Jing Yu Koh

Jiahui Yu, Yuanzhong Xu, Jing Yu Koh et al. · cmu

We present the Pathways Autoregressive Text-to-Image (Parti) model, which generates high-fidelity photorealistic images and supports content-rich synthesis involving complex compositions and world knowledge. Parti treats text-to-image generation as a sequence-to-sequence modeling problem, akin to machine translation, with sequences of image tokens as the target outputs rather than text tokens in another language. This strategy can naturally tap into the rich body of prior work on large language models, which have seen continued advances in capabilities and performance through scaling data and model sizes. Our approach is simple: First, Parti uses a Transformer-based image tokenizer, ViT-VQGAN, to encode images as sequences of discrete tokens. Second, we achieve consistent quality improvements by scaling the encoder-decoder Transformer model up to 20B parameters, with a new state-of-the-art zero-shot FID score of 7.23 and finetuned FID score of 3.22 on MS-COCO. Our detailed analysis on Localized Narratives as well as PartiPrompts (P2), a new holistic benchmark of over 1600 English prompts, demonstrate the effectiveness of Parti across a wide variety of categories and difficulty aspects. We also explore and highlight limitations of our models in order to define and exemplify key areas of focus for further improvements. See https://parti.research.google/ for high-resolution images.

Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried · cmu

We propose an efficient method to ground pretrained text-only language models to the visual domain, enabling them to process arbitrarily interleaved image-and-text data, and generate text interleaved with retrieved images. Our method leverages the abilities of language models learnt from large scale text-only pretraining, such as in-context learning and free-form text generation. We keep the language model frozen, and finetune input and output linear layers to enable cross-modality interactions. This allows our model to process arbitrarily interleaved image-and-text inputs, and generate free-form text interleaved with retrieved images. We achieve strong zero-shot performance on grounded tasks such as contextual image retrieval and multimodal dialogue, and showcase compelling interactive abilities. Our approach works with any off-the-shelf language model and paves the way towards an effective, general solution for leveraging pretrained language models in visually grounded settings.

Jing Yu Koh, Harsh Agrawal, Dhruv Batra et al. · apple-ml, cmu

We study the problem of synthesizing immersive 3D indoor scenes from one or more images. Our aim is to generate high-resolution images and videos from novel viewpoints, including viewpoints that extrapolate far beyond the input images while maintaining 3D consistency. Existing approaches are highly complex, with many separately trained stages and components. We propose a simple alternative: an image-to-image GAN that maps directly from reprojections of incomplete point clouds to full high-resolution RGB-D images. On the Matterport3D and RealEstate10K datasets, our approach significantly outperforms prior work when evaluated by humans, as well as on FID scores. Further, we show that our model is useful for generative data augmentation. A vision-and-language navigation (VLN) agent trained with trajectories spatially-perturbed by our model improves success rate by up to 1.5% over a state of the art baseline on the R2R benchmark. Our code will be made available to facilitate generative data augmentation and applications to downstream robotics and embodied AI tasks.

Kyle Sargent, Jing Yu Koh, Han Zhang et al. · cmu, stanford

Recent work has shown the possibility of training generative models of 3D content from 2D image collections on small datasets corresponding to a single object class, such as human faces, animal faces, or cars. However, these models struggle on larger, more complex datasets. To model diverse and unconstrained image collections such as ImageNet, we present VQ3D, which introduces a NeRF-based decoder into a two-stage vector-quantized autoencoder. Our Stage 1 allows for the reconstruction of an input image and the ability to change the camera position around the image, and our Stage 2 allows for the generation of new 3D scenes. VQ3D is capable of generating and reconstructing 3D-aware images from the 1000-class ImageNet dataset of 1.2 million training images. We achieve an ImageNet generation FID score of 16.8, compared to 69.8 for the next best baseline method.

