Roei Herzig

Chancharik Mitra, Brandon Huang, Trevor Darrell et al.

The combination of strong visual backbones and Large Language Model (LLM) reasoning has led to Large Multimodal Models (LMMs) becoming the current standard for a wide range of vision and language (VL) tasks. However, recent research has shown that even the most advanced LMMs still struggle to capture aspects of compositional visual reasoning, such as attributes and relationships between objects. One solution is to utilize scene graphs (SGs)--a formalization of objects and their relations and attributes that has been extensively used as a bridge between the visual and textual domains. Yet, scene graph data requires scene graph annotations, which are expensive to collect and thus not easily scalable. Moreover, finetuning an LMM based on SG data can lead to catastrophic forgetting of the pretraining objective. To overcome this, inspired by chain-of-thought methods, we propose Compositional Chain-of-Thought (CCoT), a novel zero-shot Chain-of-Thought prompting method that utilizes SG representations in order to extract compositional knowledge from an LMM. Specifically, we first generate an SG using the LMM, and then use that SG in the prompt to produce a response. Through extensive experiments, we find that the proposed CCoT approach not only improves LMM performance on several vision and language VL compositional benchmarks but also improves the performance of several popular LMMs on general multimodal benchmarks, without the need for fine-tuning or annotated ground-truth SGs. Code: https://github.com/chancharikmitra/CCoT

Jovana Kondic, Pengyuan Li, Dhiraj Joshi et al. · ibm-research

Understanding charts requires models to jointly reason over geometric visual patterns, structured numerical data, and natural language -- a capability where current vision-language models (VLMs) remain limited. We introduce ChartNet, a high-quality, million-scale multimodal dataset designed to advance chart interpretation and reasoning. ChartNet leverages a novel code-guided synthesis pipeline to generate 1.5 million diverse chart samples spanning 24 chart types and 6 plotting libraries. Each sample consists of five aligned components: plotting code, rendered chart image, data table, natural language summary, and question-answering with reasoning, providing fine-grained cross-modal alignment. To capture the full spectrum of chart comprehension, ChartNet additionally includes specialized subsets encompassing human annotated data, real-world data, safety, and grounding. Moreover, a rigorous quality-filtering pipeline ensures visual fidelity, semantic accuracy, and diversity across chart representations. Fine-tuning on ChartNet consistently improves results across benchmarks, demonstrating its utility as large-scale supervision for multimodal models. As the largest open-source dataset of its kind, ChartNet aims to support the development of foundation models with robust and generalizable capabilities for data visualization understanding. The dataset is publicly available at https://huggingface.co/datasets/ibm-granite/ChartNet

Sivan Doveh, Assaf Arbelle, Sivan Harary et al.

Vision and Language (VL) models have demonstrated remarkable zero-shot performance in a variety of tasks. However, some aspects of complex language understanding still remain a challenge. We introduce the collective notion of Structured Vision&Language Concepts (SVLC) which includes object attributes, relations, and states which are present in the text and visible in the image. Recent studies have shown that even the best VL models struggle with SVLC. A possible way of fixing this issue is by collecting dedicated datasets for teaching each SVLC type, yet this might be expensive and time-consuming. Instead, we propose a more elegant data-driven approach for enhancing VL models' understanding of SVLCs that makes more effective use of existing VL pre-training datasets and does not require any additional data. While automatic understanding of image structure still remains largely unsolved, language structure is much better modeled and understood, allowing for its effective utilization in teaching VL models. In this paper, we propose various techniques based on language structure understanding that can be used to manipulate the textual part of off-the-shelf paired VL datasets. VL models trained with the updated data exhibit a significant improvement of up to 15% in their SVLC understanding with only a mild degradation in their zero-shot capabilities both when training from scratch or fine-tuning a pre-trained model.

Roei Herzig, Ofir Abramovich, Elad Ben-Avraham et al.

Action recognition models have achieved impressive results by incorporating scene-level annotations, such as objects, their relations, 3D structure, and more. However, obtaining annotations of scene structure for videos requires a significant amount of effort to gather and annotate, making these methods expensive to train. In contrast, synthetic datasets generated by graphics engines provide powerful alternatives for generating scene-level annotations across multiple tasks. In this work, we propose an approach to leverage synthetic scene data for improving video understanding. We present a multi-task prompt learning approach for video transformers, where a shared video transformer backbone is enhanced by a small set of specialized parameters for each task. Specifically, we add a set of "task prompts", each corresponding to a different task, and let each prompt predict task-related annotations. This design allows the model to capture information shared among synthetic scene tasks as well as information shared between synthetic scene tasks and a real video downstream task throughout the entire network. We refer to this approach as "Promptonomy", since the prompts model task-related structure. We propose the PromptonomyViT model (PViT), a video transformer that incorporates various types of scene-level information from synthetic data using the "Promptonomy" approach. PViT shows strong performance improvements on multiple video understanding tasks and datasets. Project page: \url{https://ofir1080.github.io/PromptonomyViT}

Amit Alfassy, Assaf Arbelle, Oshri Halimi et al.

