Vasudev Lal

Le Xue, Manli Shu, Anas Awadalla et al. · salesforce, stanford

This paper introduces BLIP-3, an open framework for developing Large Multimodal Models (LMMs). The framework comprises meticulously curated datasets, a training recipe, model architectures, and a resulting suite of LMMs. We release 4B and 14B models, including both the pre-trained base model and the instruction fine-tuned ones. Our models undergo rigorous evaluation across a range of tasks, including both single and multi-image benchmarks. Our models demonstrate competitive performance among open-source LMMs with similar model sizes. Our resulting LMMs demonstrate competitive performance among open-source LMMs with similar model sizes, with the ability to comprehend interleaved image-text inputs. Our training code, models, and all datasets used in this work, including the three largescale datasets we create and the preprocessed ones, will be open-sourced to better support the research community.

Xiao Xu, Chenfei Wu, Shachar Rosenman et al.

Vision-Language (VL) models with the Two-Tower architecture have dominated visual-language representation learning in recent years. Current VL models either use lightweight uni-modal encoders and learn to extract, align and fuse both modalities simultaneously in a deep cross-modal encoder, or feed the last-layer uni-modal representations from the deep pre-trained uni-modal encoders into the top cross-modal encoder. Both approaches potentially restrict vision-language representation learning and limit model performance. In this paper, we propose BridgeTower, which introduces multiple bridge layers that build a connection between the top layers of uni-modal encoders and each layer of the cross-modal encoder. This enables effective bottom-up cross-modal alignment and fusion between visual and textual representations of different semantic levels of pre-trained uni-modal encoders in the cross-modal encoder. Pre-trained with only 4M images, BridgeTower achieves state-of-the-art performance on various downstream vision-language tasks. In particular, on the VQAv2 test-std set, BridgeTower achieves an accuracy of 78.73%, outperforming the previous state-of-the-art model METER by 1.09% with the same pre-training data and almost negligible additional parameters and computational costs. Notably, when further scaling the model, BridgeTower achieves an accuracy of 81.15%, surpassing models that are pre-trained on orders-of-magnitude larger datasets. Code and checkpoints are available at https://github.com/microsoft/BridgeTower.

Phillip Howard, Gadi Singer, Vasudev Lal et al.

While counterfactual data augmentation offers a promising step towards robust generalization in natural language processing, producing a set of counterfactuals that offer valuable inductive bias for models remains a challenge. Most existing approaches for producing counterfactuals, manual or automated, rely on small perturbations via minimal edits, resulting in simplistic changes. We introduce NeuroCounterfactuals, designed as loose counterfactuals, allowing for larger edits which result in naturalistic generations containing linguistic diversity, while still bearing similarity to the original document. Our novel generative approach bridges the benefits of constrained decoding, with those of language model adaptation for sentiment steering. Training data augmentation with our generations results in both in-domain and out-of-domain improvements for sentiment classification, outperforming even manually curated counterfactuals, under select settings. We further present detailed analyses to show the advantages of NeuroCounterfactuals over approaches involving simple, minimal edits.

Avinash Madasu, Vasudev Lal

Video retrieval (VR) involves retrieving the ground truth video from the video database given a text caption or vice-versa. The two important components of compositionality: objects & attributes and actions are joined using correct syntax to form a proper text query. These components (objects & attributes, actions and syntax) each play an important role to help distinguish among videos and retrieve the correct ground truth video. However, it is unclear what is the effect of these components on the video retrieval performance. We therefore, conduct a systematic study to evaluate the compositional and syntactic understanding of video retrieval models on standard benchmarks such as MSRVTT, MSVD and DIDEMO. The study is performed on two categories of video retrieval models: (i) which are pre-trained on video-text pairs and fine-tuned on downstream video retrieval datasets (Eg. Frozen-in-Time, Violet, MCQ etc.) (ii) which adapt pre-trained image-text representations like CLIP for video retrieval (Eg. CLIP4Clip, XCLIP, CLIP2Video etc.). Our experiments reveal that actions and syntax play a minor role compared to objects & attributes in video understanding. Moreover, video retrieval models that use pre-trained image-text representations (CLIP) have better syntactic and compositional understanding as compared to models pre-trained on video-text data. The code is available at https://github.com/IntelLabs/multimodal_cognitive_ai/tree/main/ICSVR

Musashi Hinck, Carolin Holtermann, Matthew Lyle Olson et al.

We uncover a surprising multilingual bias occurring in a popular class of multimodal vision-language models (VLMs). Including an image in the query to a LLaVA-style VLM significantly increases the likelihood of the model returning an English response, regardless of the language of the query. This paper investigates the causes of this loss with a two-pronged approach that combines extensive ablation of the design space with a mechanistic analysis of the models' internal representations of image and text inputs. Both approaches indicate that the issue stems in the language modelling component of the LLaVA model. Statistically, we find that switching the language backbone for a bilingual language model has the strongest effect on reducing this error. Mechanistically, we provide compelling evidence that visual inputs are not mapped to a similar space as text ones, and that intervening on intermediary attention layers can reduce this bias. Our findings provide important insights to researchers and engineers seeking to understand the crossover between multimodal and multilingual spaces, and contribute to the goal of developing capable and inclusive VLMs for non-English contexts.

Sungduk Yu, Brian L. White, Anahita Bhiwandiwalla et al.

