We investigate representational alignment between nine independently trained speech encoders and ten text encoders. We build a controlled corpus in which the same sentences are rendered by different speakers and under different emotional conditions, so that prosody varies while linguistic content is held fixed. We compare every speech–text layer pair and calibrate the strongest match with a permutation procedure that prevents the search over pairs from inflating alignment scores. Local neighborhood agreement between modalities survives calibration throughout and is strongest between late speech layers and early text layers, providing strong evidence of genuine local alignment. Global similarity measures, however, do not provide reliable evidence that the two modalities share a common geometric space as a whole. Alignment is unaffected by speaker identity and prosodic variation, so the shared structure is tied to the content of the sentence itself rather than to how it is spoken. It also does not increase with model capability or with how well a text model reproduces human judgments of meaning, suggesting that this shared structure corresponds more closely to lexical and surface-level representations than to fine-grained sentence meaning. These findings indicate where speech and text representations can be connected, suggest that cross-modal transfer may benefit from exploiting neighborhood structure rather than assuming a shared global geometry, and provide new alignment methods with an existing lexical correspondence to build on, while pointing to sentence-level meaning as the structure that future speech representations could capture so that alignment extends beyond the lexical level.

Current safety evaluations of large language models (LLMs) create a dangerous illusion of universal protection by aggregating harms under generic categories such as “Identity Hate”, obscuring vulnerabilities toward specific populations. In this work, we expose the Selective Safety Trap: a systemic failure mode where models robustly defend specific populations while leaving underrepresented communities highly vulnerable to identical adversarial attacks. To systematically audit this phenomenon, we introduce MiJaBench, a bilingual (English–Portuguese) adversarial benchmark comprising 43,961 controlled jailbreaking prompts across 16 minority groups. By evaluating 14 state-of-the-art LLMs on MiJaBench, we curate 615,454 prompt-response pairs that compose MiJaBench-Align, revealing that safety alignment is not a uniform semantic capability but a demographic hierarchy, with defense rates fluctuating by up to 42% within the same model solely based on the target group. This disparity persists across architectures and languages and is amplified by scaling, indicating that current alignment methods learn group-specific safeguards rather than a generalized notion of harm. Through targeted direct preference optimization (DPO) on a 1B-parameter baseline, we achieve strong zero-shot safety generalizations to entirely unseen demographics and complex attack strategies. We release all datasets and scripts to provide the community with a concrete pathway toward equitable, transferable safety alignment.

Instruction-following is a core capability of Large Language Models (LLMs), as they must reliably comply with instructions constraints before being able to be evaluated on downstream tasks. In this scenario, instruction-following benchmarks for low-resource languages remain limited, particularly for Portuguese, where existing evaluations are mostly based on translated English datasets and fail to capture language specific linguistic phenomena. We introduce Portuguese-IFEval, a benchmark for evaluating instruction following in Portuguese built through semantic regionalization. The benchmark contains a total of 1236 instructions across 535 prompts, each prompt have up to three verifiable instructions, combining adapted instructions with Portuguese-specific orthographic, grammatical, and discourse-level constraints. We evaluate 25 proprietary, open-source, and Portuguese-tuned LLMs under a unified protocol, establishing a comprehensive benchmark for Portuguese instruction-following evaluation. Our results show that, while English and Spanish remain the strongest-performing languages, Portuguese still presents substantial challenges, which are effectively exposed by our benchmark. Our code is available at: https://github.com/gap-ufg/Portuguese-IFEval.

Speech Retrieval-Augmented Generation (Speech RAG) systems are increasingly used in real-world settings where spoken queries vary across speakers, emotions, and acoustic conditions. However, current evaluations primarily focus on clean-condition retrieval accuracy, leaving unclear whether these systems remain semantically invariant under paralinguistic variation. We introduce ParaSpeechRAG, a counterfactual robustness benchmark that isolates non-semantic acoustic factors while preserving lexical content. The benchmark spans controlled perturbations across speaker identity, emotion and prosody, additive noise, speaking rate, and codec compression, enabling systematic evaluation of semantic invariance in speech retrieval. We evaluate three retriever families: ASR-based cascades, speech-text multimodal retrievers (CLASP), and general-purpose audio-language models (GLAP). Results reveal a consistent robustness–accuracy trade-off: multimodal retrievers achieve higher performance under clean conditions but are substantially more sensitive to paralinguistic and acoustic perturbations, particularly emotion and noise. In contrast, ASR-based cascades are more robust to these variations due to their text-centric bottleneck, though with reduced expressive capacity. These findings highlight an overlooked dimension of instability in Speech RAG systems and underscore the need for robustness-aware evaluation beyond clean-condition benchmarks.

