Jack (Hao) Bai

Training vision-language web agents with multi-step reinforcement learning is compute-intensive. We introduce AsyncWebRL, an asynchronous training system that overlaps rollout, gradient update, and policy refresh, together with a trajectory-level objective that removes an efficiency-inducing bias in multi-step GRPO. AsyncWebRL improves end-to-end training throughput and sets a new open-source state of the art on the WebGym out-of-distribution test split.

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.

Most paradigms for building foundation model agents rely on prompting or finetuning on existing demonstrations, but this is not sufficient in dynamic environments (e.g., mobile device control). In theory, while on-policy reinforcement learning (RL) should address these limitations, this approach itself is not quite effective at leveraging existing agentic data, especially when it is of low quality. An approach to address this issue is to use offline value-based RL but realizing value-based RL for agents has been elusive due to of stability and efficiency associated with running TD-learning at scale with vision-language models (VLMs). In this paper, we develop a scalable value-based RL approach called Digi-Q that makes it possible to train VLM agents with TD-learning. We situate our study in building GUI agents for Android devices. The key idea in Digi-Q is to perform TD-learning on a frozen, intermediate-layer representation of a VLM rather than training the whole VLM itself. Doing so successfully requires an initial phase of fine-tuning to prime VLM representations to feature actionable information that is critical for TD-learning. When done correctly, our approach is able to attain better performance per-unit compute FLOPS. To make maximal use of the learned Q-function, we devise a novel best-of-N policy extraction operator that imitates the best actions out of multiple candidate actions from the current policy as ranked by the value function. With no REINFORCE-style policy gradients that need careful tiuning and an efficient TD-learning approach, Digi-Q outperforms several strong prior methods on user-scale device control tasks in Android-in-the-Wild, attaining 9.9% of relative improvement over prior best-performing offline RL method in this domain.

Most paradigms for building foundation model agents rely on prompting or finetuning on existing demonstrations, but this is not sufficient in dynamic environments (e.g., mobile device control). In theory, while on-policy reinforcement learning (RL) should address these limitations, this approach itself is not quite effective at leveraging existing agentic data, especially when it is of low quality. An approach to address this issue is to use offline value-based RL but realizing value-based RL for agents has been elusive due to of stability and efficiency associated with running TD-learning at scale with vision-language models (VLMs). In this paper, we develop a scalable value-based RL approach called Digi-Q that makes it possible to train VLM agents with TD-learning. We situate our study in building GUI agents for Android devices. The key idea in Digi-Q is to perform TD-learning on a frozen, intermediate-layer representation of a VLM rather than training the whole VLM itself. Doing so successfully requires an initial phase of fine-tuning to prime VLM representations to feature actionable information that is critical for TD-learning. When done correctly, our approach is able to attain better performance per-unit compute FLOPS. To make maximal use of the learned Q-function, we devise a novel best-of-N policy extraction operator that imitates the best actions out of multiple candidate actions from the current policy as ranked by the value function. With no REINFORCE-style policy gradients that need careful tiuning and an efficient TD-learning approach, Digi-Q outperforms several strong prior methods on user-scale device control tasks in Android-in-the-Wild, attaining 9.9% of relative improvement over prior best-performing offline RL method in this domain.

Neurons in auto-regressive language models like GPT-2 can be interpreted by analyzing their activation patterns. Recent studies have shown that techniques such as dictionary learning, a form of post-hoc sparse coding, enhance this neuron-level interpretability. In our research, we are driven by the goal to fundamentally improve neural network interpretability by embedding sparse coding directly within the model architecture, rather than applying it as an afterthought. In our study, we introduce a white-box transformer-like architecture named Coding RAte TransformEr (CRATE), explicitly engineered to capture sparse, low-dimensional structures within data distributions. Our comprehensive experiments showcase significant improvements (up to 103% relative improvement) in neuron-level interpretability across a variety of evaluation metrics. Detailed investigations confirm that this enhanced interpretability is steady across different layers irrespective of the model size, underlining CRATE’s robust performance in enhancing neural network interpretability. Further analysis shows that CRATE’s increased interpretability comes from its enhanced ability to consistently and distinctively activate on relevant tokens. These findings point towards a promising direction for creating white-box foundation models that excel in neuron-level interpretation.