Detecting and attributing temperature increases driven by climate change is crucial for understanding global warming and informing adaptation strategies. However, distinguishing human-induced climate signals from natural variability remains challenging for traditional detection and attribution (D&A) methods, which rely on identifying specific "fingerprints" -- spatial patterns expected to emerge from external forcings such as greenhouse gas emissions. Deep learning offers promise in discerning these complex patterns within expansive spatial datasets, yet the lack of standardized protocols has hindered consistent comparisons across studies. To address this gap, we introduce ClimDetect, a standardized dataset comprising 1.17M daily climate snapshots paired with target climate change indicator variables. The dataset is curated from both CMIP6 climate model simulations and real-world observation-assimilated reanalysis datasets (ERA5, JRA-3Q, and MERRA-2), and is designed to enhance model accuracy in detecting climate change signals. ClimDetect integrates various input and target variables used in previous research, ensuring comparability and consistency across studies. We also explore the application of vision transformers (ViT) to climate data -- a novel approach that, to our knowledge, has not been attempted before for climate change detection tasks. Our open-access data serve as a benchmark for advancing climate science by enabling end-to-end model development and evaluation. ClimDetect is publicly accessible via Hugging Face dataset repository at: https://huggingface.co/datasets/ClimDetect/ClimDetect.

Phillip Howard, Avinash Madasu, Tiep Le et al.

While vision-language models (VLMs) have achieved remarkable performance improvements recently, there is growing evidence that these models also posses harmful biases with respect to social attributes such as gender and race. Prior studies have primarily focused on probing such bias attributes individually while ignoring biases associated with intersections between social attributes. This could be due to the difficulty of collecting an exhaustive set of image-text pairs for various combinations of social attributes. To address this challenge, we employ text-to-image diffusion models to produce counterfactual examples for probing intersectional social biases at scale. Our approach utilizes Stable Diffusion with cross attention control to produce sets of counterfactual image-text pairs that are highly similar in their depiction of a subject (e.g., a given occupation) while differing only in their depiction of intersectional social attributes (e.g., race & gender). Through our over-generate-then-filter methodology, we produce SocialCounterfactuals, a high-quality dataset containing 171k image-text pairs for probing intersectional biases related to gender, race, and physical characteristics. We conduct extensive experiments to demonstrate the usefulness of our generated dataset for probing and mitigating intersectional social biases in state-of-the-art VLMs.

Tiep Le, Vasudev Lal, Phillip Howard

Counterfactual examples have proven to be valuable in the field of natural language processing (NLP) for both evaluating and improving the robustness of language models to spurious correlations in datasets. Despite their demonstrated utility for NLP, multimodal counterfactual examples have been relatively unexplored due to the difficulty of creating paired image-text data with minimal counterfactual changes. To address this challenge, we introduce a scalable framework for automatic generation of counterfactual examples using text-to-image diffusion models. We use our framework to create COCO-Counterfactuals, a multimodal counterfactual dataset of paired image and text captions based on the MS-COCO dataset. We validate the quality of COCO-Counterfactuals through human evaluations and show that existing multimodal models are challenged by our counterfactual image-text pairs. Additionally, we demonstrate the usefulness of COCO-Counterfactuals for improving out-of-domain generalization of multimodal vision-language models via training data augmentation.

Avinash Madasu, Estelle Aflalo, Gabriela Ben Melech Stan et al.

Multi-modal retrieval has seen tremendous progress with the development of vision-language models. However, further improving these models require additional labelled data which is a huge manual effort. In this paper, we propose a framework MuMUR, that utilizes knowledge transfer from a multilingual model to boost the performance of multi-modal (image and video) retrieval. We first use state-of-the-art machine translation models to construct pseudo ground-truth multilingual visual-text pairs. We then use this data to learn a joint vision-text representation where English and non-English text queries are represented in a common embedding space based on pretrained multilingual models. We evaluate our proposed approach on a diverse set of retrieval datasets: five video retrieval datasets such as MSRVTT, MSVD, DiDeMo, Charades and MSRVTT multilingual, two image retrieval datasets such as Flickr30k and Multi30k . Experimental results demonstrate that our approach achieves state-of-the-art results on all video retrieval datasets outperforming previous models. Additionally, our framework MuMUR significantly beats other multilingual video retrieval dataset. We also observe that MuMUR exhibits strong performance on image retrieval. This demonstrates the universal ability of MuMUR to perform retrieval across all visual inputs (image and video) and text inputs (monolingual and multilingual).

Phillip Howard, Arden Ma, Vasudev Lal et al.

The extraction of aspect terms is a critical step in fine-grained sentiment analysis of text. Existing approaches for this task have yielded impressive results when the training and testing data are from the same domain. However, these methods show a drastic decrease in performance when applied to cross-domain settings where the domain of the testing data differs from that of the training data. To address this lack of extensibility and robustness, we propose a novel approach for automatically constructing domain-specific knowledge graphs that contain information relevant to the identification of aspect terms. We introduce a methodology for injecting information from these knowledge graphs into Transformer models, including two alternative mechanisms for knowledge insertion: via query enrichment and via manipulation of attention patterns. We demonstrate state-of-the-art performance on benchmark datasets for cross-domain aspect term extraction using our approach and investigate how the amount of external knowledge available to the Transformer impacts model performance.

Shachar Rosenman, Vasudev Lal, Phillip Howard

Despite impressive recent advances in text-to-image diffusion models, obtaining high-quality images often requires prompt engineering by humans who have developed expertise in using them. In this work, we present NeuroPrompts, an adaptive framework that automatically enhances a user's prompt to improve the quality of generations produced by text-to-image models. Our framework utilizes constrained text decoding with a pre-trained language model that has been adapted to generate prompts similar to those produced by human prompt engineers. This approach enables higher-quality text-to-image generations and provides user control over stylistic features via constraint set specification. We demonstrate the utility of our framework by creating an interactive application for prompt enhancement and image generation using Stable Diffusion. Additionally, we conduct experiments utilizing a large dataset of human-engineered prompts for text-to-image generation and show that our approach automatically produces enhanced prompts that result in superior image quality. We make our code and a screencast video demo of NeuroPrompts publicly available.

Phillip Howard, Avinash Madasu, Tiep Le et al.

