Spencer Whitehead

Alexander Kirillov, Eric Mintun, Nikhila Ravi et al. · meta-ai

We introduce the Segment Anything (SA) project: a new task, model, and dataset for image segmentation. Using our efficient model in a data collection loop, we built the largest segmentation dataset to date (by far), with over 1 billion masks on 11M licensed and privacy respecting images. The model is designed and trained to be promptable, so it can transfer zero-shot to new image distributions and tasks. We evaluate its capabilities on numerous tasks and find that its zero-shot performance is impressive -- often competitive with or even superior to prior fully supervised results. We are releasing the Segment Anything Model (SAM) and corresponding dataset (SA-1B) of 1B masks and 11M images at https://segment-anything.com to foster research into foundation models for computer vision.

Hao Bai, Rui Yang, Chenlu Ye et al.

Training vision-language web agents with multi-step RL is compute-intensive, with two dominant forms of inefficiency: idle GPUs in synchronous RL, and trajectories that use more steps and tokens than necessary. We present AsyncWebRL, which addresses both. On the system side, an asynchronous design overlaps rollout, gradient update, and policy refresh across iterations, paired with two web-agent-specific adaptations, namely an everlasting rollout pool and lightweight screenshot handling, that together deliver up to a $2.9\times$ end-to-end training-throughput speedup over the previously fastest open synchronous pipeline (WebGym). On the algorithmic side, we identify the per-trajectory normalizer $1/|τ_i|$ in multi-step GRPO as the root cause of trajectory-level and token-level inefficiency: because failures are systematically longer than successes, it down-weights the negative gradient on failed tokens, so the policy keeps producing verbose memory schemas. Replacing $1/|τ_i|$ with a constant $1/k$ breaks this coupling, contracting trajectories while preserving aggregate success. Together, these contributions set a new open-source state of the art on the WebGym out-of-distribution test split (+5.8% relative over the 42.9% prior best), with the largest gains on the harder slices (+42% relative on Medium, +48% relative on Hard).

Spencer Whitehead, Suzanne Petryk, Vedaad Shakib et al. · berkeley

Machine learning has advanced dramatically, narrowing the accuracy gap to humans in multimodal tasks like visual question answering (VQA). However, while humans can say "I don't know" when they are uncertain (i.e., abstain from answering a question), such ability has been largely neglected in multimodal research, despite the importance of this problem to the usage of VQA in real settings. In this work, we promote a problem formulation for reliable VQA, where we prefer abstention over providing an incorrect answer. We first enable abstention capabilities for several VQA models, and analyze both their coverage, the portion of questions answered, and risk, the error on that portion. For that, we explore several abstention approaches. We find that although the best performing models achieve over 70% accuracy on the VQA v2 dataset, introducing the option to abstain by directly using a model's softmax scores limits them to answering less than 7.5% of the questions to achieve a low risk of error (i.e., 1%). This motivates us to utilize a multimodal selection function to directly estimate the correctness of the predicted answers, which we show can increase the coverage by, for example, 2.3x from 6.8% to 15.6% at 1% risk. While it is important to analyze both coverage and risk, these metrics have a trade-off which makes comparing VQA models challenging. To address this, we also propose an Effective Reliability metric for VQA that places a larger cost on incorrect answers compared to abstentions. This new problem formulation, metric, and analysis for VQA provide the groundwork for building effective and reliable VQA models that have the self-awareness to abstain if and only if they don't know the answer.

Corentin Dancette, Spencer Whitehead, Rishabh Maheshwary et al.