Training corpuses for vision language models typically lack sufficient amounts of decision-centric data. This renders off-the-shelf VLMs sub-optimal for decision-making tasks such as in-the-wild device control through graphical user interfaces (GUIs). While training with static demonstrations has shown some promise, we show that such methods fall short when controlling real GUIs due to their failure to deal with real world stochasticity not captured in static observational data. This paper introduces a novel autonomous RL approach, called DigiRL, for training in-the-wild device control agents through fine-tuning a pre-trained VLM in two stages: offline RL to initialize the model, followed by offline-to-online RL. To do this, we build a scalable and parallelizable Android learning environment equipped with a VLM-based evaluator and develop a simple yet effective RL approach for learning in this domain. Our approach runs advantage-weighted RL with advantage estimators enhanced to account for stochasticity along with an automatic curriculum for deriving maximal learning signal. We demonstrate the effectiveness of DigiRL using the Android-in-the-Wild (AitW) dataset, where our 1.5B VLM trained with RL achieves a 49.5% absolute improvement – from 17.7% to 67.2% success rate – over supervised fine-tuning with static human demonstration data. These results significantly surpass not only the prior best agents, including AppAgent with GPT-4V (8.3% success rate) and the 17B CogAgent trained with AitW data (14.4%), but also the prior best autonomous RL approach based on filtered behavior cloning (57.8%), thereby establishing a new state-of-the-art for digital agents for in-the-wild device control.

Large vision-language models (VLMs) fine-tuned on specialized visual instruction-following data have exhibited impressive language reasoning capabilities across various scenarios. However, this fine-tuning paradigm may not be able to efficiently learn optimal decision-making agents in multi-step goal-directed tasks from interactive environments. To address this challenge, we propose an algorithmic framework that fine-tunes VLMs with reinforcement learning (RL). Specifically, our framework provides a task description and then prompts the VLM to generate chain-of-thought (CoT) reasoning, enabling the VLM to efficiently explore intermediate reasoning steps that lead to the final text-based action. Next, the open-ended text output is parsed into an executable action to interact with the environment to obtain goal-directed task rewards. Finally, our framework uses these task rewards to fine-tune the entire VLM with RL. Empirically, we demonstrate that our proposed framework enhances the decision-making capabilities of VLM agents across various tasks, enabling 7b models to outperform commercial models such as GPT4-V or Gemini. Furthermore, we find that CoT reasoning is a crucial component for performance improvement, as removing the CoT reasoning results in a significant decrease in the overall performance of our method.

In this paper, we contend that a natural objective of representation learning is to compress and transform the distribution of the data, say sets of tokens, towards a low-dimensional Gaussian mixture supported on incoherent subspaces. The goodness of such a representation can be evaluated by a principled measure, called sparse rate reduction, that simultaneously maximizes the intrinsic information gain and extrinsic sparsity of the learned representation. From this perspective, popular deep network architectures, including transformers, can be viewed as realizing iterative schemes to optimize this measure. Particularly, we derive a transformer block from alternating optimization on parts of this objective: the multi-head self-attention operator compresses the representation by implementing an approximate gradient descent step on the coding rate of the features, and the subsequent multi-layer perceptron sparsifies the features. This leads to a family of white-box transformer-like deep network architectures, named CRATE, which are mathematically fully interpretable. We show, by way of a novel connection between denoising and compression, that the inverse to the aforementioned compressive encoding can be realized by the same class of CRATE architectures. Thus, the so-derived white-box architectures are universal to both encoders and decoders. Experiments show that these networks, despite their simplicity, indeed learn to compress and sparsify representations of large-scale real-world image and text datasets, and achieve performance very close to highly engineered transformer-based models: ViT, MAE, DINO, BERT, and GPT2. We believe the proposed computational framework demonstrates great potential in bridging the gap between theory and practice of deep learning, from a unified perspective of data compression.

Open-domain dialog involves generating search queries that help obtain relevant knowledge for holding informative conversations. However, it can be challenging to determine what information to retrieve when the user is passive and does not express a clear need or request. To tackle this issue, we present a novel approach that focuses on generating internet search queries that are guided by social commonsense. Specifically, we leverage a commonsense dialog system to establish connections related to the conversation topic, which subsequently guides our query generation. Our proposed framework addresses passive user interactions by integrating topic tracking, commonsense response generation and instructiondriven query generation. Through extensive evaluations, we show that our approach1 overcomes limitations of existing query generation techniques that rely solely on explicit dialog information, and produces search queries that are more relevant, specific, and compelling, ultimately resulting in more engaging responses.