arXiv subjects

Vera Demberg

Publications and source records attributed to Vera Demberg.

At least 19 recordsLinked to original sources

arXiv2026

PragMatch: Separating Pragmatic Incongruity from Cross-Modal Mismatch in Large Vision-Language Models

Large Vision-Language Models (LVLMs) have demonstrated strong performance on multimodal benchmarks, yet it remains unclear whether they genuinely reason about relationships between images and text or rely on superficial correlations, known as shortcut learning. This question is particularly important for multimodal sarcasm detection, where successful prediction depends on recognizing pragmatic incongruity rather than treating sarcasm as simple image-text mismatch. We introduce PragMatch, a controlled benchmark of 3,000 image-text pairs derived from MMSD2.0, including original sarcastic examples and constructed literal and hard-negative pairs. We identify influential shortcut cues through systematic masking and evaluate their impact through targeted injection experiments. Our results show that LVLM predictions are sensitive to lexical, OCR-derived and stylistic cues, with injected surface signals causing substantial changes in model predictions despite unchanged underlying image-text relationships. Our findings reveal limitations in current LVLMs while PragMatch provides a systematic testbed for evaluating multimodal pragmatic reasoning beyond surface-level image-text alignment.

cs.CL↗

arXiv2026

Beyond Routing Weights: Faithful Response-Level Interpretation of Mixture-of-Experts Reward Models via Contribution Contrast

Reward models are central to learning from human preferences, yet identifying what drives their predictions remains challenging. Recent sparse Mixture-of-Experts (MoE) reward models seek to improve interpretability by routing prompts to specialized experts and characterizing experts through examples with high routing weights. However, routing weights only reveal which prompts an expert $\textit{receives}$, not how it $\textit{judges}$ responses, providing only a partial account of expert behavior. We therefore propose $\textbf{Co}$ntribution-$\textbf{Co}$ntrast ($\textbf{CoCo}$) response-level interpretation, which faithfully characterizes experts' roles using chosen-rejected response pairs with the largest contribution contrasts, jointly capturing routing and preference behavior. Across automatic and human evaluations, CoCo yields more coherent, faithful, and specialized interpretations than router-based, score-based, and sparse autoencoder-based alternatives while maintaining competitive reward modeling accuracy. To the best of our knowledge, this is the first systematic study of interpretation methods for MoE reward models.

cs.AI↗

arXiv2026

A retrieval conditioned rebinding circuit for dynamic entity tracking in large language models

To interpret context correctly and retrieve relevant information, large language models must bind entities to their attributes and update these bindings as state changes. We analyze how LLMs implement this binding process in a dynamic state tracking. Using causal interventions, we identify a retrieval conditioned rebinding mechanism, a compact attention head circuit that encodes swap relevant binding information and reinstates it at readout. Across Gemma and Llama models, this circuit supports rebinding behavior, but the representational signature of the mechanism differs across model families. In Gemma models, the binding signature is clearly expressed in the query/key subspaces of the relevant attention heads, whereas in Llama models, the binding information is carried primarily in key vectors. Overall, our results reveal an interpretable mechanism for context dependent state tracking in LLMs.

cs.CL↗

arXiv2026

Sparse Mixture-of-Experts Reward Models Learn Interpretable and Specialized Experts for Personalized Preference Modeling

Preference modeling plays a central role in reinforcement learning from human feedback (RLHF), enabling large language models (LLMs) to align with human values. However, most existing approaches assume a universal reward function, neglecting the diversity and heterogeneity of human preferences. To address this limitation without additional annotation costs, recent work has proposed learning multiple preference components from binary data and combining them to model individual preferences. Nevertheless, these components often fail to capture coherent and disentangled patterns, limiting their interpretability and effectiveness for personalization. In this work, we propose a sparse Mixture-of-Experts (MoE) reward model that encourages sparse routing and expert diversity during training on binary preference data. Across controlled and real-world experiments, sparse MoE learns interpretable routing patterns and specialized experts. It also improves test-time personalization, and post-adaptation shifts in expert weights provide a qualitative lens for analyzing how the model adapts to personalized preferences.

