Contrastive vision-language models like CLIP are used for a large variety of applications, such as zero-shot classification or as vision encoder for multi-modal models. Despite their popularity, their representations show major limitations. For instance, CLIP models learn bag-of-words representations and, as a consequence, fail to distinguish whether an image is of “a yellow submarine and a blue bus” or “a blue submarine and a yellow bus”. Previous attempts to fix this issue added hard negatives during training or modified the architecture, but failed to resolve the problem in its entirety. We suspect that the missing insights to solve the binding problem for CLIP are hidden in arguably the most important part of learning algorithms: the data. In this work, we fill this gap by rigorously identifying the influence of data properties on CLIP's ability to learn binding using a synthetic dataset. We find that common properties of natural data such as low attribute density, incomplete captions, and the saliency bias, a tendency of human captioners to describe the object that is “most salient” to them, have a detrimental effect on binding performance. In contrast to common belief, we find that neither scaling the batch size, i.e., implicitly adding more hard negatives, nor explicitly creating hard negatives enables CLIP to learn reliable binding. Only when the data expresses our identified data properties does CLIP learn almost perfect binding.

Generating a coherent sequence of images that tells a visual story, using text-to-image diffusion models, often faces the critical challenge of maintaining subject consistency across all story scenes. Existing approaches, which typically rely on fine-tuning or retraining models, are computationally expensive, time-consuming, and often interfere with the model’s pre-existing capabilities. In this paper, we follow a training-free approach and propose an efficient consistent-subject-generation method. This approach works seamlessly with pre-trained diffusion models by introducing masked cross-image attention sharing to dynamically align subject features across a batch of images, and Regional Feature Harmonization to refine visually similar details for improved subject consistency. Experimental results demonstrate that our approach successfully generates visually consistent subjects across a variety of scenarios while maintaining the creative abilities of the diffusion model.

Recent 3D CT vision-language models align volumes with reports via contrastive pretraining, but typically rely on limited public data and provide only coarse global supervision. We train a 3D CT vision-language model on 98k report-volume pairs (50k patients) collected at a single hospital, combined with public datasets, using SigLIP-style contrastive pretraining together with prompt-based disease supervision in the shared vision-text embedding space. On CT-RATE, our model achieves state-of-the-art text-to-image retrieval (R@10 31.5 vs. 22.2) and competitive disease classification (AUC 83.8 vs. 83.8), with consistent results on Rad-ChestCT (AUC 77.0 vs. 77.3). We further observe that radiologists routinely reference specific images within their reports (e.g., “series X, image Y”), linking textual descriptions to precise axial locations. We automatically mine 262k such snippet-slice pairs and introduce the task of intra-scan snippet localization – predicting the axial depth referred to by a text snippet – reducing mean absolute error to 36.3 mm at 12 mm feature resolution, compared with 67.0 mm for the best baseline. Adding this localization objective leaves retrieval and classification broadly unchanged within confidence bounds, yielding a single unified model for retrieval, classification, and intra-scan grounding.

Concept-based models like Concept Bottleneck Models (CBMs) have garnered significant interest for improving model interpretability by first predicting human-understandable concepts before mapping them to the output classes. Early approaches required costly concept annotations. To alleviate this, recent methods utilized large language models to automatically generate class-specific concept descriptions and learned mappings from a pretrained black-box model’s raw features to these concepts using vision-language models. However, these approaches assume prior knowledge of which concepts the black-box model has learned. In this work, we discover the concepts encoded by the model through unsupervised concept discovery techniques instead. We further leverage a simple input-dependent concept selection mechanism that dynamically retains a sparse set of relevant concepts of each input, enhancing both sparsity and interpretability. Our approach not only improves downstream performance, but also needs significantly fewer concepts for accurate classification. Lastly, we show how large vision-language models can guide the editing of our models' weights to correct model errors.