Aishwarya Kamath, Peter Anderson, Su Wang et al. · cmu

Recent studies in Vision-and-Language Navigation (VLN) train RL agents to execute natural-language navigation instructions in photorealistic environments, as a step towards robots that can follow human instructions. However, given the scarcity of human instruction data and limited diversity in the training environments, these agents still struggle with complex language grounding and spatial language understanding. Pretraining on large text and image-text datasets from the web has been extensively explored but the improvements are limited. We investigate large-scale augmentation with synthetic instructions. We take 500+ indoor environments captured in densely-sampled 360 degree panoramas, construct navigation trajectories through these panoramas, and generate a visually-grounded instruction for each trajectory using Marky, a high-quality multilingual navigation instruction generator. We also synthesize image observations from novel viewpoints using an image-to-image GAN. The resulting dataset of 4.2M instruction-trajectory pairs is two orders of magnitude larger than existing human-annotated datasets, and contains a wider variety of environments and viewpoints. To efficiently leverage data at this scale, we train a simple transformer agent with imitation learning. On the challenging RxR dataset, our approach outperforms all existing RL agents, improving the state-of-the-art NDTW from 71.1 to 79.1 in seen environments, and from 64.6 to 66.8 in unseen test environments. Our work points to a new path to improving instruction-following agents, emphasizing large-scale imitation learning and the development of synthetic instruction generation capabilities.

Jing Yu Koh, Stephen McAleer, Daniel Fried et al. · cmu

Autonomous agents powered by language models (LMs) have demonstrated promise in their ability to perform decision-making tasks such as web automation. However, a key limitation remains: LMs, primarily optimized for natural language understanding and generation, struggle with multi-step reasoning, planning, and using environmental feedback when attempting to solve realistic computer tasks. Towards addressing this, we propose an inference-time search algorithm for LM agents to explicitly perform exploration and multi-step planning in interactive web environments. Our approach is a form of best-first tree search that operates within the actual environment space, and is complementary with most existing state-of-the-art agents. It is the first tree search algorithm for LM agents that shows effectiveness on realistic web tasks. On the challenging VisualWebArena benchmark, applying our search algorithm on top of a GPT-4o agent yields a 39.7% relative increase in success rate compared to the same baseline without search, setting a state-of-the-art success rate of 26.4%. On WebArena, search also yields a 28.0% relative improvement over a baseline agent, setting a competitive success rate of 19.2%. Our experiments highlight the effectiveness of search for web agents, and we demonstrate that performance scales with increased test-time compute. We conduct a thorough analysis of our results to highlight improvements from search, limitations, and promising directions for future work. Our code and models are publicly released at https://jykoh.com/search-agents.

Minji Yoon, Jing Yu Koh, Bryan Hooi et al.

Multimodal learning combines multiple data modalities, broadening the types and complexity of data our models can utilize: for example, from plain text to image-caption pairs. Most multimodal learning algorithms focus on modeling simple one-to-one pairs of data from two modalities, such as image-caption pairs, or audio-text pairs. However, in most real-world settings, entities of different modalities interact with each other in more complex and multifaceted ways, going beyond one-to-one mappings. We propose to represent these complex relationships as graphs, allowing us to capture data with any number of modalities, and with complex relationships between modalities that can flexibly vary from one sample to another. Toward this goal, we propose Multimodal Graph Learning (MMGL), a general and systematic framework for capturing information from multiple multimodal neighbors with relational structures among them. In particular, we focus on MMGL for generative tasks, building upon pretrained Language Models (LMs), aiming to augment their text generation with multimodal neighbor contexts. We study three research questions raised by MMGL: (1) how can we infuse multiple neighbor information into the pretrained LMs, while avoiding scalability issues? (2) how can we infuse the graph structure information among multimodal neighbors into the LMs? and (3) how can we finetune the pretrained LMs to learn from the neighbor context in a parameter-efficient manner? We conduct extensive experiments to answer these three questions on MMGL and analyze the empirical results to pave the way for future MMGL research.

Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried

Computer use agents (CUAs) today are primarily deployed as single serial agents. This setup is suboptimal for complex long-horizon tasks that benefit from task decomposition, parallel execution, and consistent re-planning based on new information. In this paper, we argue that we should instead move towards evaluating and building multi-agent computer use (MACU) systems. These systems, which emphasize planning and parallel execution, alleviate many of the shortcomings of single-agent CUAs. We propose a general multi-agent setup in which a manager model decomposes computer use tasks as a directed acyclic graph (DAG), encoding relevant dependencies and goals for subagents. At each iteration, the manager dispatches parallel CUA subagents to carry out nodes on the ready frontier of the DAG, and continuously revises the DAG (adding, canceling, or rewriting nodes) as new findings arrive from subagents. This design treats the partially observable environment of computer use as a first class challenge: information that downstream agents may not be able to re-observe are retained and passed forward through the manager and DAG structure. We demonstrate that MACU consistently improves over strong single-agent baselines by $3.4-25.5\%$ on desktop (OSWorld) and web navigation (Online-Mind2Web, WebTailBench, Odysseys) benchmarks, exhibits more favorable test-time scaling, and solves complex long-horizon tasks where single-agent CUAs get stuck. On Odysseys, a long-horizon web navigation benchmark, MACU improves average task completion wall-clock time by ${\sim} 1.5 \times$, demonstrating its efficacy in speeding up traditionally slow CUA pipelines. Our findings highlight that multi-agent coordination is a promising axis for scaling computer use agents to work productively for longer and more effectively. We release all code and interactive visualizations at https://jykoh.com/multi-agent-computer-use.

Lawrence Keunho Jang, Jing Yu Koh, Daniel Fried et al.

Existing web agent benchmarks have largely converged on short, single-site tasks that frontier models are approaching saturation on. However, real world web use consists of long-horizon, multi-site workflows. Common web navigation tasks, such as comparing products across different domains, planning trips across multiple services, or summarizing information from multiple search queries, require sustained context and cross-site reasoning over potentially hours of browsing. To capture and evaluate such behaviors, we introduce Odysseys: a benchmark of 200 long-horizon web tasks derived from real world browsing sessions evaluated on the live Internet. We find that binary pass/fail evaluation is inadequate for long-horizon settings and introduce a rubric-based evaluation, annotating each Odysseys task with an average of 6.1 graded rubrics. We demonstrate that this yields higher agreement with humans and provides a more fine-grained signal than commonly used trajectory-level LLM-as-a-judge evaluation metrics. We tested several leading frontier models and find that the strongest models achieve a success rate of 44.5%, which leaves substantial room for future improvements. Beyond task success, we argue that efficiency is a first-class concern for long-horizon agents. We introduce a Trajectory Efficiency metric (rubric score per step) and find that even frontier agents achieve only 1.15%, marking an evident need for agents that can succeed efficiently and not simply eventually. Odysseys isolates the critical evaluation of long-horizon proficiency in open-web environments, providing a realistic benchmark to measure progress towards computer-use agents that can potentially productively operate for hours. We release our tasks, evaluation scripts, and other results at https://odysseys-website.pages.dev

Chen Henry Wu, Rishi Shah, Jing Yu Koh et al.

As language models (LMs) are used to build autonomous agents in real environments, ensuring their adversarial robustness becomes a critical challenge. Unlike chatbots, agents are compound systems with multiple components taking actions, which existing LMs safety evaluations do not adequately address. To bridge this gap, we manually create 200 targeted adversarial tasks and evaluation scripts in a realistic threat model on top of VisualWebArena, a real environment for web agents. To systematically examine the robustness of agents, we propose the Agent Robustness Evaluation (ARE) framework. ARE views the agent as a graph showing the flow of intermediate outputs between components and decomposes robustness as the flow of adversarial information on the graph. We find that we can successfully break latest agents that use black-box frontier LMs, including those that perform reflection and tree search. With imperceptible perturbations to a single image (less than 5% of total web page pixels), an attacker can hijack these agents to execute targeted adversarial goals with success rates up to 67%. We also use ARE to rigorously evaluate how the robustness changes as new components are added. We find that inference-time compute that typically improves benign performance can open up new vulnerabilities and harm robustness. An attacker can compromise the evaluator used by the reflexion agent and the value function of the tree search agent, which increases the attack success relatively by 15% and 20%. Our data and code for attacks, defenses, and evaluation are at https://github.com/ChenWu98/agent-attack