cs.LG↗

arXiv2026

MuPHI: Learning Implicit Multimodal Harm Reasoning via Semantically Grounded Reward Optimization

Understanding how harm emerges from interaction between otherwise benign image-text pairs requires intent-aware cross-modal reasoning beyond surface-level features. Existing vision-language models (VLMs) excel at literal reasoning over perceptual cues but often fail to derive harmful semantics that rely on implicit, context-dependent reasoning. To evaluate VLMs on compositional harm detection and reasoning, we introduce Multimodal Pragmatic Harm Interpretation (MuPHI), a dataset containing image-text pairs where harm is encoded in subtle multimodal cues. MuPHI spans diverse harm categories and includes annotated harm rationales for assessing VLM reasoning chains. To improve both detection and reasoning in VLMs, we propose MuPHIRM, a reasoning-augmented training framework which learns joint semantics by optimizing multi-perspective rewards. MuPHIRM improves both harm detection and reasoning quality of VLMs while demonstrating superior out-of-distribution robustness compared to both trained and inference-time baselines. Our findings suggest that reasoning-oriented reward optimization offers a promising direction towards building multimodal systems that generalize beyond benchmark-specific shortcuts.

cs.AI↗

arXiv2026

Semantic Motion Anchors: Bridging Motion and Meaning in Co-Speech Gestures

Learning a shared representation between spoken text and gesture is central to co-speech gesture retrieval, synthesis, and understanding, but remains challenging for semantically meaningful gestures whose communicative intent is not captured by motion alone. Direct contrastive alignment between transcripts and continuous motion embeddings often overemphasizes low-level kinematics and misses the symbolic content of semantic gestures. We propose semantic motion anchors, natural-language abstractions of gesture motion capturing physical form and communicative intent. Our method discretizes 3D gestures into body-hand motion primitives, verbalizes them into structured descriptions, and grounds them in the transcript to provide auxiliary contrastive supervision. On BEAT2, our method improves text-to-gesture R@1 by 8.2% over a direct text-motion baseline and outperforms prior retrieval approaches on text to gesture and gesture to text retrieval directions. Beyond aggregate retrieval metrics, semantic motion anchor supervision helps retrieve gestures that are semantically meaningful for the spoken query, rather than defaulting to generic motion patterns. A downstream retrieval-augmented gesture generation study showed that users significantly preferred gestures retrieved by our approach over a retrieval-augmented generation baseline, demonstrating that semantically grounded retrieval translates to gestures that better convey communicative intent in downstream generation.

cs.CL↗

arXiv2026

Is Cross-Lingual Transfer in Bilingual Models Human-Like? A Study with Overlapping Word Forms in Dutch and English

Bilingual speakers show cross-lingual activation during reading, especially for words with shared surface form. Cognates (friends) typically lead to facilitation, whereas interlingual homographs (false friends) cause interference or no effect. We examine whether cross-lingual activation in bilingual language models mirrors these patterns. We train Dutch-English causal Transformers under four vocabulary-sharing conditions that manipulate whether (false) friends receive shared or language-specific embeddings. Using psycholinguistic stimuli from bilingual reading studies, we evaluate the models through surprisal and embedding similarity analyses. The models largely maintain language separation, and cross-lingual effects arise primarily when embeddings are shared. In these cases, both friends and false friends show facilitation relative to controls. Regression analyses reveal that these effects are mainly driven by frequency rather than consistency in form-meaning mapping. Only when just friends share embeddings are the qualitative patterns of bilinguals reproduced. Overall, bilingual language models capture some cross-linguistic activation effects. However, their alignment with human processing seems to critically depend on how lexical overlap is encoded, possibly limiting their explanatory adequacy as models of bilingual reading.

