Publications - Current Year

Vision-language models such as CLIP are highly useful for diverse tasks due to their shared image-text embedding space. Despite this, the image and text embeddings are often poorly aligned, affecting downstream performance. Recent work has shown that this can be attributed to an information imbalance: images contain more information than their captions describe. In this work, we propose TEVI, a framework that uses captions as a signal for what to retain from image embeddings. Specifically, we use sparse autoencoders to disentangle image embeddings and train a masking module to selectively reconstruct the embedding based on a given caption. In a controlled setup with synthetic captions, we show that TEVI is effective at preserving caption-described attributes while discarding others. By applying TEVI to CLIP models trained on natural images, we further achieve improved retrieval performance across coarse-grained short-caption (MS COCO, Flickr) and fine-grained long-caption (IIW, DOCCI) benchmarks, with stronger gains on richer captions, and improved robustness on the RoCOCO benchmark.

Deep learning has become an essential part of computer vision, with deep
neural networks (DNNs) excelling in predictive performance. However, they often
fall short in other critical quality dimensions, such as robustness,
calibration, or fairness. While existing studies have focused on a subset of
these quality dimensions, none have explored a more general form of
"well-behavedness" of DNNs. With this work, we address this gap by
simultaneously studying nine different quality dimensions for image
classification. Through a large-scale study, we provide a bird's-eye view by
analyzing 326 backbone models and how different training paradigms and model
architectures affect the quality dimensions. We reveal various new insights
such that (i) vision-language models exhibit high fairness on ImageNet-1k
classification and strong robustness against domain changes; (ii)
self-supervised learning is an effective training paradigm to improve almost
all considered quality dimensions; and (iii) the training dataset size is a
major driver for most of the quality dimensions. We conclude our study by
introducing the QUBA score (Quality Understanding Beyond Accuracy), a novel
metric that ranks models across multiple dimensions of quality, enabling
tailored recommendations based on specific user needs.

Understanding how neural networks arrive at their predictions is essential for debugging, auditing, and deployment. Mechanistic interpretability pursues this goal by identifying circuits - minimal subnetworks responsible for specific behaviors. However, existing circuit discovery methods are brittle: circuits depend strongly on the chosen concept dataset and often fail to transfer out-of-distribution, raising doubts whether they capture concept or dataset-specific artifacts. We introduce Certified Circuits, which provide provable stability guarantees for circuit discovery. Our framework wraps any black-box discovery algorithm with randomized data subsampling to certify that circuit component inclusion decisions are invariant to bounded edit-distance perturbations of the concept dataset. Unstable neurons are abstained from, yielding circuits that are more compact and more accurate. On ImageNet and OOD datasets, certified circuits achieve up to 91% higher accuracy while using 45% fewer neurons, and remain reliable where baselines degrade. Certified Circuits puts circuit discovery on formal ground by producing mechanistic explanations that are provably stable and better aligned with the target concept. Code will be released soon!

We present SceneTok, a novel tokenizer for encoding view sets of scenes into a compressed and diffusable set of unstructured tokens. Existing approaches for 3D scene representation and generation commonly use 3D data structures or view-aligned fields. In contrast, we introduce the first method that encodes scene information into a small set of permutation-invariant tokens that is disentangled from the spatial grid. The scene tokens are predicted by a multi-view tokenizer given many context views and rendered into novel views by employing a light-weight rectified flow decoder. We show that the compression is 1-3 orders of magnitude stronger than for other representations while still reaching state-of-the-art reconstruction quality. Further, our representation can be rendered from novel trajectories, including ones deviating from the input trajectory, and we show that the decoder gracefully handles uncertainty. Finally, the highly-compressed set of unstructured latent scene tokens enables simple and efficient scene generation in 5 seconds, achieving a much better quality-speed trade-off than previous paradigms.