Foundation Models (FMs) have demonstrated unprecedented capabilities including zero-shot learning, high fidelity data synthesis, and out of domain generalization. However, as we show in this paper, FMs still have poor out-of-the-box performance on expert tasks (e.g. retrieval of car manuals technical illustrations from language queries), data for which is either unseen or belonging to a long-tail part of the data distribution of the huge datasets used for FM pre-training. This underlines the necessity to explicitly evaluate and finetune FMs on such expert tasks, arguably ones that appear the most in practical real-world applications. In this paper, we propose a first of its kind FETA benchmark built around the task of teaching FMs to understand technical documentation, via learning to match their graphical illustrations to corresponding language descriptions. Our FETA benchmark focuses on text-to-image and image-to-text retrieval in public car manuals and sales catalogue brochures. FETA is equipped with a procedure for completely automatic annotation extraction (code would be released upon acceptance), allowing easy extension of FETA to more documentation types and application domains in the future. Our automatic annotation leads to an automated performance metric shown to be consistent with metrics computed on human-curated annotations (also released). We provide multiple baselines and analysis of popular FMs on FETA leading to several interesting findings that we believe would be very valuable to the FM community, paving the way towards real-world application of FMs for practical expert tasks currently 'overlooked' by standard benchmarks focusing on common objects.

Elad Ben-Avraham, Roei Herzig, Karttikeya Mangalam et al.

This technical report describes the SViT approach for the Ego4D Point of No Return (PNR) Temporal Localization Challenge. We propose a learning framework StructureViT (SViT for short), which demonstrates how utilizing the structure of a small number of images only available during training can improve a video model. SViT relies on two key insights. First, as both images and videos contain structured information, we enrich a transformer model with a set of \emph{object tokens} that can be used across images and videos. Second, the scene representations of individual frames in video should "align" with those of still images. This is achieved via a "Frame-Clip Consistency" loss, which ensures the flow of structured information between images and videos. SViT obtains strong performance on the challenge test set with 0.656 absolute temporal localization error.

Dantong Niu, Amir Bar, Roei Herzig et al.

Existing video-based action recognition systems typically require dense annotation and struggle in environments when there is significant distribution shift relative to the training data. Current methods for video domain adaptation typically fine-tune the model using fully annotated data on a subset of target domain data or align the representation of the two domains using bootstrapping or adversarial learning. Inspired by the pivotal role of objects in recent supervised object-centric action recognition models, we present Object-based (yet Class-agnostic) Video Domain Adaptation (ODAPT), a simple yet effective framework for adapting the existing action recognition systems to new domains by utilizing a sparse set of frames with class-agnostic object annotations in a target domain. Our model achieves a +6.5 increase when adapting across kitchens in Epic-Kitchens and a +3.1 increase adapting between Epic-Kitchens and the EGTEA dataset. ODAPT is a general framework that can also be combined with previous unsupervised methods, offering a +5.0 boost when combined with the self-supervised multi-modal method MMSADA and a +1.7 boost when added to the adversarial-based method TA$^3$N on Epic-Kitchens.

Elad Ben-Avraham, Roei Herzig, Karttikeya Mangalam et al.

Recent action recognition models have achieved impressive results by integrating objects, their locations and interactions. However, obtaining dense structured annotations for each frame is tedious and time-consuming, making these methods expensive to train and less scalable. At the same time, if a small set of annotated images is available, either within or outside the domain of interest, how could we leverage these for a video downstream task? We propose a learning framework StructureViT (SViT for short), which demonstrates how utilizing the structure of a small number of images only available during training can improve a video model. SViT relies on two key insights. First, as both images and videos contain structured information, we enrich a transformer model with a set of \emph{object tokens} that can be used across images and videos. Second, the scene representations of individual frames in video should "align" with those of still images. This is achieved via a \emph{Frame-Clip Consistency} loss, which ensures the flow of structured information between images and videos. We explore a particular instantiation of scene structure, namely a \emph{Hand-Object Graph}, consisting of hands and objects with their locations as nodes, and physical relations of contact/no-contact as edges. SViT shows strong performance improvements on multiple video understanding tasks and datasets. Furthermore, it won in the Ego4D CVPR'22 Object State Localization challenge. For code and pretrained models, visit the project page at \url{https://eladb3.github.io/SViT/}

James Ni, Zekai Wang, Wei Lin et al.

Video generation models are rapidly improving in their ability to synthesize human actions in novel contexts, holding the potential to serve as high-level planners for contextual robot control. To realize this potential, a key research question remains open: how can a humanoid execute the human actions from generated videos in a zero-shot manner? This challenge arises because generated videos are often noisy and exhibit morphological distortions that make direct imitation difficult compared to real video. To address this, we introduce a two-stage pipeline. First, we lift video pixels into a 4D human representation and then retarget to the humanoid morphology. Second, we propose GenMimic-a physics-aware reinforcement learning policy conditioned on 3D keypoints, and trained with symmetry regularization and keypoint-weighted tracking rewards. As a result, GenMimic can mimic human actions from noisy, generated videos. We curate GenMimicBench, a synthetic human-motion dataset generated using two video generation models across a spectrum of actions and contexts, establishing a benchmark for assessing zero-shot generalization and policy robustness. Extensive experiments demonstrate improvements over strong baselines in simulation and confirm coherent, physically stable motion tracking on a Unitree G1 humanoid robot without fine-tuning. This work offers a promising path to realizing the potential of video generation models as high-level policies for robot control.

Kelvin Li, Chuyi Shang, Leonid Karlinsky et al.

While Large Multimodal Models (LMMs) have made significant progress, they remain largely text-centric, relying on language as their core reasoning modality. As a result, they are limited in their ability to handle reasoning tasks that are predominantly visual. Recent approaches have sought to address this by supervising intermediate visual steps with helper images, depth maps, or image crops. However, these strategies impose restrictive priors on what "useful" visual abstractions look like, add heavy annotation costs, and struggle to generalize across tasks. To address this critical limitation, we propose a task-agnostic mechanism that trains LMMs to discover and use visual reasoning tokens without explicit supervision. These tokens attend globally and re-encode the image in a task-adaptive way, enabling the model to extract relevant visual information without hand-crafted supervision. Our approach outperforms direct fine-tuning and achieves state-of-the-art results on a diverse range of vision-centric tasks -- including those where intermediate abstractions are hard to specify -- while also generalizing to multi-task instruction tuning.