While vision-language models (VLMs) have achieved remarkable performance improvements recently, there is growing evidence that these models also posses harmful biases with respect to social attributes such as gender and race. Prior studies have primarily focused on probing such bias attributes individually while ignoring biases associated with intersections between social attributes. This could be due to the difficulty of collecting an exhaustive set of image-text pairs for various combinations of social attributes from existing datasets. To address this challenge, we employ text-to-image diffusion models to produce counterfactual examples for probing intserctional social biases at scale. Our approach utilizes Stable Diffusion with cross attention control to produce sets of counterfactual image-text pairs that are highly similar in their depiction of a subject (e.g., a given occupation) while differing only in their depiction of intersectional social attributes (e.g., race & gender). We conduct extensive experiments using our generated dataset which reveal the intersectional social biases present in state-of-the-art VLMs.

Avinash Madasu, Vasudev Lal

Vision (image and video) - Language (VL) pre-training is the recent popular paradigm that achieved state-of-the-art results on multi-modal tasks like image-retrieval, video-retrieval, visual question answering etc. These models are trained in an unsupervised way and greatly benefit from the complementary modality supervision. In this paper, we explore if the language representations trained using vision supervision perform better than vanilla language representations on Natural Language Understanding and commonsense reasoning benchmarks. We experiment with a diverse set of image-text models such as ALBEF, BLIP, METER and video-text models like ALPRO, Frozen-in-Time (FiT), VIOLET. We compare the performance of language representations of stand-alone text encoders of these models to the language representations of text encoders learnt through vision supervision. Our experiments suggest that vanilla language representations show superior performance on most of the tasks. These results shed light on the current drawbacks of the vision-language models.

Avinash Madasu, Yossi Gandelsman, Vasudev Lal et al.

CLIP is one of the most popular foundational models and is heavily used for many vision-language tasks. However, little is known about the inner workings of CLIP. To bridge this gap we propose a study to quantify the interpretability in CLIP like models. We conduct this study on six different CLIP models from OpenAI and OpenCLIP which vary by size, type of pre-training data and patch size. Our approach begins with using the TEXTSPAN algorithm and in-context learning to break down individual attention heads into specific properties. We then evaluate how easily these heads can be interpreted using new metrics which measure property consistency within heads and property disentanglement across heads. Our findings reveal that larger CLIP models are generally more interpretable than their smaller counterparts. To further assist users in understanding the inner workings of CLIP models, we introduce CLIP-InterpreT, a tool designed for interpretability analysis. CLIP-InterpreT offers five types of analyses: property-based nearest neighbor search, per-head topic segmentation, contrastive segmentation, per-head nearest neighbors of an image, and per-head nearest neighbors of text.

Avinash Madasu, Anahita Bhiwandiwalla, Vasudev Lal

Foundational multimodal models pre-trained on large scale image-text pairs or video-text pairs or both have shown strong generalization abilities on downstream tasks. However unlike image-text models, pretraining video-text models is always not feasible due to the difficulty in collecting large-scale clean and aligned data, and exponential computational costs involved in the pretraining phase. Therefore, the pertinent question to ask is: Can image-text models be adapted to video tasks and is there any benefit to using these models over pretraining directly on videos? In this work, we focus on this question by proposing a detailed study on the generalization abilities of image-text models when evaluated on video understanding tasks in a zero-shot setting. We investigate 9 foundational image-text models on a diverse set of video tasks that include video action recognition (video AR), video retrieval (video RT), video question answering (video QA), video multiple choice (video MC) and video captioning (video CP). Our experiments show that image-text models exhibit impressive performance on video AR, video RT and video MC. Furthermore, they perform moderately on video captioning and poorly on video QA. These findings shed a light on the benefits of adapting foundational image-text models to an array of video tasks while avoiding the costly pretraining step.

Gadi Singer, Joscha Bach, Tetiana Grinberg et al.

While end-to-end learning systems are rapidly gaining capabilities and popularity, the increasing computational demands for deploying such systems, along with a lack of flexibility, adaptability, explainability, reasoning and verification capabilities, require new types of architectures. Here we introduce a classification of hybrid systems which, based on an analysis of human knowledge and intelligence, combines neural learning with various types of knowledge and knowledge sources. We present the Thrill-K architecture as a prototypical solution for integrating instantaneous knowledge, standby knowledge and external knowledge sources in a framework capable of inference, learning and intelligent control.

Sungduk Yu, Man Luo, Avinash Madusu et al.

Peer review is a critical process for ensuring the integrity of published scientific research. Confidence in this process is predicated on the assumption that experts in the relevant domain give careful consideration to the merits of manuscripts which are submitted for publication. With the recent rapid advancements in large language models (LLMs), a new risk to the peer review process is that negligent reviewers will rely on LLMs to perform the often time consuming process of reviewing a paper. However, there is a lack of existing resources for benchmarking the detectability of AI text in the domain of peer review. To address this deficiency, we introduce a comprehensive dataset containing a total of 788,984 AI-written peer reviews paired with corresponding human reviews, covering 8 years of papers submitted to each of two leading AI research conferences (ICLR and NeurIPS). We use this new resource to evaluate the ability of 18 existing AI text detection algorithms to distinguish between peer reviews fully written by humans and different state-of-the-art LLMs. Additionally, we explore a context-aware detection method called Anchor, which leverages manuscript content to detect AI-generated reviews, and analyze the sensitivity of detection models to LLM-assisted editing of human-written text. Our work reveals the difficulty of identifying AI-generated text at the individual peer review level, highlighting the urgent need for new tools and methods to detect this unethical use of generative AI. Our dataset is publicly available at: https://huggingface.co/datasets/IntelLabs/AI-Peer-Review-Detection-Benchmark.

Souvik Kundu, Anahita Bhiwandiwalla, Sungduk Yu et al.