Long-term memory in Large Language Model (LLM) agents requires selective persistence: only a subset of interactions should be consolidated beyond the current context window. Existing memory systems rely on heuristic importance rules or similarity-based novelty, which remain external to the model’s internal computation. We propose a geometric principle for memory formation: episodic memory can emerge from compression failure in latent representation space. We approximate the manifold of routine LLM activations using Sparse Autoencoders (SAEs) and define representational surprise as reconstruction error relative to this learned manifold. Deviations from routine structure yield elevated residuals, providing an unsupervised, model-internal signal for memory writing. We demonstrate this principle through ReSuME, a surprise-gated memory mechanism that commits turns to memory only when normalized reconstruction error exceeds a calibrated threshold. In long-horizon multi-turn dialogue settings, representational surprise separates routine, critical, and out-of-distribution states, and achieves a superior performance–memory trade-off against established memory-agent systems (MemGPT, EM-LLM) and unsupervised anomaly detectors, while operating without labels and at substantially lower inference cost than LLM-judge controllers. Covariance-aware normalization further enables robust cross-domain calibration. These results suggest that episodic memory gating in neural agents can be grounded in intrinsic latent geometry rather than externally engineered rules.

Prompt injection and jailbreak attacks remain a critical vulnerability for deployed large language models (LLMs), allowing adversaries to bypass safety protocols and extract sensitive information. To address this, we present Proxy Barrier (ProB), a lightweight defense that interposes a proxy LLM between the user and the target model. The proxy LLM is tasked solely to repeat the user input, and any failure indicates the presence of an attempt to reveal or override system instructions, leading the malicious request to be detected and blocked before it reaches the target model. ProB therefore requires no access to model weights or prompts, and is deployable entirely at the API level. Experiments across multiple model families demonstrate that ProB achieves state-of-the-art resilience against prompt leakage and jailbreak attacks. Notably, our approach outperforms baselines and achieves up to 98.8% defense effectiveness, and shows robust protection across both open and closed-source LLMs when suitably paired with proxy models, while also keeping response quality intact.

As large language models (LLMs) become integral to social and interactive applications, the ability to model, control, and evaluate their personality traits has become a critical area of research. This survey provides a comprehensive and structured overview of the LLM-driven personality scenario. We introduce a functional taxonomy that organizes the field by how personality is modeled (from rule-based methods to model-centric and system-level LLM techniques), across which modalities it is expressed (extending beyond text to vision, speech, and immersive virtual reality), and how it is validated (covering both qualitative and quantitative evaluation paradigms). By contextualizing current advances and systematically analyzing the limitations of existing methods including subjectivity, context dependence, limited multimodal integration, and the lack of standardized evaluation protocols, we identify key research gaps. This survey serves as a guide for future inquiry, paving the way for the development LLMs with more consistent consistent, expressive, and trustworthy personality traits.

Current Portuguese Automated Fact-Checking (AFC) research often relies on datasets lacking integrated external evidence crucial for comprehensive verification. This study addresses this gap by systematically enriching Portuguese misinformation datasets. We retrieve web evidence by simulating user information-seeking behavior, guided by core claims extracted using Large Language Models (LLMs). Additionally, we apply a semi-automated validation framework to enhance dataset reliability.Our analysis reveals that inherent dataset characteristics impact data properties, evidence retrieval, and AFC model performance. While enrichment generally improves detection, its efficacy varies, influenced by challenges such as self-reinforcing online misinformation and API limitations. This work contributes enriched datasets, associating original texts with retrieved evidence and LLM-extracted claims, to foster future evidence-based fact-checking research.The code and enriched data for this study is available at https://github.com/ju-resplande/pt_afc.

The detection of audio deepfakes (ADD) has become increasingly important due to the rapid evolution of generative speech models. However, progress in this field remains uneven across languages, particularly for low-resource languages like Portuguese, which lack high-quality datasets. In this paper, we introduce BRSpeech-DF, the first publicly available ADD dataset for Portuguese, encompassing both Brazilian and European variants. The dataset contains over 458,000 utterances, including a smaller portion of real speech from 62 speakers and a large collection of synthetic samples generated using multiple zero-shot text-to-speech (TTS) models, each conditioned on the original speaker’s voice. By providing this resource, our objective is to support the development of robust, multilingual detection systems, thereby advancing equity in speech forensics and security research. BRSpeech-DF addresses a significant gap in annotated data for underrepresented languages, facilitating more inclusive and generalizable advancements in synthetic speech detection.