cs.CL↗

arXiv2026

MIBURI: Towards Expressive Interactive Gesture Synthesis

Embodied Conversational Agents (ECAs) aim to emulate human face-to-face interaction through speech, gestures, and facial expressions. Current large language model (LLM)-based conversational agents lack embodiment and the expressive gestures essential for natural interaction. Existing solutions for ECAs often produce rigid, low-diversity motions, that are unsuitable for human-like interaction. Alternatively, generative methods for co-speech gesture synthesis yield natural body gestures but depend on future speech context and require long run-times. To bridge this gap, we present MIBURI, the first online, causal framework for generating expressive full-body gestures and facial expressions synchronized with real-time spoken dialogue. We employ body-part aware gesture codecs that encode hierarchical motion details into multi-level discrete tokens. These tokens are then autoregressively generated by a two-dimensional causal framework conditioned on LLM-based speech-text embeddings, modeling both temporal dynamics and part-level motion hierarchy in real time. Further, we introduce auxiliary objectives to encourage expressive and diverse gestures while preventing convergence to static poses. Comparative evaluations demonstrate that our causal and real-time approach produces natural and contextually aligned gestures against recent baselines. We urge the reader to explore demo videos on https://vcai.mpi-inf.mpg.de/projects/MIBURI/.

cs.CV↗

arXiv2026

Human Label Variation in Implicit Discourse Relation Recognition

There is growing recognition that many NLP tasks lack a single ground truth, as human judgments reflect diverse perspectives. To capture this variation, models have been developed to predict full annotation distributions rather than majority labels, while perspectivist models aim to reproduce the interpretations of individual annotators. In this work, we compare these approaches on Implicit Discourse Relation Recognition (IDRR), a highly ambiguous task where disagreement often arises from cognitive complexity rather than ideological bias. Our experiments show that existing annotator-specific models perform poorly in IDRR unless ambiguity is reduced, whereas models trained on label distributions yield more stable predictions. Further analysis indicates that frequent cognitively demanding cases drive inconsistency in human interpretation, posing challenges for perspectivist modeling in IDRR.

cs.CL↗

arXiv2026

Generation-Step-Aware Framework for Cross-Modal Representation and Control in Multilingual Speech-Text Models

Multilingual speech-text models rely on cross-modal language alignment to transfer knowledge between speech and text, but it remains unclear whether this reflects shared computation for the same language or modality-specific processing. We introduce a generation-step-aware framework for evaluating cross-modal computation that (i) identifies language-selective neurons for each modality at different decoding steps, (ii) decomposes them into language-representation and language-control roles, and (iii) enables cross-modal comparison via overlap measures and causal intervention. Applying our framework to SeamlessM4T, and additionally evaluating its generality on the decoder-only model Qwen2-Audio, we find that cross-modal language-representation alignment is highly model dependent. SeamlessM4T exhibits pronounced generation-step-dependent specialization, where only 5-7% of language-representation neurons are shared across modalities and overlap shifts from same-language to typologically related languages during autoregressive generation, whereas Qwen2-Audio maintains substantially larger cross-modal sharing (45-47%) together with stable, strongly language-specific alignment across decoding steps. In contrast, language-control neurons identified at later decoding steps exhibit progressively stronger cross-modal transfer from speech to text in SeamlessM4T. Together, these results demonstrate that generation-step-aware analysis can reveal both shared and model-specific patterns of cross-modal computation.