Despite recent efforts in understanding the compression impact on large language models (LLMs) in terms of their downstream task performance and trustworthiness on relatively simpler uni-modal benchmarks (for example, question answering, common sense reasoning), their detailed study on multi-modal Large Vision-Language Models (LVLMs) is yet to be unveiled. Towards mitigating this gap, we present LVLM-Compress-Bench, a framework to first thoroughly study the broad impact of compression on the generative performance of LVLMs with multi-modal input driven tasks. In specific, we consider two major classes of compression for autoregressive models, namely KV cache and weight compression, for the dynamically growing intermediate cache and static weights, respectively. We use four LVLM variants of the popular LLaVA framework to present our analysis via integrating various state-of-the-art KV and weight compression methods including uniform, outlier-reduced, and group quantization for the KV cache and weights. With this framework we demonstrate on ten different multi-modal datasets with different capabilities including recognition, knowledge, language generation, spatial awareness, visual reasoning, hallucination and visual illusion identification, toxicity, stereotypes and bias. In specific, our framework demonstrates the compression impact on both general and ethically critical metrics leveraging a combination of real world and synthetic datasets to encompass diverse societal intersectional attributes. Extensive experimental evaluations yield diverse and intriguing observations on the behavior of LVLMs at different quantization budget of KV and weights, in both maintaining and losing performance as compared to the baseline model with FP16 data format. Code will be open-sourced at https://github.com/opengear-project/LVLM-compress-bench.

Estelle Aflalo, Gabriela Ben Melech Stan, Tiep Le et al.

Large Vision Language Models (LVLMs) have achieved significant progress in integrating visual and textual inputs for multimodal reasoning. However, a recurring challenge is ensuring these models utilize visual information as effectively as linguistic content when both modalities are necessary to formulate an accurate answer. We hypothesize that hallucinations arise due to the lack of effective visual grounding in current LVLMs. Furthermore, current vision-language benchmarks are not specifically measuring the degree to which the answer require the visual input. This limitation makes it challenging to confirm that the image is truly necessary, particularly in tasks like visual question answering. In this work, we introduce FiVL, a novel method for constructing datasets designed to train LVLMs for enhanced visual grounding and also evaluate their effectiveness in achieving it. We demonstrate the value of our datasets through three approaches. First, we introduce a novel training task based on our augmented training dataset, resulting in better performance than the baseline. Second, we present benchmarks to assess the model's ability to use image as substantive evidence, rather than relying solely on linguistic priors. Finally, we identify attention heads with the strongest vision-language alignment, enabling explainability on visual-driven hallucinations. The code is available at https://github.com/IntelLabs/fivl.

Zhipeng Cai, Matthias Mueller, Reiner Birkl et al.

In the current era of generative AI breakthroughs, generating panoramic scenes from a single input image remains a key challenge. Most existing methods use diffusion-based iterative or simultaneous multi-view inpainting. However, the lack of global scene layout priors leads to subpar outputs with duplicated objects (e.g., multiple beds in a bedroom) or requires time-consuming human text inputs for each view. We propose L-MAGIC, a novel method leveraging large language models for guidance while diffusing multiple coherent views of 360 degree panoramic scenes. L-MAGIC harnesses pre-trained diffusion and language models without fine-tuning, ensuring zero-shot performance. The output quality is further enhanced by super-resolution and multi-view fusion techniques. Extensive experiments demonstrate that the resulting panoramic scenes feature better scene layouts and perspective view rendering quality compared to related works, with >70% preference in human evaluations. Combined with conditional diffusion models, L-MAGIC can accept various input modalities, including but not limited to text, depth maps, sketches, and colored scripts. Applying depth estimation further enables 3D point cloud generation and dynamic scene exploration with fluid camera motion. Code is available at https://github.com/IntelLabs/MMPano. The video presentation is available at https://youtu.be/XDMNEzH4-Ec?list=PLG9Zyvu7iBa0-a7ccNLO8LjcVRAoMn57s.

Xiao Xu, Bei Li, Chenfei Wu et al.

Two-Tower Vision-Language (VL) models have shown promising improvements on various downstream VL tasks. Although the most advanced work improves performance by building bridges between encoders, it suffers from ineffective layer-by-layer utilization of uni-modal representations and cannot flexibly exploit different levels of uni-modal semantic knowledge. In this work, we propose ManagerTower, a novel VL model architecture that gathers and combines the insights of pre-trained uni-modal experts at different levels. The managers introduced in each cross-modal layer can adaptively aggregate uni-modal semantic knowledge to facilitate more comprehensive cross-modal alignment and fusion. ManagerTower outperforms previous strong baselines both with and without Vision-Language Pre-training (VLP). With only 4M VLP data, ManagerTower achieves superior performances on various downstream VL tasks, especially 79.15% accuracy on VQAv2 Test-Std, 86.56% IR@1 and 95.64% TR@1 on Flickr30K. Code and checkpoints are available at https://github.com/LooperXX/ManagerTower.

Gabriela Ben Melech Stan, Diana Wofk, Estelle Aflalo et al.

Latent diffusion models have proven to be state-of-the-art in the creation and manipulation of visual outputs. However, as far as we know, the generation of depth maps jointly with RGB is still limited. We introduce LDM3D-VR, a suite of diffusion models targeting virtual reality development that includes LDM3D-pano and LDM3D-SR. These models enable the generation of panoramic RGBD based on textual prompts and the upscaling of low-resolution inputs to high-resolution RGBD, respectively. Our models are fine-tuned from existing pretrained models on datasets containing panoramic/high-resolution RGB images, depth maps and captions. Both models are evaluated in comparison to existing related methods.

Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo et al.

