Eleven Years of BRACIS: A Meta-Scientific Study of the Brazilian Conference on Intelligent Systems

arXiv:2608.09964 · cs.DL, cs.AI · Submitted 2026-07-26 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Eleven Years of BRACIS: A Meta-Scientific Study of the Brazilian Conference on Intelligent Systems".

Jane: The paper was written by Thales Sales Almeida, Giovana Kerche Bonás, Thiago Laitz, João Guilherme Alves Santos, Hugo Abonizio et al. from Tropic AI and Maritaca AI.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: Alright, listeners, welcome back. Today we're digging into a paper that's not about building a smarter model or a faster algorithm, but about understanding the very conference where all that work gets shared. It's called "Eleven Years of BRACIS: A Meta-Scientific Study of the Brazilian Conference on Intelligent Systems."

Jane: And Tom, I love this because it's a paper about papers. Meta-science, right? It's like stepping back and asking, "Who's publishing here, what are they publishing, and does anyone actually read it?" The authors looked at every single paper published at BRACIS from two thousand fifteen to two thousand twenty-five — that's over a thousand papers.

Tom: Over a thousand, yeah. And they didn't just count titles. They pulled the full text of each paper, grabbed citation data from Google Scholar, and even used an LLM to classify what each paper was about. That's the part that gets me excited — they're using the same tools they're studying to study the tools.

Jane: Exactly. And the title tells you it's a meta-scientific study, which just means they're applying scientific methods to understand the scientific community itself. They wanted to know things like, what topics dominate? Who are the authors? Which papers actually get cited and why?

Tom: And one of the first things they found — and this is a big one — is that the conference has shifted dramatically. In two thousand fifteen over half the papers were about proposing new algorithms. By two thousand twenty-five that number had dropped to under a quarter. Meanwhile, papers proposing new models went from twenty-seven percent to forty-six percent.

Jane: So the community moved from "here's a clever algorithm" to "here's a model we trained." And that tracks with what we see everywhere in AI — deep learning took over.

Tom: Right, and the LLM boom is right there in the data too. Before two thousand twenty zero papers at BRACIS had anything to do with large language models. By two thousand twenty-four nearly one in five papers was LLM-related.

Jane: That's a huge shift in just four years. And it makes you wonder — is BRACIS following the global trend, or is it leading something locally?

Tom: That's the question we'll dig into next. But the title alone tells you this is a conference that's willing to look in the mirror. And that's rare.

Summary: Tom: So we've set the stage — "Eleven Years of BRACIS" is a deep dive into the Brazilian Conference on Intelligent Systems. Now let's talk about what they actually found when they looked at the community itself.

Jane: And this is where it gets really interesting, Tom. The paper paints a picture of a conference with a very specific shape. They call it "hourglass-shaped." Here's what that means: eighty percent of the authors who published at BRACIS appeared in only one edition. They came, they published, they left.

Tom: But the institutions — the universities and labs — they stuck around. Nearly half of the institutions returned year after year. So you have this constant churn of individual researchers, but the institutions provide the continuity.

Jane: It's like a restaurant with a revolving door of chefs, but the kitchen itself stays the same. And that makes sense when you think about how Brazilian academia works — professors stay, students graduate and move on.

Tom: And speaking of who's publishing, the paper found that the top ten academic institutions are all Brazilian public universities. USP leads with one hundred ninety papers over the decade. That's a lot of work coming out of one place.

Jane: But here's the thing that caught my eye — the Northeast region is growing the fastest. Participation from that region grew three point four times over the decade, outpacing the conference's overall growth. So the center of gravity is shifting.

Tom: And it's not just geography. The paper also looked at gender. Only about fourteen percent of first authors were women, and that number didn't really improve over the decade. But the last author — usually the senior researcher, the advisor — that rate was higher, around twenty-five percent, though it's been drifting downward.

Lu: If I can jump in here — that gender gap is concerning, but it's consistent with what we see at other Brazilian conferences. The paper cites a study of thirteen SBC conferences that found similar rates. So BRACIS isn't an outlier; it's part of a broader pattern.

Jane: Right, Lu. And the paper doesn't just report these numbers — it connects them. The high turnover of authors, the stability of institutions, the regional growth — together they tell a story about how a national research community sustains itself.

Tom: And sustains it does. Over a thousand papers in eleven years, with a stable core of institutions anchoring the whole thing. But here's the question that keeps me up at night — does anyone actually read these papers? We'll get to citations next.