cs.CL↗

arXiv2026

System-Mediated Attention Imbalances Make Vision-Language Models Say Yes

Vision-language model (VLM) hallucination is commonly linked to imbalanced allocation of attention across input modalities: system, image and text. However, existing mitigation strategies tend towards an image-centric interpretation of these imbalances, often prioritising increased image attention while giving less consideration to the roles of the other modalities. In this study, we evaluate a more holistic, system-mediated account, which attributes these imbalances to functionally redundant system weights that reduce attention to image and textual inputs. We show that this framework offers a useful empirical perspective on the yes-bias, a common form of hallucination in which VLMs indiscriminately respond `yes'. Causally redistributing attention from the system modality to image and textual inputs substantially suppresses this bias, often outperforming existing approaches. We further present evidence suggesting that system-mediated attention imbalances contribute to the yes-bias by encouraging a default reliance on coarse input representations, which are effective for some tasks but ill-suited to others. Taken together, these findings firmly establish system attention as a key factor in VLM hallucination and highlight its potential as a lever for mitigation.

cs.CL↗

arXiv2025

How AI Responses Shape User Beliefs: The Effects of Information Detail and Confidence on Belief Strength and Stance

The growing use of AI-generated responses in everyday tools raises concern about how subtle features such as supporting detail or tone of confidence may shape people's beliefs. To understand this, we conducted a pre-registered online experiment (N = 304) investigating how the detail and confidence of AI-generated responses influence belief change. We introduce an analysis framework with two targeted measures: belief switch and belief shift. These distinguish between users changing their initial stance after AI input and the extent to which they adjust their conviction toward or away from the AI's stance, thereby quantifying not only categorical changes but also more subtle, continuous adjustments in belief strength that indicate a reinforcement or weakening of existing beliefs. Using this framework, we find that detailed responses with medium confidence are associated with the largest overall belief changes. Highly confident messages tend to elicit belief shifts but induce fewer stance reversals. Our results also show that task type (fact-checking versus opinion evaluation), prior conviction, and perceived stance agreement further modulate the extent and direction of belief change. These findings illustrate how different properties of AI responses interact with user beliefs in subtle but potentially consequential ways and raise practical as well as ethical considerations for the design of LLM-powered systems.

cs.HC↗

arXiv2025

Can LLMs subtract numbers?

We present a systematic study of subtraction in large language models (LLMs). While prior benchmarks emphasize addition and multiplication, subtraction has received comparatively little attention despite being structurally distinct as a non-commutative operation. We evaluate eight pretrained LLMs spanning four families on addition and subtraction problems. Our experiments reveal that subtraction accuracy lags behind addition by a wide margin. We find that the errors for ($a-b$) are concentrated in cases where ($a<b$). In such cases, LLMs frequently produce the correct magnitude but omit the negative sign. Probing analyses show that LLMs internally encode whether results should be negative, yet this information is often not reflected in generated outputs. We further test well-known techniques such as few-shot learning and instruction-tuning to see if they can improve the LLMs' performance. Our results suggest that while few-shot prompting yields modest gains, the instruction-tuned models achieve near-perfect accuracies in generating the negative sign. Together, these findings provide a clearer characterization of the limitations and recoverability of LLMs' arithmetic capabilities in subtraction.

cs.LG↗

arXiv2025

MUStReason: A Benchmark for Diagnosing Pragmatic Reasoning in Video-LMs for Multimodal Sarcasm Detection

Sarcasm is a specific type of irony which involves discerning what is said from what is meant. Detecting sarcasm depends not only on the literal content of an utterance but also on non-verbal cues such as speaker's tonality, facial expressions and conversational context. However, current multimodal models struggle with complex tasks like sarcasm detection, which require identifying relevant cues across modalities and pragmatically reasoning over them to infer the speaker's intention. To explore these limitations in VideoLMs, we introduce MUStReason, a diagnostic benchmark enriched with annotations of modality-specific relevant cues and underlying reasoning steps to identify sarcastic intent. In addition to benchmarking sarcasm classification performance in VideoLMs, using MUStReason we quantitatively and qualitatively evaluate the generated reasoning by disentangling the problem into perception and reasoning, we propose PragCoT, a framework that steers VideoLMs to focus on implied intentions over literal meaning, a property core to detecting sarcasm.

cs.LG↗

arXiv2025

Planning Ahead with RSA: Efficient Signalling in Dynamic Environments by Projecting User Awareness across Future Timesteps

