Institution profile

Federal University of Juiz de Fora

Academic institutionsouthamerica · br
Official website
Research library7linked papers
Opportunities0open roles
Selected work

Representative Papers

Adversarial versification in portuguese as a jailbreak operator in LLMs

Dec 17, 2025

This study identifies a critical gap in multilingual adversarial evaluation—specifically, the vulnerability of large language models (LLMs) to poetic prompt rewriting in morphologically rich, prosodically constrained languages like Portuguese. We propose the first Portuguese-specific adversarial versification jailbreak framework, integrating metrical scansion, parametric meter modeling, and Lusophone prosodic variation. Our method combines rule-based and LLM-augmented generation, prosodic scansion analysis, adaptation to the AILuminate benchmark, and cross-model robustness evaluation. Human-crafted Portuguese poetic attacks achieve a 62% success rate, while automated variants reach 43%; certain models exhibit single-turn jailbreak rates exceeding 90%. Empirical results demonstrate that mainstream alignment techniques—including RLHF and Constitutional AI—fail significantly under Portuguese prosodic perturbations. This work establishes the first systematic investigation of verse-based jailbreaking in a high-inflection, strong-metre language, revealing fundamental limitations in current multilingual safety alignment.

0 citationsRead paper

A Service Suite for Specifying Digital Twins for Industry 5.0

Nov 10, 2025

To address predictive maintenance requirements in Industry 5.0, this paper proposes DT-Create—a digital twin modeling service suite enabling real-time virtual mirroring of physical assets and dynamic decision support. Methodologically, it integrates machine learning with ontology-driven knowledge graphs to construct semantically enriched digital twins; introduces an adaptive mechanism that dynamically selects optimal prediction models based on data characteristics and supports online model updating and inference. Following the design science research paradigm, DT-Create unifies multi-source sensor data acquisition, semantic modeling, machine learning, and logical reasoning. Empirical validation demonstrates significant improvements in data interpretability, model adaptation efficiency, and decision autonomy, confirming its engineering feasibility. The core contribution is a novel adaptive twin modeling framework that orchestrates semantic representation, data processing, and model selection in tight synergy.

0 citationsRead paper

Pragmatic Frames Evoked by Gestures: A FrameNet Brasil Approach to Multimodality in Turn Organization

Sep 11, 2025

Multimodal dialogue systems lack robust modeling of how linguistic and gestural cues jointly regulate turn-taking, particularly due to the absence of systematic, functionally grounded gesture annotations. Method: We introduce the first turn-function-oriented gesture annotation schema on the real-world Frame2 corpus, uncovering previously undocumented gesture variants and their cognitive motivations. Drawing on FrameNet Brasil, we integrate mental spaces theory, conceptual blending, and metaphor theory to develop a dual-layer semantic–pragmatic annotation framework. Contribution/Results: Our analysis confirms that co-speech gestures serve as critical pragmatic turn-taking cues. The resulting annotation scheme significantly enhances both cognitive interpretability and computational utility for multimodal dialogue structure modeling. It establishes a novel paradigm for modeling human interactive cognition and advancing multimodal human–machine interaction.

0 citationsRead paper
Recent publications

Latest Papers

Adversarial versification in portuguese as a jailbreak operator in LLMs

Dec 17, 2025

This study identifies a critical gap in multilingual adversarial evaluation—specifically, the vulnerability of large language models (LLMs) to poetic prompt rewriting in morphologically rich, prosodically constrained languages like Portuguese. We propose the first Portuguese-specific adversarial versification jailbreak framework, integrating metrical scansion, parametric meter modeling, and Lusophone prosodic variation. Our method combines rule-based and LLM-augmented generation, prosodic scansion analysis, adaptation to the AILuminate benchmark, and cross-model robustness evaluation. Human-crafted Portuguese poetic attacks achieve a 62% success rate, while automated variants reach 43%; certain models exhibit single-turn jailbreak rates exceeding 90%. Empirical results demonstrate that mainstream alignment techniques—including RLHF and Constitutional AI—fail significantly under Portuguese prosodic perturbations. This work establishes the first systematic investigation of verse-based jailbreaking in a high-inflection, strong-metre language, revealing fundamental limitations in current multilingual safety alignment.

0 citationsRead paper

A Service Suite for Specifying Digital Twins for Industry 5.0

Nov 10, 2025

To address predictive maintenance requirements in Industry 5.0, this paper proposes DT-Create—a digital twin modeling service suite enabling real-time virtual mirroring of physical assets and dynamic decision support. Methodologically, it integrates machine learning with ontology-driven knowledge graphs to construct semantically enriched digital twins; introduces an adaptive mechanism that dynamically selects optimal prediction models based on data characteristics and supports online model updating and inference. Following the design science research paradigm, DT-Create unifies multi-source sensor data acquisition, semantic modeling, machine learning, and logical reasoning. Empirical validation demonstrates significant improvements in data interpretability, model adaptation efficiency, and decision autonomy, confirming its engineering feasibility. The core contribution is a novel adaptive twin modeling framework that orchestrates semantic representation, data processing, and model selection in tight synergy.

0 citationsRead paper

Pragmatic Frames Evoked by Gestures: A FrameNet Brasil Approach to Multimodality in Turn Organization

Sep 11, 2025

Multimodal dialogue systems lack robust modeling of how linguistic and gestural cues jointly regulate turn-taking, particularly due to the absence of systematic, functionally grounded gesture annotations. Method: We introduce the first turn-function-oriented gesture annotation schema on the real-world Frame2 corpus, uncovering previously undocumented gesture variants and their cognitive motivations. Drawing on FrameNet Brasil, we integrate mental spaces theory, conceptual blending, and metaphor theory to develop a dual-layer semantic–pragmatic annotation framework. Contribution/Results: Our analysis confirms that co-speech gestures serve as critical pragmatic turn-taking cues. The resulting annotation scheme significantly enhances both cognitive interpretability and computational utility for multimodal dialogue structure modeling. It establishes a novel paradigm for modeling human interactive cognition and advancing multimodal human–machine interaction.

0 citationsRead paper