Despite advances in Visual Question Answering (VQA), the ability of models to assess their own correctness remains underexplored. Recent work has shown that VQA models, out-of-the-box, can have difficulties abstaining from answering when they are wrong. The option to abstain, also called Selective Prediction, is highly relevant when deploying systems to users who must trust the system's output (e.g., VQA assistants for users with visual impairments). For such scenarios, abstention can be especially important as users may provide out-of-distribution (OOD) or adversarial inputs that make incorrect answers more likely. In this work, we explore Selective VQA in both in-distribution (ID) and OOD scenarios, where models are presented with mixtures of ID and OOD data. The goal is to maximize the number of questions answered while minimizing the risk of error on those questions. We propose a simple yet effective Learning from Your Peers (LYP) approach for training multimodal selection functions for making abstention decisions. Our approach uses predictions from models trained on distinct subsets of the training data as targets for optimizing a Selective VQA model. It does not require additional manual labels or held-out data and provides a signal for identifying examples that are easy/difficult to generalize to. In our extensive evaluations, we show this benefits a number of models across different architectures and scales. Overall, for ID, we reach 32.92% in the selective prediction metric coverage at 1% risk of error (C@1%) which doubles the previous best coverage of 15.79% on this task. For mixed ID/OOD, using models' softmax confidences for abstention decisions performs very poorly, answering <5% of questions at 1% risk of error even when faced with only 10% OOD examples, but a learned selection function with LYP can increase that to 25.38% C@1%.

Hao Bai, Alexey Taymanov, Tong Zhang et al.

We present WebGym, the largest-to-date open-source environment for training realistic visual web agents. Real websites are non-stationary and diverse, making artificial or small-scale task sets insufficient for robust policy learning. WebGym contains nearly 300,000 tasks with rubric-based evaluations across diverse, real-world websites and difficulty levels. We train agents with a simple reinforcement learning (RL) recipe, which trains on the agent's own interaction traces (rollouts), using task rewards as feedback to guide learning. To enable scaling RL, we speed up sampling of trajectories in WebGym by developing a high-throughput asynchronous rollout system, designed specifically for web agents. Our system achieves a 4-5x rollout speedup compared to naive implementations. Second, we scale the task set breadth, depth, and size, which results in continued performance improvement. Fine-tuning a strong base vision-language model, Qwen-3-VL-8B-Instruct, on WebGym results in an improvement in success rate on an out-of-distribution test set from 26.2% to 42.9%, significantly outperforming agents based on proprietary models such as GPT-4o and GPT-5-Thinking that achieve 27.1% and 29.8%, respectively. This improvement is substantial because our test set consists only of tasks on websites never seen during training, unlike many other prior works on training visual web agents.

Vardaan Pahuja, Yadong Lu, Corby Rosset et al. · microsoft-research

Recent success in large multimodal models (LMMs) has sparked promising applications of agents capable of autonomously completing complex web tasks. While open-source LMM agents have made significant advances in offline evaluation benchmarks, their performance still falls substantially short of human-level capabilities in more realistic online settings. A key bottleneck is the lack of diverse and large-scale trajectory-level datasets across various domains, which are expensive to collect. In this paper, we address this challenge by developing a scalable recipe to synthesize the largest and most diverse trajectory-level dataset to date, containing over 94K successful multimodal web trajectories, spanning 49K unique URLs, 720K screenshots, and 33M web elements. In particular, we leverage extensive web exploration and refinement to obtain diverse task intents. The average cost is 28 cents per successful trajectory, making it affordable to a wide range of users in the community. Leveraging this dataset, we train Explorer, a multimodal web agent, and demonstrate strong performance on both offline and online web agent benchmarks such as Mind2Web-Live, Multimodal-Mind2Web, and MiniWob++. Additionally, our experiments highlight data scaling as a key driver for improving web agent capabilities. We hope this study makes state-of-the-art LMM-based agent research at a larger scale more accessible.

Spencer Whitehead, Jacob Phillips, Sean Hendryx

Multimodal language models can exhibit hallucinations in their outputs, which limits their reliability. The ability to automatically detect these errors is important for mitigating them, but has been less explored and existing efforts do not localize hallucinations, instead framing this as a classification task. In this work, we first pose multimodal hallucination detection as a sequence labeling task where models must localize hallucinated text spans and present a strong baseline model. Given the high cost of human annotations for this task, we propose an approach to improve the sample efficiency of these models by creating corrupted grounding data, which we use for pre-training. Leveraging phrase grounding data, we generate hallucinations to replace grounded spans and create hallucinated text. Experiments show that pre-training on this data improves sample efficiency when fine-tuning, and that the learning signal from the grounding data plays an important role in these improvements.