Jing Yu Koh, Robert Lo, Lawrence Jang et al.

Autonomous agents capable of planning, reasoning, and executing actions on the web offer a promising avenue for automating computer tasks. However, the majority of existing benchmarks primarily focus on text-based agents, neglecting many natural tasks that require visual information to effectively solve. Given that most computer interfaces cater to human perception, visual information often augments textual data in ways that text-only models struggle to harness effectively. To bridge this gap, we introduce VisualWebArena, a benchmark designed to assess the performance of multimodal web agents on realistic \textit{visually grounded tasks}. VisualWebArena comprises of a set of diverse and complex web-based tasks that evaluate various capabilities of autonomous multimodal agents. To perform on this benchmark, agents need to accurately process image-text inputs, interpret natural language instructions, and execute actions on websites to accomplish user-defined objectives. We conduct an extensive evaluation of state-of-the-art LLM-based autonomous agents, including several multimodal models. Through extensive quantitative and qualitative analysis, we identify several limitations of text-only LLM agents, and reveal gaps in the capabilities of state-of-the-art multimodal language agents. VisualWebArena provides a framework for evaluating multimodal autonomous language agents, and offers insights towards building stronger autonomous agents for the web. Our code, baseline models, and data is publicly available at https://jykoh.com/vwa.

Raghav Kapoor, Yash Parag Butala, Melisa Russak et al. · cmu

For decades, human-computer interaction has fundamentally been manual. Even today, almost all productive work done on the computer necessitates human input at every step. Autonomous virtual agents represent an exciting step in automating many of these menial tasks. Virtual agents would empower users with limited technical proficiency to harness the full possibilities of computer systems. They could also enable the efficient streamlining of numerous computer tasks, ranging from calendar management to complex travel bookings, with minimal human intervention. In this paper, we introduce OmniACT, the first-of-a-kind dataset and benchmark for assessing an agent's capability to generate executable programs to accomplish computer tasks. Our scope extends beyond traditional web automation, covering a diverse range of desktop applications. The dataset consists of fundamental tasks such as "Play the next song", as well as longer horizon tasks such as "Send an email to John Doe mentioning the time and place to meet". Specifically, given a pair of screen image and a visually-grounded natural language task, the goal is to generate a script capable of fully executing the task. We run several strong baseline language model agents on our benchmark. The strongest baseline, GPT-4, performs the best on our benchmark However, its performance level still reaches only 15% of the human proficiency in generating executable scripts capable of completing the task, demonstrating the challenge of our task for conventional web agents. Our benchmark provides a platform to measure and evaluate the progress of language model agents in automating computer tasks and motivates future work towards building multimodal models that bridge large language models and the visual grounding of computer screens.