Brandon Huang, Hang Hua, Zhuoran Yu et al.

While Vision-language models (VLMs) have demonstrated remarkable performance across multi-modal tasks, their choice of vision encoders presents a fundamental weakness: their low-level features lack the robust structural and spatial information essential for document understanding and web agents. To bridge this gap, we introduce DAVE, a vision encoder purpose-built for VLMs and tailored for these tasks. Our training pipeline is designed to leverage abundant unlabeled data to bypass the need for costly large-scale annotations for document and web images. We begin with a self-supervised pretraining stage on unlabeled images, followed by a supervised autoregressive pretraining stage, where the model learns tasks like parsing and localization from limited, high-quality data. Within the supervised stage, we adopt two strategies to improve our encoder's alignment with both general visual knowledge and diverse document and web agentic tasks: (i) We introduce a novel model-merging scheme, combining encoders trained with different text decoders to ensure broad compatibility with different web agentic architectures. (ii) We use ensemble training to fuse features from pretrained generalist encoders (e.g., SigLIP2) with our own document and web-specific representations. Extensive experiments on classic document tasks, VQAs, web localization, and agent-based benchmarks validate the effectiveness of our approach, establishing DAVE as a strong vision encoder for document and web applications.

Chuyi Shang, Amos You, Sanjay Subramanian et al.

Recently, image-based Large Multimodal Models (LMMs) have made significant progress in video question-answering (VideoQA) using a frame-wise approach by leveraging large-scale pretraining in a zero-shot manner. Nevertheless, these models need to be capable of finding relevant information, extracting it, and answering the question simultaneously. Currently, existing methods perform all of these steps in a single pass without being able to adapt if insufficient or incorrect information is collected. To overcome this, we introduce a modular multi-LMM agent framework based on several agents with different roles, instructed by a Planner agent that updates its instructions using shared feedback from the other agents. Specifically, we propose TraveLER, a method that can create a plan to "Traverse" through the video, ask questions about individual frames to "Locate" and store key information, and then "Evaluate" if there is enough information to answer the question. Finally, if there is not enough information, our method is able to "Replan" based on its collected knowledge. Through extensive experiments, we find that the proposed TraveLER approach improves performance on several VideoQA benchmarks without the need to fine-tune on specific datasets. Our code is available at https://github.com/traveler-framework/TraveLER.

Granite Vision Team, Leonid Karlinsky, Assaf Arbelle et al.

We introduce Granite Vision, a lightweight large language model with vision capabilities, specifically designed to excel in enterprise use cases, particularly in visual document understanding. Our model is trained on a comprehensive instruction-following dataset, including document-related tasks, such as content extraction from tables, charts, diagrams, sketches, and infographics, as well as general image tasks. The architecture of Granite Vision is centered around visual modality alignment with a decoder-only, 2 billion parameter Granite large language model. Additionally, we introduce a dedicated safety classification approach in test-time that leverages a sparse set of attention vectors to identify potential harmful inputs. Despite its lightweight architecture, Granite Vision achieves strong results in standard benchmarks related to visual document understanding, as well as on the LiveXiv benchmark, which is designed to avoid test set contamination by using a constantly updated corpus of recently published Arxiv papers. We are releasing the model under the Apache-2 license, allowing for both research and commercial use, while offering complete visibility into the training data and other relevant details. See https://huggingface.co/ibm-granite/ for model weights.

Zineng Tang, Long Lian, Seun Eisape et al.

Despite the recent success of image-text contrastive models like CLIP and SigLIP, these models often struggle with vision-centric tasks that demand high-fidelity image understanding, such as counting, depth estimation, and fine-grained object recognition. These models, by performing language alignment, tend to prioritize high-level semantics over visual understanding, weakening their image understanding. On the other hand, vision-focused models are great at processing visual information but struggle to understand language, limiting their flexibility for language-driven tasks. In this work, we introduce TULIP, an open-source, drop-in replacement for existing CLIP-like models. Our method leverages generative data augmentation, enhanced image-image and text-text contrastive learning, and image/text reconstruction regularization to learn fine-grained visual features while preserving global semantic alignment. Our approach, scaling to over 1B parameters, outperforms existing state-of-the-art (SOTA) models across multiple benchmarks, establishing a new SOTA zero-shot performance on ImageNet-1K, delivering up to a $2\times$ enhancement over SigLIP on RxRx1 in linear probing for few-shot classification, and improving vision-language models, achieving over $3\times$ higher scores than SigLIP on MMVP. Our code/checkpoints are available at https://tulip-berkeley.github.io

Brandon Huang, Chancharik Mitra, Assaf Arbelle et al.

The recent success of interleaved Large Multimodal Models (LMMs) in few-shot learning suggests that in-context learning (ICL) with many examples can be promising for learning new tasks. However, this many-shot multimodal ICL setting has one crucial problem: it is fundamentally limited by the model's context length set at pretraining. The problem is especially prominent in the multimodal domain, which processes both text and images, requiring additional tokens. This motivates the need for a multimodal method to compress many shots into fewer tokens without finetuning. In this work, we enable LMMs to perform multimodal, many-shot in-context learning by leveraging Multimodal Task Vectors (MTV) -- compact implicit representations of in-context examples compressed in the model's attention heads. Specifically, we first demonstrate the existence of such MTV in LMMs and then leverage these extracted MTV to enable many-shot in-context learning for various vision-and-language tasks. Our experiments suggest that MTV can scale in performance with the number of compressed shots and generalize to similar out-of-domain tasks without additional context length for inference. Code: https://github.com/Brandon3964/MultiModal-Task-Vector

Eran Goldman, Roei Herzig, Aviv Eisenschtat et al.