Vaskar Nath, Dylan Slack, Jeff Da et al.

Techniques that learn improved representations via offline data or self-supervised objectives have shown impressive results in traditional reinforcement learning (RL). Nevertheless, it is unclear how improved representation learning can benefit reinforcement learning from human feedback (RLHF) on language models (LMs). In this work, we propose training reward models (RMs) in a contrastive, $\textit{goal-conditioned}$ fashion by increasing the representation similarity of future states along sampled preferred trajectories and decreasing the similarity along randomly sampled dispreferred trajectories. This objective significantly improves RM performance by up to 0.09 AUROC across challenging benchmarks, such as MATH and GSM8k. These findings extend to general alignment as well -- on the Helpful-Harmless dataset, we observe $2.3\%$ increase in accuracy. Beyond improving reward model performance, we show this way of training RM representations enables improved $\textit{steerability}$ because it allows us to evaluate the likelihood of an action achieving a particular goal-state (e.g., whether a solution is correct or helpful). Leveraging this insight, we find that we can filter up to $55\%$ of generated tokens during majority voting by discarding trajectories likely to end up in an "incorrect" state, which leads to significant cost savings. We additionally find that these representations can perform fine-grained control by conditioning on desired future goal-states. For example, we show that steering a Llama 3 model towards helpful generations with our approach improves helpfulness by $9.6\%$ over a supervised-fine-tuning trained baseline. Similarly, steering the model towards complex generations improves complexity by $21.6\%$ over the baseline. Overall, we find that training RMs in this contrastive, goal-conditioned fashion significantly improves performance and enables model steerability.

Ahmed Awadallah, Yash Lara, Raghav Magazine et al.

Progress in computer use agents (CUAs) has been constrained by the absence of large and high-quality datasets that capture how humans interact with a computer. While LLMs have thrived on abundant textual data, no comparable corpus exists for CUA trajectories. To address these gaps, we introduce FaraGen, a novel synthetic data generation system for multi-step web tasks. FaraGen can propose diverse tasks from frequently used websites, generate multiple solution attempts, and filter successful trajectories using multiple verifiers. It achieves high throughput, yield, and diversity for multi-step web tasks, producing verified trajectories at approximately $1 each. We use this data to train Fara-7B, a native CUA model that perceives the computer using only screenshots, executes actions via predicted coordinates, and is small enough to run on-device. We find that Fara-7B outperforms other CUA models of comparable size on benchmarks like WebVoyager, Online-Mind2Web, and WebTailBench -- our novel benchmark that better captures under-represented web tasks in pre-existing benchmarks. Furthermore, Fara-7B is competitive with much larger frontier models, illustrating key benefits of scalable data generation systems in advancing small efficient agentic models. We are making Fara-7B open-weight on Microsoft Foundry and HuggingFace, and we are releasing WebTailBench.

Erfan Shayegani, Keegan Hines, Yue Dong et al. · cmu, microsoft-research

Computer-Use Agents (CUAs) are an increasingly deployed class of agents that take actions on GUIs to accomplish user goals. In this paper, we show that CUAs consistently exhibit Blind Goal-Directedness (BGD): a bias to pursue goals regardless of feasibility, safety, reliability, or context. We characterize three prevalent patterns of BGD: (i) lack of contextual reasoning, (ii) assumptions and decisions under ambiguity, and (iii) contradictory or infeasible goals. We develop BLIND-ACT, a benchmark of 90 tasks capturing these three patterns. Built on OSWorld, BLIND-ACT provides realistic environments and employs LLM-based judges to evaluate agent behavior, achieving 93.75% agreement with human annotations. We use BLIND-ACT to evaluate nine frontier models, including Claude Sonnet and Opus 4, Computer-Use-Preview, and GPT-5, observing high average BGD rates (80.8%) across them. We show that BGD exposes subtle risks that arise even when inputs are not directly harmful. While prompting-based interventions lower BGD levels, substantial risk persists, highlighting the need for stronger training- or inference-time interventions. Qualitative analysis reveals observed failure modes: execution-first bias (focusing on how to act over whether to act), thought-action disconnect (execution diverging from reasoning), and request-primacy (justifying actions due to user request). Identifying BGD and introducing BLIND-ACT establishes a foundation for future research on studying and mitigating this fundamental risk and ensuring safe CUA deployment.

