A Simple Method to Enhance Pre-trained Language Models with Speech Tokens for Classification

This paper presents a simple method that allows to easily enhance textual pre-trained large language models with speech information, when fine-tuned for a specific classification task. A classical issue with the fusion of many embeddings from audio with text is the large length of the audio sequence compared to the text one. Our method benefits from an existing speech tokenizer trained for Audio Speech Recognition that output long sequences of tokens from a large vocabulary, making it difficult to integrate it at low cost in a large language model. By applying a simple lasso-based feature selection on multimodal Bag-of-Words representation, we retain only the most important audio tokens for the task, and adapt the language model to them with a self-supervised language modeling objective, before fine-tuning it on the downstream task. We show this helps to improve the performances compared to an unimodal model, to a bigger SpeechLM or to integrating audio via a learned representation. We show the effectiveness of our method on two recent Argumentative Fallacy Detection and Classification tasks where the use of audio was believed counterproductive, reaching state-of-the-art results. We also provide an in-depth analysis of the method, showing that even a random audio token selection helps enhancing the unimodal model. Our code is available [online](https://github.com/salocinc/EACL26SpeechTokFallacy/).

VLM2GeoVec: Toward Universal Multimodal Embeddings for Remote Sensing

Satellite imagery differs fundamentally from natural images: its aerial viewpoint, very high resolution, diverse scale variations, and abundance of small objects demand both region-level spatial reasoning and holistic scene understanding. Current remote-sensing approaches remain fragmented between dual-encoder retrieval models, which excel at large-scale cross-modal search but cannot interleave modalities, and generative assistants, which support region-level interpretation but lack scalable retrieval capabilities. We propose $\textbf{VLM2GeoVec}$, an instruction-following, single-encoder vision-language model trained contrastively to embed interleaved inputs (images, text, bounding boxes, and geographic coordinates) in a unified vector space. Our single encoder interleaves all inputs into one joint embedding trained with a contrastive loss, eliminating multi-stage pipelines and task-specific modules. To evaluate its versatility, we introduce $\textbf{RSMEB}$, a novel benchmark covering key remote-sensing embedding applications: scene classification; cross-modal search; compositional retrieval; visual-question answering; visual grounding and region-level reasoning; and semantic geospatial retrieval. On RSMEB, it achieves $\textbf{26.6%}$ P@1 on region-caption retrieval (+25 pp vs. dual-encoder baselines), $\textbf{32.5%}$ P@1 on referring-expression retrieval (+19 pp), and $\textbf{17.8%}$ P@1 on semantic geo-localization retrieval (over $3\times$ prior best), while matching or exceeding specialized baselines on conventional tasks such as scene classification and cross-modal retrieval. VLM2GeoVec unifies scalable retrieval with region-level spatial reasoning, enabling cohesive multimodal analysis in remote sensing. We will publicly release the code, checkpoints, and data upon acceptance.

MultiHateLoc: Towards Temporal Localisation of Multimodal Hate Content in Online Videos

The rapid growth of video content on platforms such as TikTok and YouTube has intensified the spread of multimodal hate speech, where harmful cues emerge subtly and asynchronously across visual, acoustic, and textual streams. Existing research primarily focuses on video-level classification, leaving the practically crucial task of temporal localisation, identifying when hateful segments occur, largely unaddressed. This challenge is even more noticeable under weak supervision, where only video-level labels are available, and static fusion or classification-based architectures struggle to capture cross-modal and temporal dynamics. To address these challenges, we propose MultiHateLoc, the first framework designed for weakly-supervised multimodal hate localisation. MultiHateLoc incorporates (1) modality-aware temporal encoders to model heterogeneous sequential patterns, including a tailored text-based preprocessing module for feature enhancement; (2) dynamic cross-modal fusion to adaptively emphasise the most informative modality at each moment and a cross-modal contrastive alignment strategy to enhance multimodal feature consistency; (3) a modality-aware MIL objective to identify discriminative segments under video-level supervision. Despite relying solely on coarse labels, MultiHateLoc produces fine-grained, interpretable frame-level predictions. Experiments on HateMM and MultiHateClip show that our method achieves state-of-the-art performance in the localisation task.

mmWEAVER: Environment-Specific mmWave Signal Synthesis from a Photo and Activity Description

