Multimodal large language models (MLLMs) must often arbitrate between visual evidence and world-knowledge priors when an image conflicts with commonsense expectations. Existing work has demonstrated that activation steering vectors can bias model outputs toward either source of information, but it remains unclear whether this arbitration mechanism generalizes across languages. To investigate this, we construct a multilingual extension of the Visual CounterFact color split and derive activation steering directions that contrast image-grounded answers with prior-based answers. Across 12 languages, we find that English-derived and shared multilingual steering directions consistently shift model outputs on non-English prompts, suggesting that visual-prior arbitration has a substantial cross-lingual component. At the same time, we find that performance gains on common knowledge questions can come at the cost of image-grounded accuracy, a trade-off that is particularly pronounced for lower-resource languages. Code and data are available at https://github.com/elffiona/cross-lingual-visual-prior-steering.

Multiparty tutoring policies must regulate turn-taking and social attention, yet collecting the interaction data needed for reinforcement learning is expensive in educational settings. We introduce an LLM-based user simulator for triadic robot–student dialogue that represents gaze and gesture as inline stage directions. Combined with phase-tagged prompting (which delineates conversational structure) and per-turn persona reinforcement, this design allows the LLMs to generate interactions that preserve key discourse structure and controllable participation dynamics. We validate six simulator variants against a real student–robot corpus (12 student pairs) using metrics for semantic similarity, cue obedience, and persona consistency. While the simulator reaches strong semantic agreement with real transcripts, it also reveals a measurable sim-to-real style gap (in verbosity and grammatical complexity). The simulator enables the creation of an RL environment where the tutor selects factorized gaze and reference actions, learning to interact with simulated students. We compare online training (APPO) to offline learning (DQN). Off-policy evaluation on real logs shows that offline policies trained purely from synthetic data exceed online-trained policies under a doubly-robust estimator, and a verbal-only ablation yields unstable OPE due to heavy-tailed importance weights. These results support a scalable workflow for prototyping socially grounded multiparty tutoring policies while relaying where real-world coverage and OPE reliability remain bottlenecks.

When captioning an image, people describe objects in diverse ways, such as by using different terms and/or including details that are perceptually noteworthy to them. Descriptions can be especially unique across languages and cultures. Modern vision-language models (VLMs) gain understanding of images with text in different languages often through training on machine translations of English captions. However, this process relies on input content written from the perception of English speakers, leading to a perceptual bias. In this work, we outline a framework to address this bias. We specifically use a small amount of native speaker data, nearest-neighbor example guidance, and multimodal LLM reasoning to augment captions to better reflect descriptions in a target language. When adding the resulting rewrites to multilingual CLIP finetuning, we improve on German and Japanese text-image retrieval case studies (up to +3.5 mean recall, +4.4 on native vs. translation errors). We also propose a mechanism to build understanding of object description variation across languages, and offer insights into cross-dataset and cross-language generalization.

Text-to-image models are appealing for customizing visual advertisements and targeting specific populations. We investigate this potential by examining the demographic bias within ads for different ad topics, and the disparate level of persuasiveness (judged by models) of ads that are identical except for gender/race of the people portrayed. We also experiment with a technique to target ads for specific countries.

We examine how multimodal large language models (MLLMs) perform logical inference grounded in visual information. We first construct a dataset of food web/chain images, along with questions that follow seven structured templates with progressively more complex reasoning involved. We show that complex reasoning about entities in the images remains challenging (even with elaborate prompts) and that visual information is underutilized.

Evaluating creativity is challenging, even for humans, not only because of its subjectivity but also because it involves complex cognitive processes. Inspired by work in marketing, we attempt to break down visual advertisement creativity into atypicality and originality. With fine-grained human annotations on these dimensions, we propose a suite of tasks specifically for such a subjective problem. We also evaluate the alignment between state-of-the-art (SoTA) vision language models (VLMs) and humans on our proposed benchmark, demonstrating both the promises and challenges of using VLMs for automatic creativity assessment.

We analyze whether object detectors trained on vision-language data learn effective visual representations for synonyms. Since many current vision-language models accept user-provided textual input, we highlight the need for such models to learn feature representations that are robust to changes in how such input is provided. Specifically, we analyze changes in synonyms used to refer to objects. Here, we study object detectors trained on vision-language data and investigate how to make their performance less dependent on whether synonyms are used to refer to an object. We propose two approaches to achieve this goal: data augmentation by back-translation and class embedding enrichment. We show the promise of such approaches, reporting improved performance on synonyms from mAP@0.5=33.87% to 37.93%.

There is a scarcity of multilingual vision-language models that properly account for the perceptual differences that are reflected in image captions across languages and cultures. In this work, through a multimodal, multilingual retrieval case study, we quantify the existing lack of model flexibility. We empirically show performance gaps between training on captions that come from native German perception and captions that have been either machine-translated or human-translated from English into German. To address these gaps, we further propose and evaluate caption augmentation strategies. While we achieve mean recall improvements (+1.3), gaps still remain, indicating an open area of future work for the community.

The use of large-scale vision-language datasets is limited for object detection due to the negative impact of label noise on localization. Prior methods have shown how such large-scale datasets can be used for pretraining, which can provide initial signal for localization, but is insufficient without clean bounding-box data for at least some categories. We propose a technique to “vet” labels extracted from noisy captions, and use them for weakly-supervised object detection (WSOD), without any bounding boxes. We analyze and annotate the types of label noise in captions in our Caption Label Noise dataset, and train a classifier that predicts if an extracted label is actually present in the image or not. Our classifier generalizes across dataset boundaries and across categories. We compare the classifier to nine baselines on five datasets, and demonstrate that it can improve WSOD without label vetting by 30% (31.2 to 40.5 mAP when evaluated on PASCAL VOC). See dataset at: https://github.com/arushirai1/CLaNDataset.

This work explores the feasibility of eliciting knowledge from language models (LMs) to decode symbolism, recognizing something (e.g.,roses) as a stand-in for another (e.g., love). We present our evaluative framework, Symbolism Analysis (SymbA), which compares LMs (e.g., RoBERTa, GPT-J) on different types of symbolism and analyze the outcomes along multiple metrics. Our findings suggest that conventional symbols are more reliably elicited from LMs while situated symbols are more challenging. Results also reveal the negative impact of the bias in pre-trained corpora. We further demonstrate that a simple re-ranking strategy can mitigate the bias and significantly improve model performances to be on par with human performances in some cases.

Speakers build rapport in the process of aligning conversational behaviors with each other. Rapport engendered with a teachable agent while instructing domain material has been shown to promote learning. Past work on lexical alignment in the field of education suffers from limitations in both the measures used to quantify alignment and the types of interactions in which alignment with agents has been studied. In this paper, we apply alignment measures based on a data-driven notion of shared expressions (possibly composed of multiple words) and compare alignment in one-on-one human-robot (H-R) interactions with the H-R portions of collaborative human-human-robot (H-H-R) interactions. We find that students in the H-R setting align with a teachable robot more than in the H-H-R setting and that the relationship between lexical alignment and rapport is more complex than what is predicted by previous theoretical and empirical work.