Adaptive agent design offers a way to improve human-AI collaboration on time-sensitive tasks in rapidly changing environments. In such cases, to ensure the human maintains an accurate understanding of critical task elements, an assistive agent must not only identify the highest priority information but also estimate how and when this information can be communicated most effectively, given that human attention represents a zero-sum cognitive resource where focus on one message diminishes awareness of other or upcoming information. We introduce a theoretical framework for adaptive signalling which meets these challenges by using principles of rational communication, formalised as Bayesian reference resolution using the Rational Speech Act (RSA) modelling framework, to plan a sequence of messages which optimise timely alignment between user belief and a dynamic environment. The agent adapts message specificity and timing to the particulars of a user and scenario based on projections of how prior-guided interpretation of messages will influence attention to the interface and subsequent belief update, across several timesteps out to a fixed horizon. In a comparison to baseline methods, we show that this effectiveness depends crucially on combining multi-step planning with a realistic model of user awareness. As the first application of RSA for communication in a dynamic environment, and for human-AI interaction in general, we establish theoretical foundations for pragmatic communication in human-agent teams, highlighting how insights from cognitive science can be capitalised to inform the design of assistive agents.

cs.AI↗

arXiv2025

Modeling Turn-Taking with Semantically Informed Gestures

In conversation, humans use multimodal cues, such as speech, gestures, and gaze, to manage turn-taking. While linguistic and acoustic features are informative, gestures provide complementary cues for modeling these transitions. To study this, we introduce DnD Gesture++, an extension of the multi-party DnD Gesture corpus enriched with 2,663 semantic gesture annotations spanning iconic, metaphoric, deictic, and discourse types. Using this dataset, we model turn-taking prediction through a Mixture-of-Experts framework integrating text, audio, and gestures. Experiments show that incorporating semantically guided gestures yields consistent performance gains over baselines, demonstrating their complementary role in multimodal turn-taking.

cs.CL↗

arXiv2025

Bridging Fairness and Explainability: Can Input-Based Explanations Promote Fairness in Hate Speech Detection?

Natural language processing (NLP) models often replicate or amplify social bias from training data, raising concerns about fairness. At the same time, their black-box nature makes it difficult for users to recognize biased predictions and for developers to effectively mitigate them. While some studies suggest that input-based explanations can help detect and mitigate bias, others question their reliability in ensuring fairness. Existing research on explainability in fair NLP has been predominantly qualitative, with limited large-scale quantitative analysis. In this work, we conduct the first systematic study of the relationship between explainability and fairness in hate speech detection, focusing on both encoder- and decoder-only models. We examine three key dimensions: (1) identifying biased predictions, (2) selecting fair models, and (3) mitigating bias during model training. Our findings show that input-based explanations can effectively detect biased predictions and serve as useful supervision for reducing bias during training, but they are unreliable for selecting fair models among candidates.Our code is available at https://github.com/Ewanwong/fairness_x_explainability.

cs.CL↗

arXiv2025

Tug-of-war between idioms' figurative and literal interpretations in LLMs

Idioms present a unique challenge for language models due to their non-compositional figurative interpretations, which often strongly diverge from the idiom's literal interpretation. In this paper, we employ causal tracing to systematically analyze how pretrained causal transformers deal with this ambiguity. We localize three mechanisms: (i) Early sublayers and specific attention heads retrieve an idiom's figurative interpretation, while suppressing its literal interpretation. (ii) When disambiguating context precedes the idiom, the model leverages it from the earliest layer and later layers refine the interpretation if the context conflicts with the retrieved interpretation. (iii) Then, selective, competing pathways carry both interpretations: an intermediate pathway prioritizes the figurative interpretation and a parallel direct route favors the literal interpretation, ensuring that both readings remain available. Our findings provide mechanistic evidence for idiom comprehension in autoregressive transformers.

cs.CL↗