Suzanne Petryk, Spencer Whitehead, Joseph E. Gonzalez et al.

The ability to judge whether a caption correctly describes an image is a critical part of vision-language understanding. However, state-of-the-art models often misinterpret the correctness of fine-grained details, leading to errors in outputs such as hallucinating objects in generated captions or poor compositional reasoning. In this work, we explore Token-Level Confidence, or TLC, as a simple yet surprisingly effective method to assess caption correctness. Specifically, we fine-tune a vision-language model on image captioning, input an image and proposed caption to the model, and aggregate either algebraic or learned token confidences over words or sequences to estimate image-caption consistency. Compared to sequence-level scores from pretrained models, TLC with algebraic confidence measures achieves a relative improvement in accuracy by 10% on verb understanding in SVO-Probes and outperforms prior state-of-the-art in image and group scores for compositional reasoning in Winoground by a relative 37% and 9%, respectively. When training data are available, a learned confidence estimator provides further improved performance, reducing object hallucination rates in MS COCO Captions by a relative 30% over the original model and setting a new state-of-the-art.

Spencer Whitehead, Hui Wu, Heng Ji et al.

Generalization to out-of-distribution data has been a problem for Visual Question Answering (VQA) models. To measure generalization to novel questions, we propose to separate them into "skills" and "concepts". "Skills" are visual tasks, such as counting or attribute recognition, and are applied to "concepts" mentioned in the question, such as objects and people. VQA methods should be able to compose skills and concepts in novel ways, regardless of whether the specific composition has been seen in training, yet we demonstrate that existing models have much to improve upon towards handling new compositions. We present a novel method for learning to compose skills and concepts that separates these two factors implicitly within a model by learning grounded concept representations and disentangling the encoding of skills from that of concepts. We enforce these properties with a novel contrastive learning procedure that does not rely on external annotations and can be learned from unlabeled image-question pairs. Experiments demonstrate the effectiveness of our approach for improving compositional and grounding performance.

Spencer Whitehead, Hui Wu, Yi Ren Fung et al.

Existing Visual Question Answering (VQA) models are often fragile and sensitive to input variations. In this paper, we propose a novel approach to address this issue based on modular networks, which creates two questions related by linguistic perturbations and regularizes the visual reasoning process between them to be consistent during training. We show that our framework markedly improves consistency and generalization ability, demonstrating the value of controlled linguistic perturbations as a useful and currently underutilized training and regularization tool for VQA models. We also present VQA Perturbed Pairings (VQA P2), a new, low-cost benchmark and augmentation pipeline to create controllable linguistic variations of VQA questions. Our benchmark uniquely draws from large-scale linguistic resources, avoiding human annotation effort while maintaining data quality compared to generative approaches. We benchmark existing VQA models using VQA P2 and provide robustness analysis on each type of linguistic variation.

Boliang Zhang, Spencer Whitehead, Lifu Huang et al.

Many name tagging approaches use local contextual information with much success, but fail when the local context is ambiguous or limited. We present a new framework to improve name tagging by utilizing local, document-level, and corpus-level contextual information. We retrieve document-level context from other sentences within the same document and corpus-level context from sentences in other topically related documents. We propose a model that learns to incorporate document-level and corpus-level contextual information alongside local contextual information via global attentions, which dynamically weight their respective contextual information, and gating mechanisms, which determine the influence of this information. Extensive experiments on benchmark datasets show the effectiveness of our approach, which achieves state-of-the-art results for Dutch, German, and Spanish on the CoNLL-2002 and CoNLL-2003 datasets.