Man-made scenes can be densely packed, containing numerous objects, often identical, positioned in close proximity. We show that precise object detection in such scenes remains a challenging frontier even for state-of-the-art object detectors. We propose a novel, deep-learning based method for precise object detection, designed for such challenging settings. Our contributions include: (1) A layer for estimating the Jaccard index as a detection quality score; (2) a novel EM merging unit, which uses our quality scores to resolve detection overlap ambiguities; finally, (3) an extensive, annotated data set, SKU-110K, representing packed retail environments, released for training and testing under such extreme settings. Detection tests on SKU-110K and counting tests on the CARPK and PUCPR+ show our method to outperform existing state-of-the-art with substantial margins. The code and data will be made available on \url{www.github.com/eg4000/SKU110K_CVPR19}.

Dantong Niu, Xudong Wang, Xinyang Han et al.

Several unsupervised image segmentation approaches have been proposed which eliminate the need for dense manually-annotated segmentation masks; current models separately handle either semantic segmentation (e.g., STEGO) or class-agnostic instance segmentation (e.g., CutLER), but not both (i.e., panoptic segmentation). We propose an Unsupervised Universal Segmentation model (U2Seg) adept at performing various image segmentation tasks -- instance, semantic and panoptic -- using a novel unified framework. U2Seg generates pseudo semantic labels for these segmentation tasks via leveraging self-supervised models followed by clustering; each cluster represents different semantic and/or instance membership of pixels. We then self-train the model on these pseudo semantic labels, yielding substantial performance gains over specialized methods tailored to each task: a +2.6 AP$^{\text{box}}$ boost vs. CutLER in unsupervised instance segmentation on COCO and a +7.0 PixelAcc increase (vs. STEGO) in unsupervised semantic segmentation on COCOStuff. Moreover, our method sets up a new baseline for unsupervised panoptic segmentation, which has not been previously explored. U2Seg is also a strong pretrained model for few-shot segmentation, surpassing CutLER by +5.0 AP$^{\text{mask}}$ when trained on a low-data regime, e.g., only 1% COCO labels. We hope our simple yet effective method can inspire more research on unsupervised universal image segmentation.

Dantong Niu, Yuvan Sharma, Haoru Xue et al.

Foundation models pre-trained on massive unlabeled datasets have revolutionized natural language and computer vision, exhibiting remarkable generalization capabilities, thus highlighting the importance of pre-training. Yet, efforts in robotics have struggled to achieve similar success, limited by either the need for costly robotic annotations or the lack of representations that effectively model the physical world. In this paper, we introduce ARM4R, an Auto-regressive Robotic Model that leverages low-level 4D Representations learned from human video data to yield a better pre-trained robotic model. Specifically, we focus on utilizing 3D point tracking representations from videos derived by lifting 2D representations into 3D space via monocular depth estimation across time. These 4D representations maintain a shared geometric structure between the points and robot state representations up to a linear transformation, enabling efficient transfer learning from human video data to low-level robotic control. Our experiments show that ARM4R can transfer efficiently from human video data to robotics and consistently improves performance on tasks across various robot environments and configurations.

Yida Yin, Zekai Wang, Yuvan Sharma et al.

Recently, Large Language Models (LLMs) have achieved remarkable success using in-context learning (ICL) in the language domain. However, leveraging the ICL capabilities within LLMs to directly predict robot actions remains largely unexplored. In this paper, we introduce RoboPrompt, a framework that enables off-the-shelf text-only LLMs to directly predict robot actions through ICL without training. Our approach first heuristically identifies keyframes that capture important moments from an episode. Next, we extract end-effector actions from these keyframes as well as the estimated initial object poses, and both are converted into textual descriptions. Finally, we construct a structured template to form ICL demonstrations from these textual descriptions and a task instruction. This enables an LLM to directly predict robot actions at test time. Through extensive experiments and analysis, RoboPrompt shows stronger performance over zero-shot and ICL baselines in simulated and real-world settings. Our project page is available at https://davidyyd.github.io/roboprompt.

Jiaxin Ge, Sanjay Subramanian, Baifeng Shi et al.

Visual Programming (VP) has emerged as a powerful framework for Visual Question Answering (VQA). By generating and executing bespoke code for each question, these methods demonstrate impressive compositional and reasoning capabilities, especially in few-shot and zero-shot scenarios. However, existing VP methods generate all code in a single function, resulting in code that is suboptimal in terms of both accuracy and interpretability. Inspired by human coding practices, we propose Recursive Visual Programming (RVP), which simplifies generated routines, provides more efficient problem solving, and can manage more complex data structures. RVP is inspired by human coding practices and approaches VQA tasks with an iterative recursive code generation approach, allowing decomposition of complicated problems into smaller parts. Notably, RVP is capable of dynamic type assignment, i.e., as the system recursively generates a new piece of code, it autonomously determines the appropriate return type and crafts the requisite code to generate that output. We show RVP's efficacy through extensive experiments on benchmarks including VSR, COVR, GQA, and NextQA, underscoring the value of adopting human-like recursive and modular programming techniques for solving VQA tasks through coding.

Nasim Borazjanizadeh, Roei Herzig, Eduard Oks et al.