Recessive dystrophic epidermolysis bullosa (RDEB) is a rare genetic skin disorder for which clinicians greatly benefit from finding similar cases using images and clinical text. However, off-the-shelf foundation models do not reliably capture clinically meaningful features for this heterogeneous, long-tail disease, and structured measurement of agreement with experts is challenging. To address these gaps, we propose evaluating embedding spaces with expert ordinal comparisons (triplet judgments), which are fast to collect and encode implicit clinical similarity knowledge. We further introduce TriDerm, a multimodal framework that learns interpretable wound representations from small cohorts by integrating wound imagery, boundary masks, and expert reports. On the vision side, TriDerm adapts visual foundation models to RDEB using wound-level attention pooling and non-contrastive representation learning. For text, we prompt large language models with comparison queries and recover medically meaningful representations via soft ordinal embeddings (SOE). We show that visual and textual modalities capture complementary aspects of wound phenotype, and that fusing both modalities yields 73.5

Large Language Models (LLMs) exhibit In-Context Learning (ICL), which enables the model to perform new tasks conditioning only on the examples provided in the context without updating the model’s weights. While ICL offers fast adaptation across natural language tasks and domains, its emergence is less straightforward for modalities beyond text. In this work, we systematically uncover properties present in LLMs that support the emergence of ICL for autoregressive models and various modalities by promoting the learning of the mechanisms needed for ICL. We identify exact token repetitions in the training data sequences as an important factor for ICL. Such repetitions further improve stability and reduce transiency in ICL performance. We analyse in detail the training dynamics of such data sequences and explain how token repetitions enhance the ICL learning mechanisms. Moreover, we emphasise the importance of the training task difficulty for the emergence of ICL. Finally, by applying our novel insights on ICL emergence, we unlock ICL capabilities across various visual datasets used for few-shot classification, and confirm the generalisability of our insights to much harder real-world examples of large-scale object classification, and a more challenging EEG classification task. Code is available at https://github.com/jelenab98/unlocking_icl

Large Language Models (LLMs) exhibit In-Context Learning (ICL), which enables the model to perform new tasks conditioning only on the examples provided in the context without updating the model’s weights. While ICL offers fast adaptation across natural language tasks and domains, its emergence is less straightforward for modalities beyond text. In this work, we systematically uncover properties present in LLMs that support the emergence of ICL for autoregressive models and various modalities by promoting the learning of the needed mechanisms for ICL. We identify exact token repetitions in the training data sequences as an important factor for ICL. Such repetitions further improve stability and reduce transiency in ICL performance. Moreover, we emphasise the significance of training task difficulty for the emergence of ICL. Finally, by applying our novel insights on ICL emergence, we unlock ICL capabilities for various visual datasets and a more challenging EEG classification task. Code is available at https://github.com/jelenab98/unlocking_icl .

Constrained Reinforcement Learning (RL) has emerged as a significant research area within RL, where integrating constraints with rewards is crucial for enhancing safety and performance across diverse control tasks. In the context of heating systems in the buildings, optimizing the energy efficiency while maintaining the residents' thermal comfort can be intuitively formulated as a constrained optimization problem. However, to solve it with RL may require large amount of data. Therefore, an accurate and versatile simulator is favored. In this paper, we propose a novel building simulator I4B which provides interfaces for different usages and apply a model-free constrained RL algorithm named constrained Soft Actor-Critic with Linear Smoothed Log Barrier function (CSAC-LB) to the heating optimization problem. Benchmarking against baseline algorithms demonstrates CSAC-LB's efficiency in data exploration, constraint satisfaction and performance.

Language models can transfer hidden biases during distillation. For example, a teacher that "likes owls" can make its student "like owls" too, even when the training data consists only of lists of numbers. This surprising phenomenon is called subliminal learning. Subliminal learning can be expected under soft distillation, where the student is trained on the teacher's full next-token distribution. But the fact that this also occurs under hard distillation-where the student only sees sampled tokens-raises a deeper question: when and how does subliminal learning actually occur? We answer this question through controlled experiments and mechanistic analysis. Our results show that subliminal learning does not need (global) token entanglement or logit leakage. Instead, it comes down to a small set of divergence tokens-rare cases where teachers with different biases would predict different tokens. Masking out these tokens mostly removes the hidden bias transfer. Mechanistically, divergence tokens reveal that early layers are critical. Surprisingly, finetuning even a single such early layer is sufficient for subliminal learning. Finally, we find that subliminal learning is fragile. Even small changes, like paraphrasing prompts, are usually sufficient to suppress it.