Deep neural networks (DNNs) are widely used, but interpreting what they actually learn remains difficult. A major obstacle is that individual neurons often encode multiple unrelated concepts, obscuring the decision process of the network. While prior work, such as sparse autoencoders, can separate these mixed signals into more meaningful, "monosemantic" features, this typically requires altering the model in ways that can degrade downstream performance. To overcome this, we introduce ELUDe (explicit, lossless, unsupervised disentanglement), a method for improving the interpretability of DNNs while preserving their functional equivalence. ELUDe breaks latent representations into clear, inspectable sub-units that behave like interpretable features, while guaranteeing that the model's outputs remain exactly the same. It requires no explicit training, no labels, and can be applied to pretrained models. ELUDe works by reorganizing how information flows between layers, re-routing concept-specific contributions while preserving the original computation by construction. Across several vision models, including DINOv2 and supervised ViT-B/16, ELUDe improves interpretability, keeps downstream accuracy unchanged, runs efficiently, and supports practical uses such as steering model representations. In short, ELUDe offers interpretability (almost) without a tradeoff: clearer, scalable, and actionable model insights with no loss in performance.

We present SemanticNVS, a camera-conditioned multi-view diffusion model for novel view synthesis (NVS), which improves generation quality and consistency by integrating pre-trained semantic feature extractors. Existing NVS methods perform well for views near the input view, however, they tend to generate semantically implausible and distorted images under long-range camera motion, revealing severe degradation. We speculate that this degradation is due to current models failing to fully understand their conditioning or intermediate generated scene content. Here, we propose to integrate pre-trained semantic feature extractors to incorporate stronger scene semantics as conditioning to achieve high-quality generation even at distant viewpoints. We investigate two different strategies, (1) warped semantic features and (2) an alternating scheme of understanding and generation at each denoising step. Experimental results on multiple datasets demonstrate the clear qualitative and quantitative (4.69%-15.26% in FID) improvement over state-of-the-art alternatives.

Semantic segmentation requires dense pixel-level annotations, which are costly and time-consuming to acquire. To address this, we present SeSAM, a framework that uses a foundational segmentation model, i.e. Segment Anything Model (SAM), with weak labels, including coarse masks, scribbles, and points. SAM, originally designed for instance-based segmentation, cannot be directly used for semantic segmentation tasks. In this work, we identify specific challenges faced by SAM and determine appropriate components to adapt it for class-based segmentation using weak labels. Specifically, SeSAM decomposes class masks into connected components, samples point prompts along object skeletons, selects SAM masks using weak-label coverage, and iteratively refines labels using pseudo-labels, enabling SAM-generated masks to be effectively used for semantic segmentation. Integrated with a semi-supervised learning framework, SeSAM balances ground-truth labels, SAM-based pseudo-labels, and high-confidence pseudo-labels, significantly improving segmentation quality. Extensive experiments across multiple benchmarks and weak annotation types show that SeSAM consistently outperforms weakly supervised baselines while substantially reducing annotation cost relative to fine supervision.

Measuring structured object understanding in vision foundation models remains challenging due to inconsistent evaluation protocols and limited part-level supervision. Semantic correspondence (SC) evaluates this capability by testing whether object parts can be matched across instances and categories under large variations in appearance, viewpoint, and geometry. To enable a systematic SC evaluation, we introduce SOCO, a new benchmark for Semantic Object Correspondence that introduces a taxonomy of correspondence types and provides consistent, functionally meaningful keypoint annotations across 100 categories and over 1M correspondence pairs. In addition, SOCO includes keypoint language descriptions, enabling the evaluation of large vision-language models (LVLMs) and their fine-grained part-level understanding. Comprehensive experiments reveal that (i) vision foundation backbones encode strong semantic structure but transfer correspondences poorly across related categories and only partially capture object-part position, (ii) LVLMs are stronger at text-prompted part localization than at visual-reference cross-image matching, exposing a gap between language-grounded localization and fine-grained visual correspondence, and (iii) correspondence performance predicts performance on dense downstream tasks, including segmentation, tracking, 3D pose estimation, and 3D detection, more strongly than ImageNet classification. Together, these findings position SOCO as a benchmark for structured, part-level representation quality in vision and multimodal foundation models.