Human reasoning relies on constructing and manipulating mental models -- simplified internal representations of situations used to understand and solve problems. Conceptual diagrams (e.g., a sketch drawn to aid reasoning) externalize these mental models, abstracting irrelevant details to efficiently capture how entities interact. In contrast, Large Language Models (LLMs) and Large MultiModal Models (LMMs) predominantly reason through text, limiting their effectiveness on complex multi-step tasks. In this paper, we propose Visual Thinking, a generalizable framework that enables LMMs to reason through multiple chains of self-generated conceptual diagrams, significantly enhancing their combinatorial planning capabilities. Our approach requires no human input beyond the natural language description of the task. It integrates textual and diagrammatic reasoning within an optimized Graph-of-Thought inference framework, enhanced by beam search and depth-wise backtracking. Evaluated on multiple challenging PDDL planning domains, our method substantially improves LMM performance (e.g., GPT-4o: 35.5% -> 90.2% in Blocksworld) and consistently outperforms text-only search-based inference methods. On more difficult domains with solution depths up to 40, it also surpasses the o1-preview reasoning model (e.g., 16 percentage points improvement in Floor Tiles). These results demonstrate the power of conceptual diagrams as a reasoning medium in LMMs.

Chancharik Mitra, Brandon Huang, Tianning Chai et al.

Generative Large Multimodal Models (LMMs) like LLaVA and Qwen-VL excel at a wide variety of vision-language (VL) tasks. Despite strong performance, LMMs' generative outputs are not specialized for vision-language classification tasks (i.e., tasks with vision-language inputs and discrete labels) such as image classification and multiple-choice VQA. One key challenge in utilizing LMMs for these tasks is the extraction of useful features from generative LMMs. To overcome this, we propose an approach that leverages multimodal feature extraction from the LMM's latent space. Toward this end, we present Sparse Attention Vectors (SAVs) -- a finetuning-free method that leverages sparse attention head activations (fewer than 5% of the heads) in LMMs as strong feature representations. With only few-shot examples, SAVs demonstrate state-of-the-art performance compared to a variety of few-shot and finetuned baselines on a collection of vision-language classification tasks. Our experiments also imply that SAVs can scale in performance with additional examples and generalize to similar tasks, establishing SAVs as both effective and robust multimodal feature representations.

Wen-Han Hsieh, Elvis Hsieh, Dantong Niu et al.

Recently, Vision-Language-Action (VLA) models have demonstrated strong performance on a range of robotic tasks. These models rely on multimodal inputs, with language instructions playing a crucial role -- not only in predicting actions, but also in robustly interpreting user intent, even when the requests are impossible to fulfill. In this work, we investigate how VLAs can recognize, interpret, and respond to false-premise instructions: natural language commands that reference objects or conditions absent from the environment. We propose Instruct-Verify-and-Act (IVA), a unified framework that (i) detects when an instruction cannot be executed due to a false premise, (ii) engages in language-based clarification or correction, and (iii) grounds plausible alternatives in perception and action. Towards this end, we construct a large-scale instruction tuning setup with structured language prompts and train a VLA model capable of handling both accurate and erroneous requests. Our approach leverages a contextually augmented, semi-synthetic dataset containing paired positive and false-premise instructions, enabling robust detection and natural language correction. Our experiments show that IVA improves false premise detection accuracy by 97.56% over baselines, while increasing successful responses in false-premise scenarios by 50.78%.

Chancharik Mitra, Yusen Luo, Raj Saravanan et al.

Vision-Language Action (VLAs) models promise to extend the remarkable success of vision-language models (VLMs) to robotics. Yet, unlike VLMs in the vision-language domain, VLAs for robotics require finetuning to contend with varying physical factors like robot embodiment, environment characteristics, and spatial relationships of each task. Existing fine-tuning methods lack specificity, adapting the same set of parameters regardless of a task's visual, linguistic, and physical characteristics. Inspired by functional specificity in neuroscience, we hypothesize that it is more effective to finetune sparse model representations specific to a given task. In this work, we introduce Robotic Steering, a finetuning approach grounded in mechanistic interpretability that leverages few-shot demonstrations to identify and selectively finetune task-specific attention heads aligned with the physical, visual, and linguistic requirements of robotic tasks. Through comprehensive on-robot evaluations with a Franka Emika robot arm, we demonstrate that Robotic Steering outperforms LoRA while achieving superior robustness under task variation, reduced computational cost, and enhanced interpretability for adapting VLAs to diverse robotic tasks.

Dantong Niu, Yuvan Sharma, Baifeng Shi et al.

Robotic manipulation policies often struggle to generalize to novel objects, limiting their real-world utility. In contrast, cognitive science suggests that children develop generalizable dexterous manipulation skills by mastering a small set of simple toys and then applying that knowledge to more complex items. Inspired by this, we study if similar generalization capabilities can also be achieved by robots. Our results indicate robots can learn generalizable grasping using randomly assembled objects that are composed from just four shape primitives: spheres, cuboids, cylinders, and rings. We show that training on these "toys" enables robust generalization to real-world objects, yielding strong zero-shot performance. Crucially, we find the key to this generalization is an object-centric visual representation induced by our proposed detection pooling mechanism. Evaluated in both simulation and on physical robots, our model achieves a 67% real-world grasping success rate on the YCB dataset, outperforming state-of-the-art approaches that rely on substantially more in-domain data. We further study how zero-shot generalization performance scales by varying the number and diversity of training toys and the demonstrations per toy. We believe this work offers a promising path to scalable and generalizable learning in robotic manipulation. Demonstration videos, code, checkpoints and our dataset are available on our project page: https://lego-grasp.github.io/ .

Tianning Chai, Chancharik Mitra, Brandon Huang et al.