Improvements: Tom: We're back with "Eleven Years of BRACIS," and now we're getting to the part that really matters — what does this paper tell us about how the conference could be better?

Jane: And the biggest red flag they found is access. BRACIS proceedings are published behind IEEE and Springer paywalls. That means if you're not at a university with a subscription, you can't read most of these papers.

Tom: And get this — only seven point four percent of the papers have an arXiv preprint. So for the vast majority of this work, there's no free version anywhere. The paper calls this a "likely bottleneck" for the conference's impact.

Meng: As someone who works in industry, I can tell you that's a real problem. If I can't access a paper easily, I'm probably not going to cite it, and I'm definitely not going to build on it. Paywalls are a friction point that kills adoption.

Jane: Exactly, Meng. And the data backs that up. The paper found that papers with an arXiv preprint get significantly more citations — a median of ten versus four for papers without one. That's a massive difference.

Tom: But here's the nuance — the paper is careful to say this is correlation, not causation. Maybe authors who post to arXiv are just more likely to promote their work, or maybe they're at institutions that value that kind of sharing. But either way, the signal is clear.

Lu: And there's another improvement they highlight. The paper found that only twenty-two percent of BRACIS papers cite other BRACIS papers. The intra-conference citation network is sparse. That means the community isn't building on itself as much as it could.

Tom: Right, Lu. And when they do cite each other, it's concentrated on a few hub papers — BERTimbau gets cited by forty-one other BRACIS papers. That's a Portuguese-language BERT model, by the way, and it's the conference's biggest success story with over a thousand citations.

Jane: So the improvement here is twofold. First, make the papers open access so people can actually read them. Second, encourage more internal citation — build on the work that's already been done at the conference rather than starting from scratch every time.

Meng: And honestly, the open access piece is the one that would have the biggest immediate impact. The paper shows that artifact release — sharing code and data — went from nine percent of papers in two thousand fifteen to fifty-seven percent in two thousand twenty-three. That's a huge cultural shift. But it doesn't matter if the paper itself is locked away.

Tom: So the recipe is clear: open the papers, keep sharing the code, and cite each other more. Simple to say, harder to do. But the data gives us a roadmap.

Conclusion: Tom: And that brings us to the end of our look at "Eleven Years of BRACIS: A Meta-Scientific Study of the Brazilian Conference on Intelligent Systems." What a ride this paper has been.

Jane: It really has, Tom. We started by understanding what the conference publishes — mostly machine learning, NLP, and computer vision, with LLM research exploding in the last few years. Then we looked at who publishes — a stable core of Brazilian public universities, with a revolving door of individual authors and the Northeast region growing fastest.

Tom: And finally, we saw the impact — or the lack of it. Citations are heavily concentrated, with the top one percent of papers carrying twenty-seven percent of all citations. And the biggest barrier to more impact is simply that most of this work is locked behind paywalls.

Jane: But the paper ends on a constructive note. It suggests moving to an open-access model, which several other Brazilian computer science conferences already use. And it releases the dataset and pipeline so anyone can run this same analysis on future editions.

Lu: That's the part I find most valuable. This isn't just a one-time audit. It's a tool that lets the community track its own health over time. Every year, you could run this pipeline and see how the conference is evolving.

Meng: And from a practical standpoint, the arXiv finding is the one I'll remember. If you want your BRACIS paper to get cited, put it on arXiv. That's a concrete, actionable takeaway for any researcher.

Tom: Absolutely, Meng. And it's a testament to the authors — they didn't just criticize the conference, they gave it a path forward. That's what good meta-science should do.

Jane: So here's to BRACIS, to Brazilian AI research, and to the people who take the time to study their own community. We'll be back next time with another paper, but for now, this one's going on the shelf.

Tom: And remember, listeners — if you take one thing from this episode, let it be this: open access matters. It's not just about being generous. It's about making your work actually count. See you next time.