Jing Yu Koh, Daniel Fried, Ruslan Salakhutdinov

We propose a method to fuse frozen text-only large language models (LLMs) with pre-trained image encoder and decoder models, by mapping between their embedding spaces. Our model demonstrates a wide suite of multimodal capabilities: image retrieval, novel image generation, and multimodal dialogue. Ours is the first approach capable of conditioning on arbitrarily interleaved image and text inputs to generate coherent image (and text) outputs. To achieve strong performance on image generation, we propose an efficient mapping network to ground the LLM to an off-the-shelf text-to-image generation model. This mapping network translates hidden representations of text into the embedding space of the visual models, enabling us to leverage the strong text representations of the LLM for visual outputs. Our approach outperforms baseline generation models on tasks with longer and more complex language. In addition to novel image generation, our model is also capable of image retrieval from a prespecified dataset, and decides whether to retrieve or generate at inference time. This is done with a learnt decision module which conditions on the hidden representations of the LLM. Our model exhibits a wider range of capabilities compared to prior multimodal language models. It can process image-and-text inputs, and produce retrieved images, generated images, and generated text -- outperforming non-LLM based generation models across several text-to-image tasks that measure context dependence.

Jiahui Yu, Xin Li, Jing Yu Koh et al.

Pretraining language models with next-token prediction on massive text corpora has delivered phenomenal zero-shot, few-shot, transfer learning and multi-tasking capabilities on both generative and discriminative language tasks. Motivated by this success, we explore a Vector-quantized Image Modeling (VIM) approach that involves pretraining a Transformer to predict rasterized image tokens autoregressively. The discrete image tokens are encoded from a learned Vision-Transformer-based VQGAN (ViT-VQGAN). We first propose multiple improvements over vanilla VQGAN from architecture to codebook learning, yielding better efficiency and reconstruction fidelity. The improved ViT-VQGAN further improves vector-quantized image modeling tasks, including unconditional, class-conditioned image generation and unsupervised representation learning. When trained on ImageNet at \(256\times256\) resolution, we achieve Inception Score (IS) of 175.1 and Fr'echet Inception Distance (FID) of 4.17, a dramatic improvement over the vanilla VQGAN, which obtains 70.6 and 17.04 for IS and FID, respectively. Based on ViT-VQGAN and unsupervised pretraining, we further evaluate the pretrained Transformer by averaging intermediate features, similar to Image GPT (iGPT). This ImageNet-pretrained VIM-L significantly beats iGPT-L on linear-probe accuracy from 60.3% to 73.2% for a similar model size. VIM-L also outperforms iGPT-XL which is trained with extra web image data and larger model size.

Jing Yu Koh, Honglak Lee, Yinfei Yang et al.

People navigating in unfamiliar buildings take advantage of myriad visual, spatial and semantic cues to efficiently achieve their navigation goals. Towards equipping computational agents with similar capabilities, we introduce Pathdreamer, a visual world model for agents navigating in novel indoor environments. Given one or more previous visual observations, Pathdreamer generates plausible high-resolution 360 visual observations (RGB, semantic segmentation and depth) for viewpoints that have not been visited, in buildings not seen during training. In regions of high uncertainty (e.g. predicting around corners, imagining the contents of an unseen room), Pathdreamer can predict diverse scenes, allowing an agent to sample multiple realistic outcomes for a given trajectory. We demonstrate that Pathdreamer encodes useful and accessible visual, spatial and semantic knowledge about human environments by using it in the downstream task of Vision-and-Language Navigation (VLN). Specifically, we show that planning ahead with Pathdreamer brings about half the benefit of looking ahead at actual observations from unobserved parts of the environment. We hope that Pathdreamer will help unlock model-based approaches to challenging embodied navigation tasks such as navigating to specified objects and VLN.

Wonkwang Lee, Whie Jung, Han Zhang et al.

Learning to predict the long-term future of video frames is notoriously challenging due to inherent ambiguities in the distant future and dramatic amplifications of prediction error through time. Despite the recent advances in the literature, existing approaches are limited to moderately short-term prediction (less than a few seconds), while extrapolating it to a longer future quickly leads to destruction in structure and content. In this work, we revisit hierarchical models in video prediction. Our method predicts future frames by first estimating a sequence of semantic structures and subsequently translating the structures to pixels by video-to-video translation. Despite the simplicity, we show that modeling structures and their dynamics in the discrete semantic structure space with a stochastic recurrent estimator leads to surprisingly successful long-term prediction. We evaluate our method on three challenging datasets involving car driving and human dancing, and demonstrate that it can generate complicated scene structures and motions over a very long time horizon (i.e., thousands frames), setting a new standard of video prediction with orders of magnitude longer prediction time than existing approaches. Full videos and codes are available at https://1konny.github.io/HVP/.

