluating approved exams, suggesting their potential as reliable automated evaluators despite the inherently subjective nature of legal writing assessment. The source code and the benchmark -- containing questions, evaluation guidelines, model-generated responses, and their respective automated evaluations -- are publicly available.

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.