Thales Sales Almeida, Giovana Kerche Bonás, Thiago Laitz, João Guilherme Alves Santos, Hugo Abonizio, Roseval Malaquias Junior, Marcos Piau, Celio Larcher, Ramon Pires, Rodrigo Nogueira

Tropic AI · Maritaca AI

cs.DL, cs.AI

Submitted: 2026-07-26

Updated: 2026-08-12

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 63/100

Terminology

Multi-Agent System — 7× in this paper · explained in 6 episodes
Benchmark — 5× in this paper · explained in 2 episodes
Reinforcement Learning — 5× in this paper · explained in 4 episodes
Verification — 5× in this paper · explained in 2 episodes
Core — 3× in this paper · explained in 2 episodes
Embeddings — 3× in this paper · explained in 2 episodes
Transformer — 3× in this paper · explained in 2 episodes
Clustering — 2× in this paper · explained in 2 episodes
Data Augmentation — 2× in this paper · explained in 7 episodes
Deep Reinforcement Learning — 2× in this paper · explained in 2 episodes
Explainability — 2× in this paper · explained in 3 episodes
Artificial Intelligence (AI) — 1× in this paper · explained in 2 episodes
As Artificial Intelligence — 1× in this paper
Meta-learning — 1× in this paper · explained in 2 episodes
Trace — 1× in this paper · explained in 2 episodes
alpha — 1× in this paper · explained in 1 episode

Summary

Summary

This paper presents a meta-scientific study of the Brazilian Conference on Intelligent Systems (BRACIS), the main national venue for Artificial Intelligence research in Brazil, hosted by the Brazilian Computer Society since 2012. The authors build a per-paper record of all 1,066 accepted papers from 2015 to 2025, using DBLP metadata, 6,765 Google Scholar citations, and the paper full texts, to investigate what BRACIS publishes, who publishes it, and which work gets cited.

The study is organized around three pillars: thematic landscape, community structure, and scientific impact. The paper states: We organize the analysis around the three-pillar layout, similarly to Nunes et al. (2026)’s: thematic landscape, community structure, and scientific impact.

Thematic Landscape: The paper finds that Machine Learning, Natural Language Processing, and Computer Vision jointly account for about half of all BRACIS papers. Specifically, these three areas have 174, 171, and 162 papers respectively, followed by Optimization and Metaheuristics, Data Mining, and Bioinformatics and Healthcare AI. These nine areas capture 87% of all papers. The contribution mix has shifted over the decade: Algorithm contributions fell from 59% in 2015 to 23% in 2025, while Model contributions rose from 27% to 46% and Empirical roughly doubled from 12% to 26%. Large Language Model research rose from zero before 2020 to 19% by 2024, and Portuguese-language work is prominent, with portuguese-language appears in 51 papers and brazilian-portuguese in 24.

Community Structure: The paper reports that 80.5% of BRACIS authors appear in only one edition while year-over-year retention sits at 16%. Across 1,066 papers there are 2,623 unique authors, with 2,112 (80.5%) appearing in only one edition and just 49 (1.9%) appearing in five or more editions. Institutional retention is much higher: Per-edition institutional retention averages 46.5% across 319 distinct institutions, nearly three times the 16% author retention rate. Team sizes have grown modestly: The median paper had 3 authors from 2015 to 2022 and 4 from 2023 onward. The female first-author rate is 13.3% on average, while the female last-author rate averages 25.7% and declines over the decade. The Brazilian Northeast region is the fastest-growing, with participation growing 3.4 times from 2015 to 2025. USP leads volume with 190 papers, and the top 10 academic institutions are uniformly Brazilian public universities. Co-authorship concentrates within Brazilian macroregions, with the Southeast carrying 475 within-region co-authored papers, the Northeast 315, and the South 214.

Scientific Impact: The paper finds that BRACIS citations are heavily skewed: the top 1% of papers carry 27% of all citations. The corpus has a mean of 6.5 citations per paper and a median of 2, with 26% of papers receiving no citations. The corpus h-index is 30, g-index 56, and i10 162. Only four papers exceed 100 citations: BERTimbau (1,008), Hyper-Parameter Tuning of Decision Trees (166), Sabía (153), and Hate Speech Classification (134). The most-cited papers cluster on Portuguese-language NLP and older applied deep learning. Intra-BRACIS citations are rare: 22% of papers cite at least one other BRACIS paper; the mean intra-venue in-degree is 0.31, with BERTimbau cited by 41 other BRACIS papers.

Openness Practices: The paper reports that Artifact sharing rose from 8.9% in 2015 to 57.3% in 2023, with 37.5% of eligible papers linking to at least one public code or data artifact. However, arXiv adoption at BRACIS reaches 7.4%, with only 79 of 1,066 papers having an arXiv preprint. Industry coauthorship has roughly doubled but remains rare at 7.2% of papers. Reflective sections (Future Work, Limitations) adoption more than tripled across the decade, rising from 12.5% in 2015 to 51.7% in 2025 for any reflective section.