One of the key shortcomings in current text-to-image (T2I) models is their inability to consistently generate images which faithfully follow the spatial relationships specified in the text prompt. In this paper, we offer a comprehensive investigation of this limitation, while also developing datasets and methods that support algorithmic solutions to improve spatial reasoning in T2I models. We find that spatial relationships are under-represented in the image descriptions found in current vision-language datasets. To alleviate this data bottleneck, we create SPRIGHT, the first spatially focused, large-scale dataset, by re-captioning 6 million images from 4 widely used vision datasets and through a 3-fold evaluation and analysis pipeline, show that SPRIGHT improves the proportion of spatial relationships in existing datasets. We show the efficacy of SPRIGHT data by showing that using only $\sim$0.25% of SPRIGHT results in a 22% improvement in generating spatially accurate images while also improving FID and CMMD scores. We also find that training on images containing a larger number of objects leads to substantial improvements in spatial consistency, including state-of-the-art results on T2I-CompBench with a spatial score of 0.2133, by fine-tuning on <500 images. Through a set of controlled experiments and ablations, we document additional findings that could support future work that seeks to understand factors that affect spatial consistency in text-to-image models.

Gabriela Ben Melech Stan, Estelle Aflalo, Raanan Yehezkel Rohekar et al.

In the rapidly evolving landscape of artificial intelligence, multi-modal large language models are emerging as a significant area of interest. These models, which combine various forms of data input, are becoming increasingly popular. However, understanding their internal mechanisms remains a complex task. Numerous advancements have been made in the field of explainability tools and mechanisms, yet there is still much to explore. In this work, we present a novel interactive application aimed towards understanding the internal mechanisms of large vision-language models. Our interface is designed to enhance the interpretability of the image patches, which are instrumental in generating an answer, and assess the efficacy of the language model in grounding its output in the image. With our application, a user can systematically investigate the model and uncover system limitations, paving the way for enhancements in system capabilities. Finally, we present a case study of how our application can aid in understanding failure mechanisms in a popular large multi-modal model: LLaVA.

Musashi Hinck, Matthew L. Olson, David Cobbley et al.

We train a suite of multimodal foundation models (MMFM) using the popular LLaVA framework with the recently released Gemma family of large language models (LLMs). Of particular interest is the 2B parameter Gemma model, which provides opportunities to construct capable small-scale MMFMs. In line with findings from other papers in this space, we test the effect of ablating three design features: pretraining the connector, utilizing a more powerful image backbone, and increasing the size of the language backbone. The resulting models, which we call LLaVA-Gemma, exhibit moderate performance on an array of evaluations, but fail to improve past the current comparably sized SOTA models. Closer analysis of performance shows mixed effects; skipping pretraining tends to reduce performance, larger vision models sometimes improve performance, and increasing language model size has inconsistent effects. We publicly release training recipes, code and weights for our models for the LLaVA-Gemma models.

Prafulla Kumar Choubey, Xin Su, Man Luo et al.

Knowledge graphs (KGs) generated by large language models (LLMs) are becoming increasingly valuable for Retrieval-Augmented Generation (RAG) applications that require knowledge-intensive reasoning. However, existing KG extraction methods predominantly rely on prompt-based approaches, which are inefficient for processing large-scale corpora. These approaches often suffer from information loss, particularly with long documents, due to the lack of specialized design for KG construction. Additionally, there is a gap in evaluation datasets and methodologies for ontology-free KG construction. To overcome these limitations, we propose SynthKG, a multi-step, document-level ontology-free KG synthesis workflow based on LLMs. By fine-tuning a smaller LLM on the synthesized document-KG pairs, we streamline the multi-step process into a single-step KG generation approach called Distill-SynthKG, substantially reducing the number of LLM inference calls. Furthermore, we re-purpose existing question-answering datasets to establish KG evaluation datasets and introduce new evaluation metrics. Using KGs produced by Distill-SynthKG, we also design a novel graph-based retrieval framework for RAG. Experimental results demonstrate that Distill-SynthKG not only surpasses all baseline models in KG quality -- including models up to eight times larger -- but also consistently excels in retrieval and question-answering tasks. Our proposed graph retrieval framework also outperforms all KG-retrieval methods across multiple benchmark datasets. We release the SynthKG dataset and Distill-SynthKG model publicly to support further research and development.

Matthew Lyle Olson, Neale Ratzlaff, Musashi Hinck et al.

Although capable of generating creative text, Large Language Models (LLMs) are poor judges of what constitutes "creativity". In this work, we show that we can leverage this knowledge of how to write creatively in order to better judge what is creative. We take a mechanistic approach that extracts differences in the internal states of an LLM when prompted to respond "boringly" or "creatively" to provide a robust measure of creativity that corresponds strongly with human judgment. We also show these internal state differences can be applied to enhance the creativity of generated text at inference time.

Neale Ratzlaff, Man Luo, Xin Su et al.

Multimodal models typically combine a powerful large language model (LLM) with a vision encoder and are then trained on multimodal data via instruction tuning. While this process adapts LLMs to multimodal settings, it remains unclear whether this adaptation compromises their original language reasoning capabilities. In this work, we explore the effects of multimodal instruction tuning on language reasoning performance. We focus on LLaVA, a leading multimodal framework that integrates LLMs such as Vicuna or Mistral with the CLIP vision encoder. We compare the performance of the original LLMs with their multimodal-adapted counterparts across eight language reasoning tasks. Our experiments yield several key insights. First, the impact of multimodal learning varies between Vicuna and Mistral: we observe a degradation in language reasoning for Mistral but improvements for Vicuna across most tasks. Second, while multimodal instruction learning consistently degrades performance on mathematical reasoning tasks (e.g., GSM8K), it enhances performance on commonsense reasoning tasks (e.g., CommonsenseQA). Finally, we demonstrate that a training-free model merging technique can effectively mitigate the language reasoning degradation observed in multimodal-adapted Mistral and even improve performance on visual tasks.