Aligning Large Language Models (LLMs) and Large Multimodal Models (LMMs) to human preferences is a central challenge in improving the quality of the models' generative outputs for real-world applications. A common approach is to use reward modeling to encode preferences, enabling alignment via post-training using reinforcement learning. However, traditional reward modeling is not easily adaptable to new preferences because it requires a separate reward model, commonly trained on large preference datasets. To address this, we introduce Activation Reward Models (Activation RMs) -- a novel few-shot reward modeling method that leverages activation steering to construct well-aligned reward signals using minimal supervision and no additional model finetuning. Activation RMs outperform existing few-shot reward modeling approaches such as LLM-as-a-judge with in-context learning, voting-based scoring, and token probability scoring on standard reward modeling benchmarks. Furthermore, we demonstrate the effectiveness of Activation RMs in mitigating reward hacking behaviors, highlighting their utility for safety-critical applications. Toward this end, we propose PreferenceHack, a novel few-shot setting benchmark, the first to test reward models on reward hacking in a paired preference format. Finally, we show that Activation RM achieves state-of-the-art performance on this benchmark, surpassing even GPT-4o.

Nasim Borazjanizadeh, Roei Herzig, Trevor Darrell et al.

Large Language Models (LLMs) have recently achieved impressive performance in math and reasoning benchmarks. However, they often struggle with logic problems and puzzles that are relatively easy for humans. To further investigate this, we introduce a new benchmark, SearchBench, which contains 11 unique search problems inspired by intuitive puzzles. Each SearchBench problem type is equipped with automated pipelines to generate an arbitrary number of instances and analyze the feasibility, correctness, and optimality of LLM-generated solutions. We show that using step-by-step, language-only reasoning, even the most advanced LLMs fail to solve SearchBench; for example, OpenAI's frontier models GPT-4 and o1-preview solve only 1.4% and 18.6% of problems, respectively. The reason is that SearchBench problems require considering multiple pathways and performing backtracking, posing a significant challenge to auto-regressive models. Interestingly, performance is significantly boosted when we prompt models to generate a complete A* search algorithm - a comparatively more cognitively difficult task. This approach effectively offloads the iterative search and backtracking process from the models, which they struggle with in text. This in-context learning baseline is further enhanced via a Multi-Stage-Multi-Try (MSMT) inference method, increasing GPT-4's rate of correct solutions to over 57%.

Dantong Niu, Yuvan Sharma, Giscard Biamby et al.

In recent years, instruction-tuned Large Multimodal Models (LMMs) have been successful at several tasks, including image captioning and visual question answering; yet leveraging these models remains an open question for robotics. Prior LMMs for robotics applications have been extensively trained on language and action data, but their ability to generalize in different settings has often been less than desired. To address this, we introduce LLARVA, a model trained with a novel instruction tuning method that leverages structured prompts to unify a range of robotic learning tasks, scenarios, and environments. Additionally, we show that predicting intermediate 2-D representations, which we refer to as "visual traces", can help further align vision and action spaces for robot learning. We generate 8.5M image-visual trace pairs from the Open X-Embodiment dataset in order to pre-train our model, and we evaluate on 12 different tasks in the RLBench simulator as well as a physical Franka Emika Panda 7-DoF robot. Our experiments yield strong performance, demonstrating that LLARVA - using 2-D and language representations - performs well compared to several contemporary baselines, and can generalize across various robot environments and configurations.

Irene Huang, Wei Lin, M. Jehanzeb Mirza et al.

Compositional Reasoning (CR) entails grasping the significance of attributes, relations, and word order. Recent Vision-Language Models (VLMs), comprising a visual encoder and a Large Language Model (LLM) decoder, have demonstrated remarkable proficiency in such reasoning tasks. This prompts a crucial question: have VLMs effectively tackled the CR challenge? We conjecture that existing CR benchmarks may not adequately push the boundaries of modern VLMs due to the reliance on an LLM-only negative text generation pipeline. Consequently, the negatives produced either appear as outliers from the natural language distribution learned by VLMs' LLM decoders or as improbable within the corresponding image context. To address these limitations, we introduce ConMe -- a compositional reasoning benchmark and a novel data generation pipeline leveraging VLMs to produce `hard CR Q&A'. Through a new concept of VLMs conversing with each other to collaboratively expose their weaknesses, our pipeline autonomously generates, evaluates, and selects challenging compositional reasoning questions, establishing a robust CR benchmark, also subsequently validated manually. Our benchmark provokes a noteworthy, up to 33%, decrease in CR performance compared to preceding benchmarks, reinstating the CR challenge even for state-of-the-art VLMs.

Sivan Doveh, Assaf Arbelle, Sivan Harary et al.

Vision and Language (VL) models offer an effective method for aligning representation spaces of images and text, leading to numerous applications such as cross-modal retrieval, visual question answering, captioning, and more. However, the aligned image-text spaces learned by all the popular VL models are still suffering from the so-called `object bias' - their representations behave as `bags of nouns', mostly ignoring or downsizing the attributes, relations, and states of objects described/appearing in texts/images. Although some great attempts at fixing these `compositional reasoning' issues were proposed in the recent literature, the problem is still far from being solved. In this paper, we uncover two factors limiting the VL models' compositional reasoning performance. These two factors are properties of the paired VL dataset used for finetuning and pre-training the VL model: (i) the caption quality, or in other words `image-alignment', of the texts; and (ii) the `density' of the captions in the sense of mentioning all the details appearing on the image. We propose a fine-tuning approach for automatically treating these factors leveraging a standard VL dataset (CC3M). Applied to CLIP, we demonstrate its significant compositional reasoning performance increase of up to $\sim27\%$ over the base model, up to $\sim20\%$ over the strongest baseline, and by $6.7\%$ on average.