Reconstructing unbounded outdoor scenes from sparse outward-facing views poses significant challenges due to minimal view overlap. Previous methods often lack cross-scene understanding and their primitive-centric formulations overload local features to compensate for missing global context, resulting in blurriness in unseen parts of the scene. We propose sshELF, a fast, single-shot pipeline for sparse-view 3D scene reconstruction via hierarchal extrapolation of latent features. Our key insights is that disentangling information extrapolation from primitive decoding allows efficient transfer of structural patterns across training scenes. Our method: (1) learns cross-scene priors to generate intermediate virtual views to extrapolate to unobserved regions, (2) offers a two-stage network design separating virtual view generation from 3D primitive decoding for efficient training and modular model design, and (3) integrates a pre-trained foundation model for joint inference of latent features and texture, improving scene understanding and generalization. sshELF can reconstruct 360 degree scenes from six sparse input views and achieves competitive results on synthetic and real-world datasets. We find that sshELF faithfully reconstructs occluded regions, supports real-time rendering, and provides rich latent features for downstream applications. The code will be released.

Training high-quality CLIP models typically requires enormous datasets, which limits the development of domain-specific models – especially in areas that even the largest CLIP models do not cover well – and drives up training costs. This poses challenges for scientific research that needs fine-grained control over the training procedure of CLIP models. In this work, we show that by employing smart web search strategies enhanced with knowledge graphs, a robust CLIP model can be trained from scratch with considerably less data. Specifically, we demonstrate that an expert foundation model for living organisms can be built using just 10M images. Moreover, we introduce EntityNet, a dataset comprising 33M images paired with 46M text descriptions, which enables the training of a generic CLIP model in significantly reduced time.

While artificial neural networks excel in unsupervised learning of non-sparse structure, classical statistical regression techniques offer better interpretability, in particular when sparseness is enforced by ℓ_1 regularization, enabling identification of which factors drive observed dynamics. We investigate how these two types of approaches can be optimally combined, exemplarily considering two-photon calcium imaging data where sparse autoregressive dynamics are to be extracted. We propose embedding a vector autoregressive (VAR) model as an interpretable regression technique into a convolutional autoencoder, which provides dimension reduction for tractable temporal modeling. A skip connection separately addresses non-sparse static spatial information, selectively channeling sparse structure into the ℓ_1-regularized VAR. ℓ_1-estimation of regression parameters is enabled by differentiating through the piecewise linear solution path. This is contrasted with approaches where the autoencoder does not adapt to the VAR model. Having an embedded statistical model also enables a testing approach for comparing temporal sequences from the same observational unit. Additionally, contribution maps visualize which spatial regions drive the learned dynamics.

We study how to train visually grounded vision-language models (VLMs) for radiology without manual spatial annotations. We introduce RefRad2D, a large-scale bilingual (German/English) dataset of 1.2M CT and MR image-text pairs derived from clinical practice, with task-specific VQA and spatial grounding subsets generated automatically via LLM-based curation and automated segmentation. Trained on this data, our model RadGrounder jointly performs report generation, visual question answering, and spatial grounding via bounding-box detection or segmentation. On external VQA benchmarks (Slake, VQA-RAD), RadGrounder achieves competitive results with specialized medical VLMs. Adding our clinical data to the training mixture improves open-ended VQA over fine-tuning on the downstream datasets alone, showing the transferability of our dataset. Crucially, adding grounding supervision does not degrade language quality, enabling spatially verifiable outputs at no cost to VQA performance.