Avinash Madasu, Vasudev Lal, Phillip Howard

Vision-Language Models (VLMs) are increasingly deployed in diverse cultural contexts, yet their internal biases remain poorly understood. In this work, we propose a novel framework to systematically evaluate how VLMs encode cultural differences and biases related to race, gender, and physical traits across countries. We introduce three retrieval-based tasks: (1) Race to Country retrieval, which examines the association between individuals from specific racial groups (East Asian, White, Middle Eastern, Latino, South Asian, and Black) and different countries; (2) Personal Traits to Country retrieval, where images are paired with trait-based prompts (e.g., Smart, Honest, Criminal, Violent) to investigate potential stereotypical associations; and (3) Physical Characteristics to Country retrieval, focusing on visual attributes like skinny, young, obese, and old to explore how physical appearances are culturally linked to nations. Our findings reveal persistent biases in VLMs, highlighting how visual representations may inadvertently reinforce societal stereotypes.

Matthew Lyle Olson, Musashi Hinck, Neale Ratzlaff et al.

The ImageNet hierarchy provides a structured taxonomy of object categories, offering a valuable lens through which to analyze the representations learned by deep vision models. In this work, we conduct a comprehensive analysis of how vision models encode the ImageNet hierarchy, leveraging Sparse Autoencoders (SAEs) to probe their internal representations. SAEs have been widely used as an explanation tool for large language models (LLMs), where they enable the discovery of semantically meaningful features. Here, we extend their use to vision models to investigate whether learned representations align with the ontological structure defined by the ImageNet taxonomy. Our results show that SAEs uncover hierarchical relationships in model activations, revealing an implicit encoding of taxonomic structure. We analyze the consistency of these representations across different layers of the popular vision foundation model DINOv2 and provide insights into how deep vision models internalize hierarchical category information by increasing information in the class token through each layer. Our study establishes a framework for systematic hierarchical analysis of vision model representations and highlights the potential of SAEs as a tool for probing semantic structure in deep networks.

Gabriela Ben-Melech Stan, Estelle Aflalo, Man Luo et al.

Large Vision Language Models (LVLMs) have demonstrated remarkable reasoning capabilities over textual and visual inputs. However, these models remain prone to generating misinformation. Identifying and mitigating ungrounded responses is crucial for developing trustworthy AI. Traditional explainability methods such as gradient-based relevancy maps, offer insight into the decision process of models, but are often computationally expensive and unsuitable for real-time output validation. In this work, we introduce FastRM, an efficient method for predicting explainable Relevancy Maps of LVLMs. Furthermore, FastRM provides both quantitative and qualitative assessment of model confidence. Experimental results demonstrate that FastRM achieves a 99.8% reduction in computation time and a 44.4% reduction in memory footprint compared to traditional relevancy map generation. FastRM allows explainable AI to be more practical and scalable, thereby promoting its deployment in real-world applications and enabling users to more effectively evaluate the reliability of model outputs.

Matthew Lyle Olson, Musashi Hinck, Neale Ratzlaff et al.

Sparse Autoencoders (SAEs) have emerged as a popular tool for interpreting the hidden states of large language models (LLMs). By learning to reconstruct activations from a sparse bottleneck layer, SAEs discover interpretable features from the high-dimensional internal representations of LLMs. Despite their popularity with language models, SAEs remain understudied in the visual domain. In this work, we provide an extensive evaluation the representational power of SAEs for vision models using a broad range of image-based tasks. Our experimental results demonstrate that SAE features are semantically meaningful, improve out-of-distribution generalization, and enable controllable generation across three vision model architectures: vision embedding models, multi-modal LMMs and diffusion models. In vision embedding models, we find that learned SAE features can be used for OOD detection and provide evidence that they recover the ontological structure of the underlying model. For diffusion models, we demonstrate that SAEs enable semantic steering through text encoder manipulation and develop an automated pipeline for discovering human-interpretable attributes. Finally, we conduct exploratory experiments on multi-modal LLMs, finding evidence that SAE features reveal shared representations across vision and language modalities. Our study provides a foundation for SAE evaluation in vision models, highlighting their strong potential improving interpretability, generalization, and steerability in the visual domain.

Man Luo, David Cobbley, Xin Su et al.

Computer use agents (CUA) are systems that automatically interact with graphical user interfaces (GUIs) to complete tasks. CUA have made significant progress with the advent of large vision-language models (VLMs). However, these agents typically rely on cloud-based inference with substantial compute demands, raising critical privacy and scalability concerns, especially when operating on personal devices. In this work, we take a step toward privacy-preserving and resource-efficient agents by developing a lightweight vision-language model that runs entirely on local machines. To train this compact agent, we introduce an LLM-as-Judge framework that automatically evaluates and filters synthetic interaction trajectories, producing high-quality data for reinforcement learning without human annotation. Experiments on the OS-World benchmark demonstrate that our fine-tuned local model outperforms existing baselines, highlighting a promising path toward private, efficient, and generalizable GUI agents.

Matthew Lyle Olson, Neale Ratzlaff, Musashi Hinck et al.

In the past year, large (>100B parameter) mixture-of-expert (MoE) models have become increasingly common in the open domain. While their advantages are often framed in terms of efficiency, prior work has also explored functional differentiation through routing behavior. We investigate whether expert routing in large MoE models is influenced by the semantics of the inputs. To test this, we design two controlled experiments. First, we compare activations on sentence pairs with a shared target word used in the same or different senses. Second, we fix context and substitute the target word with semantically similar or dissimilar alternatives. Comparing expert overlap across these conditions reveals clear, statistically significant evidence of semantic routing in large MoE models.

Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck et al.