Realistic signal generation and dataset augmentation are essential for advancing mmWave radar applications such as activity recognition and pose estimation, which rely heavily on diverse, and environment-specific signal datasets. However, mmWave signals are inherently complex, sparse, and high-dimensional, making physical simulation computationally expensive. This paper presents mmWeaver, a novel framework that synthesizes realistic, environment-specific complex mmWave signals by modeling them as continuous functions using Implicit Neural Representations (INRs), achieving up to 49-fold compression. mmWeaver incorporates hypernetworks that dynamically generate INR parameters based on environmental context (extracted from RGB-D images) and human motion features (derived from text-to-pose generation via MotionGPT), enabling efficient and adaptive signal synthesis. By conditioning on these semantic and geometric priors, mmWeaver generates diverse I/Q signals at multiple resolutions, preserving phase information critical for downstream tasks such as point cloud estimation and activity classification. Extensive experiments show that mmWeaver achieves a complex SSIM of 0.88 and a PSNR of 35 dB, outperforming existing methods in signal realism while improving activity recognition accuracy by up to 7% and reducing human pose estimation error by up to 15%, all while operating 6-35 times faster than simulation-based approaches.

Super Suffixes: Bypassing Text Generation Alignment and Guard Models Simultaneously

The rapid deployment of Large Language Models (LLMs) has created an urgent need for enhanced security and privacy measures in Machine Learning (ML). LLMs are increasingly being used to process untrusted text inputs and even generate executable code, often while having access to sensitive system controls. To address these security concerns, several companies have introduced guard models, which are smaller, specialized models designed to protect text generation models from adversarial or malicious inputs. In this work, we advance the study of adversarial inputs by introducing Super Suffixes, suffixes capable of overriding multiple alignment objectives across various models with different tokenization schemes. We demonstrate their effectiveness, along with our joint optimization technique, by successfully bypassing the protection mechanisms of Llama Prompt Guard 2 on five different text generation models for malicious text and code generation. To the best of our knowledge, this is the first work to reveal that Llama Prompt Guard 2 can be compromised through joint optimization. Additionally, by analyzing the changing similarity of a model's internal state to specific concept directions during token sequence processing, we propose an effective and lightweight method to detect Super Suffix attacks. We show that the cosine similarity between the residual stream and certain concept directions serves as a distinctive fingerprint of model intent. Our proposed countermeasure, DeltaGuard, significantly improves the detection of malicious prompts generated through Super Suffixes. It increases the non-benign classification rate to nearly 100%, making DeltaGuard a valuable addition to the guard model stack and enhancing robustness against adversarial prompt attacks.

Pose-Based Sign Language Spotting via an End-to-End Encoder Architecture

Automatic Sign Language Recognition (ASLR) has emerged as a vital field for bridging the gap between deaf and hearing communities. However, the problem of sign-to-sign retrieval or detecting a specific sign within a sequence of continuous signs remains largely unexplored. We define this novel task as Sign Language Spotting. In this paper, we present a first step toward sign language retrieval by addressing the challenge of detecting the presence or absence of a query sign video within a sentence-level gloss or sign video. Unlike conventional approaches that rely on intermediate gloss recognition or text-based matching, we propose an end-to-end model that directly operates on pose keypoints extracted from sign videos. Our architecture employs an encoder-only backbone with a binary classification head to determine whether the query sign appears within the target sequence. By focusing on pose representations instead of raw RGB frames, our method significantly reduces computational cost and mitigates visual noise. We evaluate our approach on the Word Presence Prediction dataset from the WSLP 2025 shared task, achieving 61.88\% accuracy and 60.00\% F1-score. These results demonstrate the effectiveness of our pose-based framework for Sign Language Spotting, establishing a strong foundation for future research in automatic sign language retrieval and verification. Code is available at https://github.com/EbimoJohnny/Pose-Based-Sign-Language-Spotting

Refining Diffusion Models for Motion Synthesis with an Acceleration Loss to Generate Realistic IMU Data

We propose a text-to-IMU (inertial measurement unit) motion-synthesis framework to obtain realistic IMU data by fine-tuning a pretrained diffusion model with an acceleration-based second-order loss (L_acc). L_acc enforces consistency in the discrete second-order temporal differences of the generated motion, thereby aligning the diffusion prior with IMU-specific acceleration patterns. We integrate L_acc into the training objective of an existing diffusion model, finetune the model to obtain an IMU-specific motion prior, and evaluate the model with an existing text-to-IMU framework that comprises surface modelling and virtual sensor simulation. We analysed acceleration signal fidelity and differences between synthetic motion representation and actual IMU recordings. As a downstream application, we evaluated Human Activity Recognition (HAR) and compared the classification performance using data of our method with the earlier diffusion model and two additional diffusion model baselines. When we augmented the earlier diffusion model objective with L_acc and continued training, L_acc decreased by 12.7% relative to the original model. The improvements were considerably larger in high-dynamic activities (i.e., running, jumping) compared to low-dynamic activities~(i.e., sitting, standing). In a low-dimensional embedding, the synthetic IMU data produced by our refined model shifts closer to the distribution of real IMU recordings. HAR classification trained exclusively on our refined synthetic IMU data improved performance by 8.7% compared to the earlier diffusion model and by 7.6% over the best-performing comparison diffusion model. We conclude that acceleration-aware diffusion refinement provides an effective approach to align motion generation and IMU synthesis and highlights how flexible deep learning pipelines are for specialising generic text-to-motion priors to sensor-specific tasks.