Correlational Findings: The paper finds that Among openness signals, only arXiv preprint posting correlates with citations. arXiv-posted papers carry mean 13.2 citations against 8.5 for non-arXiv, with median 10 against 4 and Mann-Whitney p < 10−4. Artifact release and industry coauthorship show no such correlation. Empirical papers have the highest median citation count (median 6) compared to Model and Algorithm papers (median 4). Best-paper nominee status is uncorrelated with eventual citation impact: removing BERTimbau, a 2020 nominee, drops the nominee mean to 5.1, identical to non-nominees.

Conclusion: The paper concludes that BRACIS emerges as an established regional venue with a distinct role in the Brazilian AI ecosystem, with a diversified thematic base, an hourglass community structure, and heavily concentrated citations. The authors note that BRACIS proceedings sit behind IEEE and Springer paywalls... yet only 7.4% of papers have one, pointing to access as a likely bottleneck. They release the dataset and pipeline for future analysis.

Improvements for AI systems

Based on this paper, here are the specific improvements I can make to AI systems and what the improved systems can do:

Improvement: Build a system that automatically identifies papers likely to benefit from arXiv posting by analyzing features correlated with the 7.4% preprint rate and the significant citation premium (mean 13.2 vs 8.5 citations, p<10−4).

What it can do: Automatically flag high-potential BRACIS papers (e.g., Portuguese-NLP resources, empirical studies with artifacts) and prompt authors to post preprints, potentially increasing citation impact by 55% for those papers.

Improvement: Train a domain-adapted model using the paper's finding that Portuguese-language work (BERTimbau, Sabía, Juru, LegalBert-pt) dominates the citation top-20 and achieves 8-168 citations/year, while international venues under-publish such work.

Improvement: Use the 8.9%→57.3% artifact-release growth trend and the finding that artifact release correlates with higher raw citation means (12.5 vs 7.2) to build a classifier that predicts which papers lack public code/data.

Improvement: Leverage the hourglass community structure (80.5% single-edition authors, 46.5% institutional retention) and the sparse co-authorship matrix (only 2 strong pairs: UFRPE-UFPE, IFCE-UFC) to build a graph-based recommender.

Improvement: Use the citations-per-year metric (cpy) to identify papers like Juru (13.0 cpy) and LegalBert-pt (10.7 cpy) that will likely enter the top-20 within 2-3 years, even though they're too recent for total-citation rankings.

Improvement: Train a model on the 31.7% of papers with reflective sections (Future Work/Limitations) to detect missing sections in new submissions.

Improvement: Model the counterfactual: if BRACIS moved to open access (as the paper recommends), estimate citation gains using the arXiv premium (55% mean increase) and the 92.6% of papers currently without preprints.

Improvement: Use the 13.8% female first-author rate and 25.7% female last-author rate (declining to 18.9% by 2025) to build a monitoring dashboard.

Improvement: Use the LLM-topic growth curve (0%→19% in 4 years) and the Algorithm→Model/Empirical shift (59%→23% Algorithm) to forecast future BRACIS topic distributions.

Improvement: Use the sparse intra-BRACIS citation graph (324 edges, 22% of papers citing prior BRACIS work) to build a recommendation engine.

Abstract

The Brazilian Conference on Intelligent Systems (BRACIS) is the main national venue for Artificial Intelligence research in Brazil, hosted by the Brazilian Computer Society since 2012 and publishing work from institutions across the country. Across eleven years, from 2015 to 2025, we build a per-paper record of all 1,066 accepted papers from DBLP metadata, 6,765 Google Scholar citations, and the paper full texts, and use it to ask what BRACIS publishes, who publishes it, and which work gets cited. Large Language Model research grows from zero before 2020 to 19% of papers in 2024, on top of a base of Machine Learning, Computer Vision, and Optimization work. The community is hourglass-shaped: 80.5% of 2,623 authors appear in a single edition, while institutions return at nearly three times the author rate. Citations are heavily concentrated, with the top 1% of papers carrying 27% of the total. Openness practices have grown, with artifact release rising from 8.9% of papers in 2015 to 57.3% in 2023, and we find a notable correlation between having an arXiv preprint and higher citation counts. Since proceedings sit behind IEEE and Springer paywalls and only 7.4% of papers have a preprint, most BRACIS work is hard to reach for readers without institutional access.