Matyáš Boháček

Maty Bohacek, Nino Scherrer, Nicholas Dufour et al. · stanford

The evaluation of large language models relies heavily on standardized benchmarks. These benchmarks provide useful aggregated metrics, but can obscure (i) particular sub-areas where the models are weak ("model gaps") and (ii) imbalanced coverage in the benchmarks themselves ("benchmark gaps"). To automatically uncover both types of gaps, we propose a simple new method using concept activations from sparse autoencoders, to identify fine-grained gaps on a per-concept basis. The method also benefits from grounding evaluation in the model's internal representations, as well as easy comparison across benchmarks. We applied the method to five popular open-source models and more than a dozen benchmarks, as illustrative examples. As validation of the approach, we found that our automatic, unsupervised method was able to recover model gaps that have been previously documented in the literature (e.g. relating to sycophancy), in addition to identifying novel model gaps. We were also able to automatically uncover benchmark gaps: core concepts that should fall within the scope of a given benchmark. Our "competency gaps" method can be used to complement existing benchmarks, by providing a concept-level decomposition of model behavior, and by helping benchmark developers iterate upon benchmark design. Code is available at https://competency-gaps.github.io.

Matyáš Boháček, Marek Hrúz

Today's sign language recognition models require large training corpora of laboratory-like videos, whose collection involves an extensive workforce and financial resources. As a result, only a handful of such systems are publicly available, not to mention their limited localization capabilities for less-populated sign languages. Utilizing online text-to-video dictionaries, which inherently hold annotated data of various attributes and sign languages, and training models in a few-shot fashion hence poses a promising path for the democratization of this technology. In this work, we collect and open-source the UWB-SL-Wild few-shot dataset, the first of its kind training resource consisting of dictionary-scraped videos. This dataset represents the actual distribution and characteristics of available online sign language data. We select glosses that directly overlap with the already existing datasets WLASL100 and ASLLVD and share their class mappings to allow for transfer learning experiments. Apart from providing baseline results on a pose-based architecture, we introduce a novel approach to training sign language recognition models in a few-shot scenario, resulting in state-of-the-art results on ASLLVD-Skeleton and ASLLVD-Skeleton-20 datasets with top-1 accuracy of $30.97~\%$ and $95.45~\%$, respectively.

Matyáš Boháček, Michal Bravanský, Filip Trhlík et al.

We present the Verifee Dataset: a novel dataset of news articles with fine-grained trustworthiness annotations. We develop a detailed methodology that assesses the texts based on their parameters encompassing editorial transparency, journalist conventions, and objective reporting while penalizing manipulative techniques. We bring aboard a diverse set of researchers from social, media, and computer sciences to overcome barriers and limited framing of this interdisciplinary problem. We collect over $10,000$ unique articles from almost $60$ Czech online news sources. These are categorized into one of the $4$ classes across the credibility spectrum we propose, raging from entirely trustworthy articles all the way to the manipulative ones. We produce detailed statistics and study trends emerging throughout the set. Lastly, we fine-tune multiple popular sequence-to-sequence language models using our dataset on the trustworthiness classification task and report the best testing F-1 score of $0.52$. We open-source the dataset, annotation methodology, and annotators' instructions in full length at https://verifee.ai/research to enable easy build-up work. We believe similar methods can help prevent disinformation and educate in the realm of media literacy.

Raina Panda, Daniel Fein, Arpita Singhal et al.

Artistic style transfer in generative models remains a significant challenge, as existing methods often introduce style only via model fine-tuning, additional adapters, or prompt engineering, all of which can be computationally expensive and may still entangle style with subject matter. In this paper, we introduce a training- and inference-light, interpretable method for representing and transferring artistic style. Our approach leverages an art-specific Sparse Autoencoder (SAE) on top of latent embeddings of generative image models. Trained on artistic data, our SAE learns an emergent, largely disentangled set of stylistic and compositional concepts, corresponding to style-related elements pertaining brushwork, texture, and color palette, as well as semantic and structural concepts. We call it LouvreSAE and use it to construct style profiles: compact, decomposable steering vectors that enable style transfer without any model updates or optimization. Unlike prior concept-based style transfer methods, our method requires no fine-tuning, no LoRA training, and no additional inference passes, enabling direct steering of artistic styles from only a few reference images. We validate our method on ArtBench10, achieving or surpassing existing methods on style evaluations (VGG Style Loss and CLIP Score Style) while being 1.7-20x faster and, critically, interpretable.