Recent knowledge distillation (KD) methods for semantic segmentation introduce increasingly complex hand-crafted objectives, yet are typically evaluated under fixed iteration schedules. These objectives substantially increase per-iteration cost, meaning equal iteration counts do not correspond to equal training budgets. It is therefore unclear whether reported gains reflect stronger distillation signals or simply greater compute. We show that iteration-based comparisons are misleading: when wall-clock compute is matched, canonical logit- and feature-based KD outperform recent segmentation-specific methods. Under extended training, feature-based distillation achieves state-of-the-art ResNet-18 performance on Cityscapes and ADE20K. A PSPNet ResNet-18 student closely approaches its ResNet-101 teacher despite using only one quarter of the parameters, reaching 99% of the teacher's mIoU on Cityscapes (79.0 vs. 79.8) and 92% on ADE20K. Our results challenge the prevailing assumption that KD for segmentation requires task-specific mechanisms and suggest that scaling, rather than complex hand-crafted objectives, should guide future method design.

Current world models operate at a single level of abstraction, with most prioritizing perceptual fidelity while lacking the spatial reasoning and semantic understanding required for real-world downstream tasks. We present a hierarchical driving world model that factorizes future prediction across two levels operating at distinct temporal and abstraction scales: a high-level predictor that forecasts coarse scene structure over extended temporal horizons, and a low-level generator that produces detailed predictions conditioned on the high-level output. This decomposition yields high perceptual fidelity while also capturing strong spatial and semantic representations. We further show that pretraining with a diffusion forcing objective yields substantially richer internal representations than the standard teacher forcing objective, while teacher forcing – predicting only the next frame from clean context – produces more stable autoregressive rollouts. We therefore introduce a generic two-stage training paradigm that pretrains the model with diffusion forcing and fine-tunes with teacher forcing, combining the representational benefits of the former with the rollout stability of the latter. Our approach achieves state-of-the-art results across the standard suite of driving world model evaluations on established benchmarks, including long-horizon generation fidelity, steering responsiveness evaluated on counterfactual scenarios, and internal representation quality. Project page with code, demo, checkpoints and qualitative results: https://lmb-freiburg.github.io/orbis2.github.io/

Standard LLM evaluations only test capabilities or dispositions that evaluators designed them for, missing unexpected differences such as behavioral shifts between model revisions or emergent misaligned tendencies. Model diffing addresses this limitation by automatically surfacing systematic behavioral differences. Recent approaches include LLM-based methods that generate natural language descriptions and sparse autoencoder (SAE)-based methods that identify interpretable features. However, no systematic comparison of these approaches exists nor are there established evaluation criteria. We address this gap by proposing evaluation metrics for key desiderata (generalization, interestingness, and abstraction level) and use these to compare existing methods. Our results show that an improved LLM-based baseline performs comparably to the SAE-based method while typically surfacing more abstract behavioral differences.

Feed-forward 3D reconstruction models such as DUSt3R, VGGT, and Depth Anything 3 (DA3) are transformer-based foundation models that infer camera geometry and dense scene structure in a single forward pass. Trained at scale in a supervised fashion, they raise a central question: do these models build upon geometric principles akin to traditional multi-view pipelines, or do they primarily rely on learned priors arising from the large-scale training setup? We find that epipolar geometry emerges within the intermediate layers of all three models and is causally linked to correspondence patterns in attention heads. To study this, we perform a systematic analysis of their internal representations across three real-world datasets and a controlled synthetic dataset. We quantify geometric understanding by probing intermediate features, analyzing attention patterns to identify correspondence matching patterns, and performing targeted interventions at the attention level. Further, we assess the role of learned priors by applying challenging input-level perturbations, such as occlusions, scene ambiguities, and varying camera configurations, and compare them against classical multi-stage reconstruction pipelines.