Roei Herzig, Alon Mendelson, Leonid Karlinsky et al.

Vision and language models (VLMs) have demonstrated remarkable zero-shot (ZS) performance in a variety of tasks. However, recent works have shown that even the best VLMs struggle to capture aspects of compositional scene understanding, such as object attributes, relations, and action states. In contrast, obtaining structured annotations, such as scene graphs (SGs), that could improve these models is time-consuming and costly, and thus cannot be used on a large scale. Here we ask whether small SG datasets can provide sufficient information for enhancing structured understanding of pretrained VLMs. We show that it is indeed possible to improve VLMs when learning from SGs by integrating components that incorporate structured information into both visual and textual representations. For the visual side, we incorporate a special "SG Component" in the image transformer trained to predict SG information, while for the textual side, we utilize SGs to generate fine-grained captions that highlight different compositional aspects of the scene. Our method improves the performance of several popular VLMs on multiple VL datasets with only a mild degradation in ZS capabilities.

Sivan Harary, Eli Schwartz, Assaf Arbelle et al.

The ability to generalize learned representations across significantly different visual domains, such as between real photos, clipart, paintings, and sketches, is a fundamental capacity of the human visual system. In this paper, different from most cross-domain works that utilize some (or full) source domain supervision, we approach a relatively new and very practical Unsupervised Domain Generalization (UDG) setup of having no training supervision in neither source nor target domains. Our approach is based on self-supervised learning of a Bridge Across Domains (BrAD) - an auxiliary bridge domain accompanied by a set of semantics preserving visual (image-to-image) mappings to BrAD from each of the training domains. The BrAD and mappings to it are learned jointly (end-to-end) with a contrastive self-supervised representation model that semantically aligns each of the domains to its BrAD-projection, and hence implicitly drives all the domains (seen or unseen) to semantically align to each other. In this work, we show how using an edge-regularized BrAD our approach achieves significant gains across multiple benchmarks and a range of tasks, including UDG, Few-shot UDA, and unsupervised generalization across multi-domain datasets (including generalization to unseen domains and classes).

Roei Herzig, Elad Ben-Avraham, Karttikeya Mangalam et al.

Recently, video transformers have shown great success in video understanding, exceeding CNN performance; yet existing video transformer models do not explicitly model objects, although objects can be essential for recognizing actions. In this work, we present Object-Region Video Transformers (ORViT), an \emph{object-centric} approach that extends video transformer layers with a block that directly incorporates object representations. The key idea is to fuse object-centric representations starting from early layers and propagate them into the transformer-layers, thus affecting the spatio-temporal representations throughout the network. Our ORViT block consists of two object-level streams: appearance and dynamics. In the appearance stream, an "Object-Region Attention" module applies self-attention over the patches and \emph{object regions}. In this way, visual object regions interact with uniform patch tokens and enrich them with contextualized object information. We further model object dynamics via a separate "Object-Dynamics Module", which captures trajectory interactions, and show how to integrate the two streams. We evaluate our model on four tasks and five datasets: compositional and few-shot action recognition on SomethingElse, spatio-temporal action detection on AVA, and standard action recognition on Something-Something V2, Diving48 and Epic-Kitchen100. We show strong performance improvement across all tasks and datasets considered, demonstrating the value of a model that incorporates object representations into a transformer architecture. For code and pretrained models, visit the project page at \url{https://roeiherz.github.io/ORViT/}

Amir Bar, Xin Wang, Vadim Kantorov et al.

Recent self-supervised pretraining methods for object detection largely focus on pretraining the backbone of the object detector, neglecting key parts of detection architecture. Instead, we introduce DETReg, a new self-supervised method that pretrains the entire object detection network, including the object localization and embedding components. During pretraining, DETReg predicts object localizations to match the localizations from an unsupervised region proposal generator and simultaneously aligns the corresponding feature embeddings with embeddings from a self-supervised image encoder. We implement DETReg using the DETR family of detectors and show that it improves over competitive baselines when finetuned on COCO, PASCAL VOC, and Airbus Ship benchmarks. In low-data regimes DETReg achieves improved performance, e.g., when training with only 1% of the labels and in the few-shot learning settings.

Achiya Jerbi, Roei Herzig, Jonathan Berant et al.

Object detection is a fundamental task in computer vision, requiring large annotated datasets that are difficult to collect, as annotators need to label objects and their bounding boxes. Thus, it is a significant challenge to use cheaper forms of supervision effectively. Recent work has begun to explore image captions as a source for weak supervision, but to date, in the context of object detection, captions have only been used to infer the categories of the objects in the image. In this work, we argue that captions contain much richer information about the image, including attributes of objects and their relations. Namely, the text represents a scene of the image, as described recently in the literature. We present a method that uses the attributes in this "textual scene graph" to train object detectors. We empirically demonstrate that the resulting model achieves state-of-the-art results on several challenging object detection datasets, outperforming recent approaches.

Amir Bar, Roei Herzig, Xiaolong Wang et al.

Videos of actions are complex signals containing rich compositional structure in space and time. Current video generation methods lack the ability to condition the generation on multiple coordinated and potentially simultaneous timed actions. To address this challenge, we propose to represent the actions in a graph structure called Action Graph and present the new ``Action Graph To Video'' synthesis task. Our generative model for this task (AG2Vid) disentangles motion and appearance features, and by incorporating a scheduling mechanism for actions facilitates a timely and coordinated video generation. We train and evaluate AG2Vid on the CATER and Something-Something V2 datasets, and show that the resulting videos have better visual quality and semantic consistency compared to baselines. Finally, our model demonstrates zero-shot abilities by synthesizing novel compositions of the learned actions. For code and pretrained models, see the project page https://roeiherz.github.io/AG2Video

Joanna Materzynska, Tete Xiao, Roei Herzig et al.