ConStruct: Structural Distillation of Foundation Models for Prototype-Based Weakly Supervised Histopathology Segmentation

Weakly supervised semantic segmentation (WSSS) in histopathology relies heavily on classification backbones, yet these models often localize only the most discriminative regions and struggle to capture the full spatial extent of tissue structures. Vision-language models such as CONCH offer rich semantic alignment and morphology-aware representations, while modern segmentation backbones like SegFormer preserve fine-grained spatial cues. However, combining these complementary strengths remains challenging, especially under weak supervision and without dense annotations. We propose a prototype learning framework for WSSS in histopathological images that integrates morphology-aware representations from CONCH, multi-scale structural cues from SegFormer, and text-guided semantic alignment to produce prototypes that are simultaneously semantically discriminative and spatially coherent. To effectively leverage these heterogeneous sources, we introduce text-guided prototype initialization that incorporates pathology descriptions to generate more complete and semantically accurate pseudo-masks. A structural distillation mechanism transfers spatial knowledge from SegFormer to preserve fine-grained morphological patterns and local tissue boundaries during prototype learning. Our approach produces high-quality pseudo masks without pixel-level annotations, improves localization completeness, and enhances semantic consistency across tissue types. Experiments on BCSS-WSSS datasets demonstrate that our prototype learning framework outperforms existing WSSS methods while remaining computationally efficient through frozen foundation model backbones and lightweight trainable adapters.

Script Gap: Evaluating LLM Triage on Indian Languages in Native vs Roman Scripts in a Real World Setting

Large Language Models (LLMs) are increasingly deployed in high-stakes clinical applications in India. In many such settings, speakers of Indian languages frequently communicate using romanized text rather than native scripts, yet existing research rarely evaluates this orthographic variation using real-world data. We investigate how romanization impacts the reliability of LLMs in a critical domain: maternal and newborn healthcare triage. We benchmark leading LLMs on a real-world dataset of user-generated queries spanning five Indian languages and Nepali. Our results reveal consistent degradation in performance for romanized messages, with F1 scores trailing those of native scripts by 5-12 points. At our partner maternal health organization in India, this gap could cause nearly 2 million excess errors in triage. Crucially, this performance gap by scripts is not due to a failure in clinical reasoning. We demonstrate that LLMs often correctly infer the semantic intent of romanized queries. Nevertheless, their final classification outputs remain brittle in the presence of orthographic noise in romanized inputs. Our findings highlight a critical safety blind spot in LLM-based health systems: models that appear to understand romanized input may still fail to act on it reliably.

FRIEDA: Benchmarking Multi-Step Cartographic Reasoning in Vision-Language Models

Cartographic reasoning is the skill of interpreting geographic relationships by aligning legends, map scales, compass directions, map texts, and geometries across one or more map images. Although essential as a concrete cognitive capability and for critical tasks such as disaster response and urban planning, it remains largely unevaluated. Building on progress in chart and infographic understanding, recent large vision language model studies on map visual question-answering often treat maps as a special case of charts. In contrast, map VQA demands comprehension of layered symbology (e.g., symbols, geometries, and text labels) as well as spatial relations tied to orientation and distance that often span multiple maps and are not captured by chart-style evaluations. To address this gap, we introduce FRIEDA, a benchmark for testing complex open-ended cartographic reasoning in LVLMs. FRIEDA sources real map images from documents and reports in various domains and geographical areas. Following classifications in Geographic Information System (GIS) literature, FRIEDA targets all three categories of spatial relations: topological (border, equal, intersect, within), metric (distance), and directional (orientation). All questions require multi-step inference, and many require cross-map grounding and reasoning. We evaluate eleven state-of-the-art LVLMs under two settings: (1) the direct setting, where we provide the maps relevant to the question, and (2) the contextual setting, where the model may have to identify the maps relevant to the question before reasoning. Even the strongest models, Gemini-2.5-Pro and GPT-5-Think, achieve only 38.20% and 37.20% accuracy, respectively, far below human performance of 84.87%. These results reveal a persistent gap in multi-step cartographic reasoning, positioning FRIEDA as a rigorous benchmark to drive progress on spatial intelligence in LVLMs.