Matyas Bohacek, Ignacio Vilanova Echavarri

Generative Artificial Intelligence (GAI) has experienced exponential growth in recent years, partly facilitated by the abundance of large-scale open-source datasets. These datasets are often built using unrestricted and opaque data collection practices. While most literature focuses on the development and applications of GAI models, the ethical and legal considerations surrounding the creation of these datasets are often neglected. In addition, as datasets are shared, edited, and further reproduced online, information about their origin, legitimacy, and safety often gets lost. To address this gap, we introduce the Compliance Rating Scheme (CRS), a framework designed to evaluate dataset compliance with critical transparency, accountability, and security principles. We also release an open-source Python library built around data provenance technology to implement this framework, allowing for seamless integration into existing dataset-processing and AI training pipelines. The library is simultaneously reactive and proactive, as in addition to evaluating the CRS of existing datasets, it equally informs responsible scraping and construction of new datasets.

Matyas Bohacek, Hany Farid

Trained on massive amounts of human-generated content, AI-generated image synthesis is capable of reproducing semantically coherent images that match the visual appearance of its training data. We show that when retrained on even small amounts of their own creation, these generative-AI models produce highly distorted images. We also show that this distortion extends beyond the text prompts used in retraining, and that once affected, the models struggle to fully heal even after retraining on only real images.

Matyáš Boháček, Hany Farid

The 2022 Russian invasion of Ukraine is being fought on two fronts: a brutal ground war and a duplicitous disinformation campaign designed to conceal and justify Russia's actions. This campaign includes at least one example of a deep-fake video purportedly showing Ukrainian President Zelenskyy admitting defeat and surrendering. In anticipation of future attacks of this form, we describe a facial and gestural behavioral model that captures distinctive characteristics of Zelenskyy's speaking style. Trained on over eight hours of authentic video from four different settings, we show that this behavioral model can distinguish Zelenskyy from deep-fake imposters.This model can play an important role -- particularly during the fog of war -- in distinguishing the real from the fake.

Matyáš Boháček, Zhuo Cao, Marek Hrúz

Sign language recognition could significantly improve the user experience for d/Deaf people with the general consumer technology, such as IoT devices or videoconferencing. However, current sign language recognition architectures are usually computationally heavy and require robust GPU-equipped hardware to run in real-time. Some models aim for lower-end devices (such as smartphones) by minimizing their size and complexity, which leads to worse accuracy. This highly scrutinizes accurate in-the-wild applications. We build upon the SPOTER architecture, which belongs to the latter group of light methods, as it came close to the performance of large models employed for this task. By substituting its original third-party pose estimation module with the MediaPipe library, we achieve an overall state-of-the-art result on the WLASL100 dataset. Significantly, our method beats previous larger architectures while still being twice as computationally efficient and almost $11$ times faster on inference when compared to a relevant benchmark. To demonstrate our method's combined efficiency and precision, we built an online demo that enables users to translate sign lemmas of American sign language in their browsers. This is the first publicly available online application demonstrating this task to the best of our knowledge.

Matyas Bohacek, Hany Farid

AI-generated video generation continues its journey through the uncanny valley to produce content that is increasingly perceptually indistinguishable from reality. To better protect individuals, organizations, and societies from its malicious applications, we describe an effective and robust technique for distinguishing real from AI-generated human motion using multi-modal semantic embeddings. Our method is robust to the types of laundering that typically confound more low- to mid-level approaches, including resolution and compression attacks. This method is evaluated against DeepAction, a custom-built, open-sourced dataset of video clips with human actions generated by seven text-to-video AI models and matching real footage. The dataset is available under an academic license at https://www.huggingface.co/datasets/faridlab/deepaction_v1.

Vaclav Knapp, Matyas Bohacek

In video understanding tasks, particularly those involving human motion, synthetic data generation often suffers from uncanny features, diminishing its effectiveness for training. Tasks such as sign language translation, gesture recognition, and human motion understanding in autonomous driving have thus been unable to exploit the full potential of synthetic data. This paper proposes a method for generating synthetic human action video data using pose transfer (specifically, controllable 3D Gaussian avatar models). We evaluate this method on the Toyota Smarthome and NTU RGB+D datasets and show that it improves performance in action recognition tasks. Moreover, we demonstrate that the method can effectively scale few-shot datasets, making up for groups underrepresented in the real training data and adding diverse backgrounds. We open-source the method along with RANDOM People, a dataset with videos and avatars of novel human identities for pose transfer crowd-sourced from the internet.