Human action is naturally compositional: humans can easily recognize and perform actions with objects that are different from those used in training demonstrations. In this paper, we study the compositionality of action by looking into the dynamics of subject-object interactions. We propose a novel model which can explicitly reason about the geometric relations between constituent objects and an agent performing an action. To train our model, we collect dense object box annotations on the Something-Something dataset. We propose a novel compositional action recognition task where the training combinations of verbs and nouns do not overlap with the test set. The novel aspects of our model are applicable to activities with prominent object interaction dynamics and to objects which can be tracked using state-of-the-art approaches; for activities without clearly defined spatial object-agent interactions, we rely on baseline scene-level spatio-temporal representations. We show the effectiveness of our approach not only on the proposed compositional action recognition task, but also in a few-shot compositional setting which requires the model to generalize across both object appearance and action category.

Roei Herzig, Amir Bar, Huijuan Xu et al.

Generating realistic images of complex visual scenes becomes challenging when one wishes to control the structure of the generated images. Previous approaches showed that scenes with few entities can be controlled using scene graphs, but this approach struggles as the complexity of the graph (the number of objects and edges) increases. In this work, we show that one limitation of current methods is their inability to capture semantic equivalence in graphs. We present a novel model that addresses these issues by learning canonical graph representations from the data, resulting in improved image generation for complex visual scenes. Our model demonstrates improved empirical performance on large scene graphs, robustness to noise in the input scene graph, and generalization on semantically equivalent graphs. Finally, we show improved performance of the model on three different benchmarks: Visual Genome, COCO, and CLEVR.

Eli Brosh, Matan Friedmann, Ilan Kadar et al.

Accurate vehicle localization is a crucial step towards building effective Vehicle-to-Vehicle networks and automotive applications. Yet standard grade GPS data, such as that provided by mobile phones, is often noisy and exhibits significant localization errors in many urban areas. Approaches for accurate localization from imagery often rely on structure-based techniques, and thus are limited in scale and are expensive to compute. In this paper, we present a scalable visual localization approach geared for real-time performance. We propose a hybrid coarse-to-fine approach that leverages visual and GPS location cues. Our solution uses a self-supervised approach to learn a compact road image representation. This representation enables efficient visual retrieval and provides coarse localization cues, which are fused with vehicle ego-motion to obtain high accuracy location estimates. As a benchmark to evaluate the performance of our visual localization approach, we introduce a new large-scale driving dataset based on video and GPS data obtained from a large-scale network of connected dash-cams. Our experiments confirm that our approach is highly effective in challenging urban environments, reducing localization error by an order of magnitude.

Moshiko Raboh, Roei Herzig, Gal Chechik et al.

Reasoning about complex visual scenes involves perception of entities and their relations. Scene graphs provide a natural representation for reasoning tasks, by assigning labels to both entities (nodes) and relations (edges). Unfortunately, reasoning systems based on SGs are typically trained in a two-step procedure: First, training a model to predict SGs from images; Then, a separate model is created to reason based on predicted SGs. In many domains, it is preferable to train systems jointly in an end-to-end manner, but SGs are not commonly used as intermediate components in visual reasoning systems because being discrete and sparse, scene-graph representations are non-differentiable and difficult to optimize. Here we propose Differentiable Scene Graphs (DSGs), an image representation that is amenable to differentiable end-to-end optimization, and requires supervision only from the downstream tasks. DSGs provide a dense representation for all regions and pairs of regions, and do not spend modelling capacity on areas of the images that do not contain objects or relations of interest. We evaluate our model on the challenging task of identifying referring relationships (RR) in three benchmark datasets, Visual Genome, VRD and CLEVR. We describe a multi-task objective, and train in an end-to-end manner supervised by the downstream RR task. Using DSGs as an intermediate representation leads to new state-of-the-art performance.

Roei Herzig, Elad Levi, Huijuan Xu et al.

Events defined by the interaction of objects in a scene are often of critical importance; yet important events may have insufficient labeled examples to train a conventional deep model to generalize to future object appearance. Activity recognition models that represent object interactions explicitly have the potential to learn in a more efficient manner than those that represent scenes with global descriptors. We propose a novel inter-object graph representation for activity recognition based on a disentangled graph embedding with direct observation of edge appearance. We employ a novel factored embedding of the graph structure, disentangling a representation hierarchy formed over spatial dimensions from that found over temporal variation. We demonstrate the effectiveness of our model on the Charades activity recognition benchmark, as well as a new dataset of driving activities focusing on multi-object interactions with near-collision events. Our model offers significantly improved performance compared to baseline approaches without object-graph representations, or with previous graph-based models.

Roei Herzig, Moshiko Raboh, Gal Chechik et al.

Machine understanding of complex images is a key goal of artificial intelligence. One challenge underlying this task is that visual scenes contain multiple inter-related objects, and that global context plays an important role in interpreting the scene. A natural modeling framework for capturing such effects is structured prediction, which optimizes over complex labels, while modeling within-label interactions. However, it is unclear what principles should guide the design of a structured prediction model that utilizes the power of deep learning components. Here we propose a design principle for such architectures that follows from a natural requirement of permutation invariance. We prove a necessary and sufficient characterization for architectures that follow this invariance, and discuss its implication on model design. Finally, we show that the resulting model achieves new state of the art results on the Visual Genome scene graph labeling benchmark, outperforming all recent approaches.