Outdoor thermal comfort is influenced not only by meteorological variables air temperature, radiation and humidity at regional and local scales but also by local parameters such as mean radiant temperature and wind patterns, which vary at meter-scale within cities. All these factors can be affected by ongoing climate change. Hence, modelling future thermal comfort requires a multi-scale approach. Thermal comfort in outdoor settings can be quantified and described by thermal indices such as the Universal Thermal Climate Index (UTCI), which reflects the human response to environmental and physiological forcing. To date, several microscale modelling approaches have been proposed to model the meteorological and geometric variables that contribute to the UTCI, but they are all highly detailed, complex and computationally intensive. As a result, only individual heat waves, short case studies or single points have been modelled to estimate future heat stress conditions in cities. This study introduces a novel and efficient deep-learning model that instantly and accurately predicts thermal comfort maps across entire cities and for long periods. This model is unique in its adoption of a solitary deep learning architecture, avoiding the use of sub-models that separately model, for example, air temperature or wind speed. We will refer to this model as the Unified Human Thermal Comfort Neural Network (UHTC-NN). Training and evaluation of the UHTC-NN is based on a machine learning model from a previous study, which combines four sub-models modelling air temperature, mean radiant temperature, wind speed, and relative humidity into UTCI. The UHTC-NN has a mean absolute error of 0.5 K compared to its preceding model. The UHTC-NN enables new applications of thermal comfort modelling, including meter-scale urban climate projection to support climate adaptation management in cities. In a case study, we apply UHTC-NN to downscale 15 EURO-CORDEX climate projections with a 3-hour resolution over 30 years to generate high-resolution (1x1 m) street-level outdoor thermal comfort maps for the city of Freiburg, Germany. We compare the changes in UTCI frequency distribution and uncertainties of three different Representative Concentration Pathways (RCP2.6, 4.5 and 8.5) for the years 2070-2099 with the historical climate (1990-2019). Our study models the entire city center of Freiburg, with a domain size of 2.5x2.5 km, covering various aspects of the city's urban form. We show that the average number of hours per year with strong to extreme heat stress (UTCI >= 32°C) will increase up to three and six times for RCP2.6 and RCP8.5, respectively. The number of night-time hours with UTCI >= 20°C will increase by a factor of two and five, respectively for RCP2.6 and RCP8.5, compared to the 1990-2019 period. In addition, the 80th UTCI percentile shifts by 2°C and 4°C for RCP2.6 and RCP8.5, respectively. The presented high-resolution urban climate simulations allow us to identify intra-urban variability and daytime / nocturnal hot-spots where climate change will have the greatest impacts on outdoor thermal comfort. Such urban climate simulations therefore allow for an effective selection of areas where climate adaptation needs to be prioritized.

Compositional generalization, the ability to generate novel combinations of known concepts, is a key ingredient for visual generative models. Yet, not all mechanisms that enable or inhibit it are fully understood. In this work, we conduct a systematic study of how various design choices influence compositional generalization in image and video generation in a positive or negative way. Through controlled experiments, we identify two key factors: (i) whether the training objective operates on a discrete or continuous distribution, and (ii) to what extent conditioning provides information about the constituent concepts during training. Building on these insights, we show that relaxing the MaskGIT discrete loss with an auxiliary continuous JEPA-based objective can improve compositional performance in discrete models like MaskGIT.

Reinforcement Learning (RL) has been widely applied to many control tasks and substantially improved the performances compared to conventional control methods in many domains where the reward function is well defined. However, for many real-world problems, it is often more convenient to formulate optimization problems in terms of rewards and constraints simultaneously. Optimizing such constrained problems via reward shaping can be difficult as it requires tedious manual tuning of reward functions with several interacting terms. Recent formulations which include constraints mostly require a pre-training phase, which often needs human expertise to collect data or assumes having a sub-optimal policy readily available. We propose a new constrained RL method called CSAC-LB (Constrained Soft Actor-Critic with Log Barrier Function), which achieves competitive performance without any pre-training by applying a linear smoothed log barrier function to an additional safety critic. It implements an adaptive penalty for policy learning and alleviates the numerical issues that are known to complicate the application of the log barrier function method. As a result, we show that with CSAC-LB, we achieve state-of-the-art performance on several constrained control tasks with different levels of difficulty and evaluate our methods in a locomotion task on a real quadruped robot platform.