Large Multi-Modal Models (LMMs) have demonstrated impressive capabilities as general-purpose chatbots able to engage in conversations about visual inputs. However, their responses are influenced by societal biases present in their training datasets, leading to undesirable differences in how the model responds when presented with images depicting people of different demographics. In this work, we propose a training-free debiasing framework for LMMs that intervenes on the model's representations during text generation by constructing a steering vector that reduces reference on protected attributes. Our framework introduces two complementary methods: (1) a dataset-based approach that constructs a steering vector by contrasting model activations on biased and neutral inputs, and (2) a novel optimization-based approach designed for low-resource settings, which constructs the steering vector using a single step of gradient-based perturbation without requiring additional data. Our experiments show that these interventions effectively reduce the propensity of LMMs to generate text related to protected attributes while maintaining sentiment and fluency. Furthermore, we demonstrate that debiased LMMs achieve comparable accuracy to their unmodified counterparts on downstream tasks, indicating that bias mitigation can be achieved without sacrificing model performance.

Neale Ratzlaff, Matthew Lyle Olson, Musashi Hinck et al.

Large Vision Language Models (LVLMs) such as LLaVA have demonstrated impressive capabilities as general-purpose chatbots that can engage in conversations about a provided input image. However, their responses are influenced by societal biases present in their training datasets, leading to undesirable differences in how the model responds when presented with images depicting people of different demographics. In this work, we propose a novel debiasing framework for LVLMs by directly ablating biased attributes during text generation to avoid generating text related to protected attributes, or even representing them internally. Our method requires no training and a relatively small amount of representative biased outputs (~1000 samples). Our experiments show that not only can we can minimize the propensity of LVLMs to generate text related to protected attributes, but we can even use synthetic data to inform the ablation while retaining captioning performance on real data such as COCO. Furthermore, we find the resulting generations from a debiased LVLM exhibit similar accuracy as a baseline biased model, showing that debiasing effects can be achieved without sacrificing model performance.

Gabriela Ben Melech Stan, Estelle Aflalo, Avinash Madasu et al.

Training models on synthetic data has emerged as an increasingly important strategy for improving the performance of generative AI. This approach is particularly helpful for large multimodal models (LMMs) due to the relative scarcity of high-quality paired image-text data compared to language-only data. While a variety of methods have been proposed for generating large multimodal datasets, they do not tailor the synthetic data to address specific deficiencies in the reasoning abilities of LMMs which will be trained with the generated dataset. In contrast, humans often learn in a more efficient manner by seeking out examples related to the types of reasoning where they have failed previously. Inspired by this observation, we propose a new approach for synthetic data generation which is grounded in the analysis of an existing LMM's reasoning failures. Our methodology leverages frontier models to automatically analyze errors produced by a weaker LMM and propose new examples which can be used to correct the reasoning failure via additional training, which are then further filtered to ensure high quality. We generate a large multimodal instruction tuning dataset containing over 553k examples using our approach and conduct extensive experiments demonstrating its utility for improving the performance of LMMs on multiple downstream tasks. Our results show that models trained on our synthetic data can even exceed the performance of LMMs trained on an equivalent amount of additional real data, demonstrating the high value of generating synthetic data targeted to specific reasoning failure modes in LMMs. We will make our dataset and code publicly available.

Avinash Madasu, Vasudev Lal, Phillip Howard

CLIP is one of the most popular foundation models and is heavily used for many vision-language tasks, yet little is known about its inner workings. As CLIP is increasingly deployed in real-world applications, it is becoming even more critical to understand its limitations and embedded social biases to mitigate potentially harmful downstream consequences. However, the question of what internal mechanisms drive both the impressive capabilities as well as problematic shortcomings of CLIP has largely remained unanswered. To bridge this gap, we study the conceptual consistency of text descriptions for attention heads in CLIP-like models. Specifically, we propose Concept Consistency Score (CCS), a novel interpretability metric that measures how consistently individual attention heads in CLIP models align with specific concepts. Our soft-pruning experiments reveal that high CCS heads are critical for preserving model performance, as pruning them leads to a significantly larger performance drop than pruning random or low CCS heads. Notably, we find that high CCS heads capture essential concepts and play a key role in out-of-domain detection, concept-specific reasoning, and video-language understanding. Moreover, we prove that high CCS heads learn spurious correlations which amplify social biases. These results position CCS as a powerful interpretability metric exposing the paradox of performance and social biases in CLIP models.

Raanan Y. Rohekar, Yaniv Gurwicz, Sungduk Yu et al.

Are generative pre-trained transformer (GPT) models, trained only to predict the next token, implicitly learning a world model from which sequences are generated one token at a time? We address this question by deriving a causal interpretation of the attention mechanism in GPT and presenting a causal world model that arises from this interpretation. Furthermore, we propose that GPT models, at inference time, can be utilized for zero-shot causal structure learning for input sequences, and introduce a corresponding confidence score. Empirical tests were conducted in controlled environments using the setups of the Othello and Chess strategy games. A GPT, pre-trained on real-world games played with the intention of winning, was tested on out-of-distribution synthetic data consisting of sequences of random legal moves. We find that the GPT model is likely to generate legal next moves for out-of-distribution sequences for which a causal structure is encoded in the attention mechanism with high confidence. In cases where it generates illegal moves, it also fails to capture a causal structure.

Xin Su, Man Luo, Kris W Pan et al.

Multimodal retrieval augmented generation (RAG) plays a crucial role in domains such as knowledge-based visual question answering (KB-VQA), where external knowledge is needed to answer a question. However, existing multimodal LLMs (MLLMs) are not designed for context-augmented generation, limiting their effectiveness in such tasks. While synthetic data generation has recently gained attention for training MLLMs, its application for context-augmented generation remains underexplored. To address this gap, we introduce SK-VQA, a large-scale synthetic multimodal dataset containing over 2 million visual question-answer pairs, each associated with context documents containing information necessary to determine the final answer. Compared to previous datasets, SK-VQA contains 11x more unique questions, exhibits greater domain diversity, and covers a broader spectrum of image sources. Through human evaluations, we confirm the high quality of the generated question-answer pairs and their contextual relevance. Extensive experiments show that SK-VQA serves both as a challenging KB-VQA benchmark and as an effective training resource for adapting MLLMs to context-augmented generation. Our results further indicate that models trained on SK-VQA demonstrate enhanced generalization in both context-aware VQA and multimodal RAG settings. SK-VQA is publicly available via Hugging Face Hub.