We present Rewis3d, a framework that leverages recent advances in feed-forward 3D reconstruction to significantly improve weakly supervised semantic segmentation on 2D images. Obtaining dense, pixel-level annotations remains a costly bottleneck for training segmentation models. Alleviating this issue, sparse annotations offer an efficient weakly-supervised alternative. However, they still incur a performance gap. To address this, we introduce a novel approach that leverages 3D scene reconstruction as an auxiliary supervisory signal. Our key insight is that 3D geometric structure recovered from 2D videos provides strong cues that can propagate sparse annotations across entire scenes. Specifically, a dual student-teacher architecture enforces semantic consistency between 2D images and reconstructed 3D point clouds, using state-of-the-art feed-forward reconstruction to generate reliable geometric supervision. Extensive experiments demonstrate that Rewis3d achieves state-of-the-art performance in sparse supervision, outperforming existing approaches by 2-7% without requiring additional labels or inference overhead.

Building performant Vision-Language Models (VLMs) requires carefully curating large-scale training datasets, yet the community lacks systematic benchmarks for evaluating such curation strategies. We introduce DataComp for VLMs (DCVLM), a benchmark for controlled data-centric experiments to improve VLM training. As part of DCVLM, we collect 160 datasets spanning four data types -- image-caption pairs, multimodal interleaved documents, text-only, and instruction-tuning data -- into a corpus of 6T multimodal tokens. DCVLM allows participants to test curation strategies (filtering, mixing, formatting, sampling) across 1B-8B models and 6.25B-200B token budgets. Models are then evaluated on a carefully selected suite of up to 52 downstream benchmarks across 9 domains. We conduct extensive experiments on DCVLM and find that data mixing, not filtering, is key to a high-quality training dataset: instruction-heavy mixtures scale better than caption-heavy ones, with gains widening at larger scales. The resulting dataset, DCVLM-Baseline, enables training an 8B VLM to 63.6% accuracy on our 33-task core suite with 200B training tokens. Compared to FineVision, the state-of-the-art open VLM training dataset, this represents an improvement of +5.4pp. DCVLM and all accompanying artifacts will be made publicly available at www.datacomp.ai/dcvlm/.

Most vision models are trained on RGB images processed through ISP pipelines optimized for human perception, which can discard sensor-level information useful for machine reasoning. RAW images preserve unprocessed scene data, enabling models to leverage richer cues for both object detection and object description, capturing fine-grained details, spatial relationships, and contextual information often lost in processed images. To support research in this domain, we introduce RAWDet-7, a large-scale dataset of ~25k training and 7.6k test RAW images collected across diverse cameras, lighting conditions, and environments, densely annotated for seven object categories following MS-COCO and LVIS conventions. In addition, we provide object-level descriptions derived from the corresponding high-resolution sRGB images, facilitating the study of object-level information preservation under RAW image processing and low-bit quantization. The dataset allows evaluation under simulated 4-bit, 6-bit, and 8-bit quantization, reflecting realistic sensor constraints, and provides a benchmark for studying detection performance, description quality & detail, and generalization in low-bit RAW image processing. Dataset & code upon acceptance.

We introduce Contrast Sensitive Flow (CSFlow), a weighting scheme that connects the human eye's Contrast Sensitivity Function (CSF) to the iterative denoising steps of flow matching. Because real-world images concentrate signal at low spatial frequencies, these components reach high signal-to-noise ratio earlier during continuous diffusion than high-frequency components. When generating images with diffusion or flow matching models, this induces a soft autoregressive structure in Fourier space, where coarse image content stabilizes before fine detail. Meanwhile, the human visual system is unequally sensitive to spatial frequencies: very low and very high frequencies require significantly higher contrast to be perceived. We for the first time merge these observations through two contributions: (1) a metric that estimates which frequencies are generated at each reverse flow interval and (2) timestep weights obtained by aligning the frequencies generated at each noise level with human contrast sensitivity. We validate our contributions experimentally showing that these weights can improve generative performance by lowering FID by 4.7%, increasing Inception Score by 2.2% and improving GenEval scores by 2.5% using inference-only timestep modification or short fine-tuning. Qualitatively, we find that our CSFlow weights lead to better visual realism and less cartoonish appearance of generated images.