Tomas Peterka, Matyas Bohacek

The most effective misinformation campaigns are multimodal, often combining text with images and videos taken out of context -- or fabricating them entirely -- to support a given narrative. Contemporary methods for detecting misinformation, whether in deepfakes or text articles, often miss the interplay between multiple modalities. Built around a large language model, the system proposed in this paper addresses these challenges. It analyzes both the article's text and the provenance metadata of included images and videos to determine whether they are relevant. We open-source the system prototype and interactive web interface.

Saad Hassan, Matyas Bohacek, Chaelin Kim et al.

Searching for unfamiliar American Sign Language (ASL) signs is challenging for learners because, unlike spoken languages, they cannot type a text-based query to look up an unfamiliar sign. Advances in isolated sign recognition have enabled the creation of video-based dictionaries, allowing users to submit a video and receive a list of the closest matching signs. Previous HCI research using Wizard-of-Oz prototypes has explored interface designs for ASL dictionaries. Building on these studies, we incorporate their design recommendations and leverage state-of-the-art sign-recognition technology to develop an automated video-based dictionary. We also present findings from an observational study with twelve novice ASL learners who used this dictionary during video-comprehension and question-answering tasks. Our results address human-AI interaction challenges not covered in previous WoZ research, including recording and resubmitting signs, unpredictable outputs, system latency, and privacy concerns. These insights offer guidance for designing and deploying video-based ASL dictionary systems.

Vaclav Knapp, Matyas Bohacek

Recent pose-transfer methods aim to generate temporally consistent and fully controllable videos of human action where the motion from a reference video is reenacted by a new identity. We evaluate three state-of-the-art pose-transfer methods -- AnimateAnyone, MagicAnimate, and ExAvatar -- by generating videos with actions and identities outside the training distribution and conducting a participant study about the quality of these videos. In a controlled environment of 20 distinct human actions, we find that participants, presented with the pose-transferred videos, correctly identify the desired action only 42.92% of the time. Moreover, the participants find the actions in the generated videos consistent with the reference (source) videos only 36.46% of the time. These results vary by method: participants find the splatting-based ExAvatar more consistent and photorealistic than the diffusion-based AnimateAnyone and MagicAnimate.

Matyas Bohacek, Thomas Fel, Maneesh Agrawala et al.

Despite their impressive performance, generative image models trained on large-scale datasets frequently fail to produce images with seemingly simple concepts -- e.g., human hands or objects appearing in groups of four -- that are reasonably expected to appear in the training data. These failure modes have largely been documented anecdotally, leaving open the question of whether they reflect idiosyncratic anomalies or more structural limitations of these models. To address this, we introduce a systematic approach for identifying and characterizing "conceptual blindspots" -- concepts present in the training data but absent or misrepresented in a model's generations. Our method leverages sparse autoencoders (SAEs) to extract interpretable concept embeddings, enabling a quantitative comparison of concept prevalence between real and generated images. We train an archetypal SAE (RA-SAE) on DINOv2 features with 32,000 concepts -- the largest such SAE to date -- enabling fine-grained analysis of conceptual disparities. Applied to four popular generative models (Stable Diffusion 1.5/2.1, PixArt, and Kandinsky), our approach reveals specific suppressed blindspots (e.g., bird feeders, DVD discs, and whitespaces on documents) and exaggerated blindspots (e.g., wood background texture and palm trees). At the individual datapoint level, we further isolate memorization artifacts -- instances where models reproduce highly specific visual templates seen during training. Overall, we propose a theoretically grounded framework for systematically identifying conceptual blindspots in generative models by assessing their conceptual fidelity with respect to the underlying data-generating process.

Tomas Peterka, Matyas Bohacek

Out-of-context and misattributed imagery is the leading form of media manipulation in today's misinformation and disinformation landscape. The existing methods attempting to detect this practice often only consider whether the semantics of the imagery corresponds to the text narrative, missing manipulation so long as the depicted objects or scenes somewhat correspond to the narrative at hand. To tackle this, we introduce News Media Provenance Dataset, a dataset of news articles with provenance-tagged images. We formulate two tasks on this dataset, location of origin relevance (LOR) and date and time of origin relevance (DTOR), and present baseline results on six large language models (LLMs). We identify that, while the zero-shot performance on LOR is promising, the performance on DTOR hinders, leaving room for specialized architectures and future work.