Jerry Tang, Meng Du, Vy A. Vo et al.

Encoding models have been used to assess how the human brain represents concepts in language and vision. While language and vision rely on similar concept representations, current encoding models are typically trained and tested on brain responses to each modality in isolation. Recent advances in multimodal pretraining have produced transformers that can extract aligned representations of concepts in language and vision. In this work, we used representations from multimodal transformers to train encoding models that can transfer across fMRI responses to stories and movies. We found that encoding models trained on brain responses to one modality can successfully predict brain responses to the other modality, particularly in cortical regions that represent conceptual meaning. Further analysis of these encoding models revealed shared semantic dimensions that underlie concept representations in language and vision. Comparing encoding models trained using representations from multimodal and unimodal transformers, we found that multimodal transformers learn more aligned representations of concepts in language and vision. Our results demonstrate how multimodal transformers can provide insights into the brain's capacity for multimodal processing.

Gabriela Ben Melech Stan, Diana Wofk, Scottie Fox et al.

This research paper proposes a Latent Diffusion Model for 3D (LDM3D) that generates both image and depth map data from a given text prompt, allowing users to generate RGBD images from text prompts. The LDM3D model is fine-tuned on a dataset of tuples containing an RGB image, depth map and caption, and validated through extensive experiments. We also develop an application called DepthFusion, which uses the generated RGB images and depth maps to create immersive and interactive 360-degree-view experiences using TouchDesigner. This technology has the potential to transform a wide range of industries, from entertainment and gaming to architecture and design. Overall, this paper presents a significant contribution to the field of generative AI and computer vision, and showcases the potential of LDM3D and DepthFusion to revolutionize content creation and digital experiences. A short video summarizing the approach can be found at https://t.ly/tdi2.

Phillip Howard, Junlin Wang, Vasudev Lal et al.

Comparative knowledge (e.g., steel is stronger and heavier than styrofoam) is an essential component of our world knowledge, yet understudied in prior literature. In this paper, we harvest the dramatic improvements in knowledge capabilities of language models into a large-scale comparative knowledge base. While the ease of acquisition of such comparative knowledge is much higher from extreme-scale models like GPT-4, compared to their considerably smaller and weaker counterparts such as GPT-2, not even the most powerful models are exempt from making errors. We thus ask: to what extent are models at different scales able to generate valid and diverse comparative knowledge? We introduce NeuroComparatives, a novel framework for comparative knowledge distillation overgenerated from language models such as GPT-variants and LLaMA, followed by stringent filtering of the generated knowledge. Our framework acquires comparative knowledge between everyday objects, producing a corpus of up to 8.8M comparisons over 1.74M entity pairs - 10X larger and 30% more diverse than existing resources. Moreover, human evaluations show that NeuroComparatives outperform existing resources in terms of validity (up to 32% absolute improvement). Our acquired NeuroComparatives leads to performance improvements on five downstream tasks. We find that neuro-symbolic manipulation of smaller models offers complementary benefits to the currently dominant practice of prompting extreme-scale language models for knowledge distillation.

Estelle Aflalo, Meng Du, Shao-Yen Tseng et al.

Breakthroughs in transformer-based models have revolutionized not only the NLP field, but also vision and multimodal systems. However, although visualization and interpretability tools have become available for NLP models, internal mechanisms of vision and multimodal transformers remain largely opaque. With the success of these transformers, it is increasingly critical to understand their inner workings, as unraveling these black-boxes will lead to more capable and trustworthy models. To contribute to this quest, we propose VL-InterpreT, which provides novel interactive visualizations for interpreting the attentions and hidden representations in multimodal transformers. VL-InterpreT is a task agnostic and integrated tool that (1) tracks a variety of statistics in attention heads throughout all layers for both vision and language components, (2) visualizes cross-modal and intra-modal attentions through easily readable heatmaps, and (3) plots the hidden representations of vision and language tokens as they pass through the transformer layers. In this paper, we demonstrate the functionalities of VL-InterpreT through the analysis of KD-VLP, an end-to-end pretraining vision-language multimodal transformer-based model, in the tasks of Visual Commonsense Reasoning (VCR) and WebQA, two visual question answering benchmarks. Furthermore, we also present a few interesting findings about multimodal transformer behaviors that were learned through our tool.

Yongfei Liu, Chenfei Wu, Shao-yen Tseng et al.

Self-supervised vision-and-language pretraining (VLP) aims to learn transferable multi-modal representations from large-scale image-text data and to achieve strong performances on a broad scope of vision-language tasks after finetuning. Previous mainstream VLP approaches typically adopt a two-step strategy relying on external object detectors to encode images in a multi-modal Transformer framework, which suffer from restrictive object concept space, limited image context and inefficient computation. In this paper, we propose an object-aware end-to-end VLP framework, which directly feeds image grid features from CNNs into the Transformer and learns the multi-modal representations jointly. More importantly, we propose to perform object knowledge distillation to facilitate learning cross-modal alignment at different semantic levels. To achieve that, we design two novel pretext tasks by taking object features and their semantic labels from external detectors as supervision: 1.) Object-guided masked vision modeling task focuses on enforcing object-aware representation learning in the multi-modal Transformer; 2.) Phrase-region alignment task aims to improve cross-modal alignment by utilizing the similarities between noun phrases and object labels in the linguistic space. Extensive experiments on a wide range of vision-language tasks demonstrate the efficacy of our proposed framework, and we achieve competitive or superior performances over the existing pretraining strategies.