Explainable artificial intelligence (XAI) aims to provide human-interpretable insights into the behavior of deep neural networks (DNNs), typically by estimating a simplified causal structure of the model. In existing work, this causal structure often includes relationships where the presence of a concept is associated with a strong activation of a neuron. For example, attribution methods primarily identify input pixels that contribute most to a prediction, and feature visualization methods reveal inputs that cause high activation of a target neuron - the former implicitly assuming that the relevant information resides in the input, and the latter that neurons encode the presence of concepts. However, a largely overlooked type of causal relationship is that of encoded absences, where the absence of a concept increases neural activation. In this work, we show that such missing but relevant concepts are common and that mainstream XAI methods struggle to reveal them when applied in their standard form. To address this, we propose two simple extensions to attribution and feature visualization techniques that uncover encoded absences. Across experiments, we show how mainstream XAI methods can be used to reveal and explain encoded absences, how ImageNet models exploit them, and that debiasing can be improved when considering them.

End-to-end autonomous driving has gained significant attention for its potential to learn robust behavior in interactive scenarios and scale with data. Popular architectures often build on separate modules for perception and planning connected through latent representations, such as bird's eye view feature grids, to maintain end-to-end differentiability. This paradigm emerged mostly on open-loop datasets, with evaluation focusing not only on driving performance, but also intermediate perception tasks. Unfortunately, architectural advances that excel in open-loop often fail to translate to scalable learning of robust closed-loop driving. In this paper, we systematically re-examine the impact of common architectural patterns on closed-loop performance: (1) high-resolution perceptual representations, (2) disentangled trajectory representations, and (3) generative planning. Crucially, our analysis evaluates the combined impact of these patterns, revealing both unexpected limitations as well as underexplored synergies. Building on these insights, we introduce BevAD, a novel lightweight and highly scalable end-to-end driving architecture. BevAD achieves 72.7% success rate on the Bench2Drive benchmark and demonstrates strong data-scaling behavior using pure imitation learning. Our code and models are publicly available here: dmholtz.github.io/bevad/

The purpose of predictive modeling on relational data is to predict future or missing values in a relational database, for example, future purchases of a user, risk of readmission of the patient, or the likelihood that a financial transaction is fraudulent. Typically powered by machine learning methods, predictive models are used in recommendations, financial fraud detection, supply chain optimization, and other systems, providing billions of predictions every day. However, training a machine learning model requires manual work to extract the required training examples - prediction entities and target labels - from the database, which is slow, laborious, and prone to mistakes. Here, we present the Predictive Query Language (PQL), an SQL-inspired declarative language for defining predictive tasks on relational databases. PQL allows specifying a predictive task in a single declarative query, enabling the automatic computation of training labels for a large variety of machine learning tasks, such as regression, classification, time-series forecasting, and recommender systems. PQL is already successfully integrated and used in a collection of use cases as part of a predictive AI platform. The versatility of the language can be demonstrated through its many ongoing use cases, including financial fraud, item recommendations, and workload prediction. We demonstrate its versatile design through two implementations; one for small-scale, low-latency use and one that can handle large-scale databases.

Large Vision-Language Models (LVLMs) map visual inputs into dense token sequences, imposing a quadratic computational bottleneck for inference. Elastic visual-token compression addresses this by training a single model that can run at multiple visual-token budgets. However, existing approaches struggle under aggressive compression. Spatial-only compression, as in nested pooling, behaves as an imperfect low-pass filter and induces spectral aliasing that obscures fine-grained detail. Query-only compression, as in nested query resampling, replaces explicit grid-aligned tokens with non-local summaries and substantially degrades spatial grounding. To resolve this representational conflict, we introduce PARCEL (Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding), a visual tokenization architecture that dynamically partitions the labor of feature extraction. PARCEL establishes spatial pool tokens as low-frequency layout anchors and conditions elastic query tokens on these anchors through Pool-Conditioned Query Resampling. This encourages query tokens to focus on complementary visual features rather than redundant spatial mapping. Extensive evaluations across 27 benchmarks show that PARCEL improves the performance-efficiency Pareto frontier, consistently outperforming existing matryoshka baselines across visual-token budgets while preserving the "train once, deploy anywhere" paradigm.