Han Zhang, Jing Yu Koh, Jason Baldridge et al.

The output of text-to-image synthesis systems should be coherent, clear, photo-realistic scenes with high semantic fidelity to their conditioned text descriptions. Our Cross-Modal Contrastive Generative Adversarial Network (XMC-GAN) addresses this challenge by maximizing the mutual information between image and text. It does this via multiple contrastive losses which capture inter-modality and intra-modality correspondences. XMC-GAN uses an attentional self-modulation generator, which enforces strong text-image correspondence, and a contrastive discriminator, which acts as a critic as well as a feature encoder for contrastive learning. The quality of XMC-GAN's output is a major step up from previous models, as we show on three challenging datasets. On MS-COCO, not only does XMC-GAN improve state-of-the-art FID from 24.70 to 9.33, but--more importantly--people prefer XMC-GAN by 77.3 for image quality and 74.1 for image-text alignment, compared to three other recent models. XMC-GAN also generalizes to the challenging Localized Narratives dataset (which has longer, more detailed descriptions), improving state-of-the-art FID from 48.70 to 14.12. Lastly, we train and evaluate XMC-GAN on the challenging Open Images data, establishing a strong benchmark FID score of 26.91.

Jing Yu Koh, Jason Baldridge, Honglak Lee et al.

Localized Narratives is a dataset with detailed natural language descriptions of images paired with mouse traces that provide a sparse, fine-grained visual grounding for phrases. We propose TReCS, a sequential model that exploits this grounding to generate images. TReCS uses descriptions to retrieve segmentation masks and predict object labels aligned with mouse traces. These alignments are used to select and position masks to generate a fully covered segmentation canvas; the final image is produced by a segmentation-to-image generator using this canvas. This multi-step, retrieval-based approach outperforms existing direct text-to-image generation models on both automatic metrics and human evaluations: overall, its generated images are more photo-realistic and better match descriptions.

Jing Yu Koh, Duc Thanh Nguyen, Quang-Trung Truong et al.

Fully-automatic execution is the ultimate goal for many Computer Vision applications. However, this objective is not always realistic in tasks associated with high failure costs, such as medical applications. For these tasks, semi-automatic methods allowing minimal effort from users to guide computer algorithms are often preferred due to desirable accuracy and performance. Inspired by the practicality and applicability of the semi-automatic approach, this paper proposes a novel deep neural network architecture, namely SideInfNet that effectively integrates features learnt from images with side information extracted from user annotations. To evaluate our method, we applied the proposed network to three semantic segmentation tasks and conducted extensive experiments on benchmark datasets. Experimental results and comparison with prior work have verified the superiority of our model, suggesting the generality and effectiveness of the model in semi-automatic semantic segmentation.

Jing Yu Koh, Wojciech Samek, Klaus-Robert Müller et al.

Semantic boundary and edge detection aims at simultaneously detecting object edge pixels in images and assigning class labels to them. Systematic training of predictors for this task requires the labeling of edges in images which is a particularly tedious task. We propose a novel strategy for solving this task, when pixel-level annotations are not available, performing it in an almost zero-shot manner by relying on conventional whole image neural net classifiers that were trained using large bounding boxes. Our method performs the following two steps at test time. Firstly it predicts the class labels by applying the trained whole image network to the test images. Secondly, it computes pixel-wise scores from the obtained predictions by applying backprop gradients as well as recent visualization algorithms such as deconvolution and layer-wise relevance propagation. We show that high pixel-wise scores are indicative for the location of semantic boundaries, which suggests that the semantic boundary problem can be approached without using edge labels during the training phase.