Manling Li, Alireza Zareian, Qi Zeng et al.

We introduce a new task, MultiMedia Event Extraction (M2E2), which aims to extract events and their arguments from multimedia documents. We develop the first benchmark and collect a dataset of 245 multimedia news articles with extensively annotated events and arguments. We propose a novel method, Weakly Aligned Structured Embedding (WASE), that encodes structured representations of semantic information from textual and visual data into a common embedding space. The structures are aligned across modalities by employing a weakly supervised training strategy, which enables exploiting available resources without explicit cross-media annotation. Compared to uni-modal state-of-the-art methods, our approach achieves 4.0% and 9.8% absolute F-score gains on text event argument role labeling and visual event extraction. Compared to state-of-the-art multimedia unstructured representations, we achieve 8.3% and 5.0% absolute F-score gains on multimedia event extraction and argument role labeling, respectively. By utilizing images, we extract 21.4% more event mentions than traditional text-only methods.

Maxwell Crouse, Ibrahim Abdelaziz, Bassem Makni et al.

Automated theorem provers have traditionally relied on manually tuned heuristics to guide how they perform proof search. Deep reinforcement learning has been proposed as a way to obviate the need for such heuristics, however, its deployment in automated theorem proving remains a challenge. In this paper we introduce TRAIL, a system that applies deep reinforcement learning to saturation-based theorem proving. TRAIL leverages (a) a novel neural representation of the state of a theorem prover and (b) a novel characterization of the inference selection process in terms of an attention-based action policy. We show through systematic analysis that these mechanisms allow TRAIL to significantly outperform previous reinforcement-learning-based theorem provers on two benchmark datasets for first-order logic automated theorem proving (proving around 15% more theorems).

Pavan Kapanipathi, Veronika Thost, Siva Sankalp Patel et al.

Textual entailment is a fundamental task in natural language processing. Most approaches for solving the problem use only the textual content present in training data. A few approaches have shown that information from external knowledge sources like knowledge graphs (KGs) can add value, in addition to the textual content, by providing background knowledge that may be critical for a task. However, the proposed models do not fully exploit the information in the usually large and noisy KGs, and it is not clear how it can be effectively encoded to be useful for entailment. We present an approach that complements text-based entailment models with information from KGs by (1) using Personalized PageR- ank to generate contextual subgraphs with reduced noise and (2) encoding these subgraphs using graph convolutional networks to capture KG structure. Our technique extends the capability of text models exploiting structural and semantic information found in KGs. We evaluate our approach on multiple textual entailment datasets and show that the use of external knowledge helps improve prediction accuracy. This is particularly evident in the challenging BreakingNLI dataset, where we see an absolute improvement of 5-20% over multiple text-based entailment models.

Qingyun Wang, Zhihao Zhou, Lifu Huang et al.

We present a paper abstract writing system based on an attentive neural sequence-to-sequence model that can take a title as input and automatically generate an abstract. We design a novel Writing-editing Network that can attend to both the title and the previously generated abstract drafts and then iteratively revise and polish the abstract. With two series of Turing tests, where the human judges are asked to distinguish the system-generated abstracts from human-written ones, our system passes Turing tests by junior domain experts at a rate up to 30% and by non-expert at a rate up to 80%.

Di Lu, Spencer Whitehead, Lifu Huang et al.

Current image captioning approaches generate descriptions which lack specific information, such as named entities that are involved in the images. In this paper we propose a new task which aims to generate informative image captions, given images and hashtags as input. We propose a simple but effective approach to tackle this problem. We first train a convolutional neural networks - long short term memory networks (CNN-LSTM) model to generate a template caption based on the input image. Then we use a knowledge graph based collective inference algorithm to fill in the template with specific named entities retrieved via the hashtags. Experiments on a new benchmark dataset collected from Flickr show that our model generates news-style image descriptions with much richer information. Our model outperforms unimodal baselines significantly with various evaluation metrics.