Large vision-language models (LVLMs) tend to hallucinate, especially when visual inputs are corrupted at test time. We show that such corruptions act as additional distribution shifts, significantly amplifying hallucination rates in real-world applications. To address this, we propose CLIP-guided Test-Time Training (ClipTTT), a method to adapt LVLMs under degraded conditions on the fly with a single test sample. Specifically, we leverage the image-text alignment strength of a pre-trained CLIP model as a stable guidance signal to identify reliable self-supervision targets, enabling rapid adaptation without altering the base LVLMs. Extensive experiments on standard hallucination benchmarks, with 15 common corruptions, demonstrate that ClipTTT effectively mitigates hallucinations and improves descriptive faithfulness under visual corruptions.

We present ActionPlan, a unified motion diffusion framework that bridges real-time streaming with high-quality offline generation within a single model. The core idea is to introduce a per-frame action plan: the model predicts frame-level text latents that act as dense semantic anchors throughout denoising, and uses them to denoise the full motion sequence with combined semantic and motion cues. To support this structured workflow, we design latent-specific diffusion steps, allowing each motion latent to be denoised independently and sampled in flexible orders at inference. As a result, ActionPlan can run in a history-conditioned, future-aware mode for real-time streaming, while also supporting high-quality offline generation. The same mechanism further enables zero-shot motion editing and in-betweening without additional models. Experiments demonstrate that our real-time streaming is 5.25x faster while also achieving 18% motion quality improvement over the best previous method in terms of FID.

Social media platforms have become primary arenas for climate communication, generating millions of images and posts that - if systematically analysed - can reveal which communication strategies mobilise public concern and which fall flat. We aim to facilitate such research by analysing how computer vision methods can be used for social media discourse analysis. This analysis includes application-based taxonomy design, model selection, prompt engineering, and validation. We benchmark six promptable vision-language models and 15 zero-shot CLIP-like models on two datasets from X (formerly Twitter) - a 1,038-image expert-annotated set and a larger corpus of over 1.2 million images, with 50,000 labels manually validated - spanning five annotation dimensions: animal content, climate change consequences, climate action, image setting, and image type. Among the models benchmarked, Gemini-3.1-flash-lite outperforms all others across all super-categories and both datasets, while the gap to open-weight models of moderate size remains relatively small. Beyond instance-level metrics, we advocate for distributional evaluation: VLM predictions can reliably recover population level trends even when per-image accuracy is moderate, making them a viable starting point for discourse analysis at scale. We find that chain-of-thought reasoning reduces rather than improves performance, and that annotation dimension specific prompt design improves performance. We release tweet IDs and labels along with our code at github.com/KathPra/Codebooks2VLMs.git.

Precise control of 3D facial expressions from text is crucial for virtual avatars, animation, and human-computer interaction, yet existing text-to-3D methods jointly generate identity, expression, and texture, making fine-grained expression control difficult. We instead formulate text-driven expression synthesis as a regression problem in the disentangled parameter space of a 3D Morphable Model (3DMM). This setting, however, requires paired data linking detailed language to precise expression parameters, which are missing from existing resources. To fill this gap, we introduce Txt2Emote, a benchmark of diverse 3D facial expressions with fine-grained textual annotations obtained from GPT-4o and a high-fidelity face tracker, providing both explicit descriptions detailing facial features and implicit descriptions referencing the situational context behind the expression. Leveraging this dataset, we present EmoteGPT, a text-to-3D expression framework based on a Multimodal Large Language Model (MLLM) with a dedicated <expr> token to semantically ground expression representations, which are then decoded into 3DMM parameters. We further improve EmoteGPT by augmenting training with large-scale image-to-3DMM data, enabling it to surpass state-of-the-art text-to-3D face synthesis methods on emotion recognition metrics and in perceived expressiveness. Integrated into avatar pipelines, our method enables photorealistic and stylized 3D avatars, as well as expressive 3D-consistent 2D face synthesis from textual input.</expr>

Vision Transformers (ViTs) have become a dominant architecture in computer vision, yet producing stable and high-resolution attribution maps for these models remains challenging. Architectural components such as patch embeddings and attention routing often introduce structured artifacts in pixel-level explanations, causing many existing methods to rely on coarse patch-level attributions. We introduce DAVE \textit{(\underline{D}istribution-aware \underline{A}ttribution via \underline{V}iT Gradient D\underline{E}composition)}, a mathematically grounded attribution method for ViTs based on a structured decomposition of the input gradient. By exploiting architectural properties of ViTs, DAVE isolates locally equivariant and stable components of the effective input--output mapping. It separates these from architecture-induced artifacts and other sources of instability.

Reinforcement learning (RL) with verifiable rewards (RLVR) has demonstrated the great potential of enhancing the reasoning abilities in multimodal large language models (MLLMs). However, the reliance on language-centric priors and expensive manual annotations prevents MLLMs' intrinsic visual understanding and scalable reward designs. In this work, we introduce SSL-R1, a generic self-supervised RL framework that derives verifiable rewards directly from images. To this end, we revisit self-supervised learning (SSL) in visual domains and reformulate widely-used SSL tasks into a set of verifiable visual puzzles for RL post-training, requiring neither human nor external model supervision. Training MLLMs on these tasks substantially improves their performance on multimodal understanding and reasoning benchmarks, highlighting the potential of leveraging vision-centric self-supervised tasks for MLLM post-training. We think this work will provide useful experience in devising effective self-supervised verifiable rewards to enable RL at scale. Project page: github.com/Jiahao000/SSL-R1.

Large vision-language models (LVLMs) have demonstrated impressive performance in various multimodal understanding and reasoning tasks. However, they still struggle with object hallucinations, i.e., the claim of nonexistent objects in the visual input. To address this challenge, we propose Region-aware Chain-of-Verification (R-CoV), a visual chain-of-verification method to alleviate object hallucinations in LVLMs in a post-hoc manner. Motivated by how humans comprehend intricate visual information -- often focusing on specific image regions or details within a given sample -- we elicit such region-level processing from LVLMs themselves and use it as a chaining cue to detect and alleviate their own object hallucinations. Specifically, our R-CoV consists of six steps: initial response generation, entity extraction, coordinate generation, region description, verification execution, and final response generation. As a simple yet effective method, R-CoV can be seamlessly integrated into various LVLMs in a training-free manner and without relying on external detection models. Extensive experiments on several widely used hallucination benchmarks across multiple LVLMs demonstrate that R-CoV can significantly alleviate object hallucinations in LVLMs. Project page: github.com/Jiahao000/R-CoV.

The pervasive growth of digital content, specifically short videos on social media platforms, has significantly altered how topics are discussed and understood in public discourse. In this work, we advance automated visual theme detection by assessing zero-shot and clustering capabilities on social media data. (1) We evaluated the capabilities of notable VLMs such as VideoChatGPT, PandaGPT, and VideoLLava using zero-shot image classification and compared their performance to the baseline provided by frame-wise CLIP image classification. (2) By treating clustering as a minimum cost multicut problem, we aim to uncover insightful patterns in an unsupervised manner. For both analysis strategies, we provide extensive evaluations and practical guidance to practitioners. While VLMs are currently not able to detect climate change specific classes, the clustering results are distinct visual frames. %Given that VLMs are not currently capable to grasp the climate change discourse, we focus the clustering evaluation of image embedding models. We find that both ConvNeXt V2 and DINOv2 produce meaningful clusters, with DINOv2 focusing more on style differences and abstract categories, while ConvNeXt V2 clusters differ in more fine-grained ways. Code available at github.com/KathPra/ClimateVID.git.

Transparency of neural networks' internal reasoning is at the heart of interpretability research, adding to trust, safety, and understanding of these models. The field of mechanistic interpretability has recently focused on studying task-specific computational graphs, defined by connections (edges) between model components. Such edge-based circuits have been defined in the context of large language models, yet vision-based approaches so far only consider neuron-based circuits. These tell which information is encoded, but not how it is routed through the complex wiring of a neural network. In this work, we investigate whether useful mechanistic circuits can be identified through computational graphs in vision transformers. We propose an effective method for Automatic Visual Circuit Discovery (Vi-CD) that recovers class-specific circuits for classification, identifies circuits underlying typographic attacks in CLIP, and discovers circuits that lend themselves for steering to correct harmful model behavior. Overall, we find that insightful and actionable edge-based circuits can be recovered from vision transformers, adding transparency to the internal computations of these models.