Adjacent
These are non-canonical entries — useful supporting context like reused foundation models, general multimodal methods, agent infrastructure, or non-browser search agents that materially inform GUI research. They are intentionally excluded from the main browse, the stats, and the per-environment / per-keyword groupings.
43 entries.
Ziyan Jiang, Manoj K. Ramanathan, Zheng Chen, Yilun Du +1 more
AgentVQA is a diagnostic visual benchmark probing visual understanding capabilities required by agents, including UI element localization, temporal motion tracking, and state change interpretation across 2,400 questions.
Yanzhe Chen, Ziyi Yang, Jifeng Zhu, Qiming Huang +13 more
Presents PAPAV, a capability-centric framework unifying multimodal computer-use agents and robotic systems across perception, anticipation, planning, action, and verification across 64 benchmarks.
Jiongxiao Wang, Han Ding, Youzhi Luo, Yawei Wang +3 more
Introduces Pitfall Anticipation World Model (PAWM), which selectively intervenes to predict impending failures within a bounded future horizon, improving Claude Sonnet 4.5 performance on WebArena-Lite and ScienceWorld.
Bytedance Seed
Seed1.8 is a foundation model for generalized real-world agency that unifies search, code generation and execution, and GUI interaction under one agentic interface. The model card emphasizes strong language-vision performance plus latency- and cost-aware inference modes for deployment.
Haozhan Shen, Shilin Yan, Hongwei Xue, Shuaiqi Lu +4 more
MM-CondChain is a benchmark for visually grounded deep compositional reasoning built from multi-layer conditional chains whose steps are programmatically verified through VPIR. It spans natural images, charts, and GUI trajectories, and shows that even the strongest MLLMs remain weak on deep chained reasoning.
Ching-Yu Kao, Xinfeng Li, Shenyu Dai, Tianze Qiu +3 more
This paper studies documentation-embedded instruction injection in high-privilege LLM agents and frames the failure mode as the Trusted Executor Dilemma. It introduces ReadSecBench, shows exfiltration success up to 85%, and finds that both rule-based and LLM-based defenses still fail to catch the attacks reliably.
Yinjie Wang, Xuyang Chen, Xiaolong Jin, Mengdi Wang +1 more
OpenClaw-RL is an asynchronous RL framework that treats next-state signals from live interactions as a universal learning source. It combines scalar rewards from a process-reward judge with hindsight-guided on-policy distillation, and trains agents across conversations, terminals, GUI tasks, SWE, and tool use in one online loop.
Ido Levy, Segev Shlomov, Avi Yaeli, Nir Mashkif +4 more
TabAgent replaces an LLM agent's repeated closed-set decisions, such as tool shortlisting, skill selection, routing and verification, with a small CatBoost classifier trained on features extracted from the agent's own successful execution traces. It is evaluated on seven agent benchmarks including AppWorld, tau2-bench, SkillsBench and Mind2Web, and on AppWorld tool shortlisting it recovers essentially all tools the GPT-5.4 shortlister used on successful runs (Recall@9) at about 1,000x lower latency and 100,000x lower cost per decision.
Renjun Xu, Yang Yan
This survey reviews the emerging agent-skills ecosystem for LLMs, covering architectural foundations such as SKILL.md and MCP, methods for acquiring and refining skills, deployment patterns for agent systems, and the security problems introduced by portable, dynamically loaded capabilities.
Kaichen He, Zihao Wang, Muyao Li, Anji Liu +1 more
CrossAgent studies how a single agent model can switch among heterogeneous action spaces, including APIs, GUI events, and lower-level commands, without hand-written routing rules. Its training pipeline combines supervised fine-tuning with multi-turn GRPO and reports state-of-the-art results on 800+ Minecraft tasks, making it relevant to GUI work as a broader action-space unification result rather than a direct GUI paper.
Shuyue Hu, Haoyang Yan, Yiqun Zhang, Yang Chen +2 more
This paper argues that stronger single-agent foundation models do not automatically become strong multi-agent systems, and evaluates 41 open models on seven single-agent and multi-agent benchmarks to show the gap directly. It uses GUI interaction as one example of native single-agent capability, but its main contribution is a broader multi-agent intelligence agenda rather than GUI research itself.
Yueqi Song, Ketan Ramaneti, Zaid Sheikh, Ziru Chen +17 more
Agent Data Protocol (ADP) standardizes heterogeneous agent trajectories into a lightweight schema and conversion pipeline so diverse agent datasets can plug into multiple SFT pipelines without per-dataset engineering. Converting 13 existing datasets into ADP and fine-tuning on the unified corpus improves base models by about 20% on average across coding, browsing, tool-use, and research benchmarks.
Shrey Pandit, Xuan-Phi Nguyen, Yifei Ming, Austin Xu +3 more
This paper synthesizes training data for deep-research web agents by progressively increasing question difficulty until a baseline agent fails, then using that agent again for validation and filtering. The resulting corpus is aimed at long-horizon online-tool use rather than browser-native GUI interaction, but it is relevant as adjacent training-data work for agent systems.
Sayash Kapoor, Benedikt Stroebl, Peter Kirgis, Nitya Nadgir +27 more
HAL provides standardized infrastructure for evaluating agents across models, scaffolds, and benchmarks rather than introducing a new agent. It reports results from 21,730 rollouts across 9 models and 9 benchmarks, tracks costs and full traces, and uses LLM-aided log inspection to surface behaviors such as benchmark gaming and unsafe actions.
Neil Kale, Chen Bo Calvin Zhang, Kevin Zhu, Ankit Aich +4 more
Stress-tests LLM agent monitoring systems for detecting covert misbehavior using a monitor red-teaming (MRT) workflow varying agent/monitor awareness and adversarial evasion strategies, evaluated on SHADE-Arena for tool-calling agents and CUA-SHADE-Arena for computer-use agents.
Ziyang Luo, Zhiqi Shen, Wenzhuo Yang, Zirui Zhao +6 more
MCP-Universe introduces the first comprehensive benchmark for evaluating large language models (LLMs) through interactions with real-world Model Context Protocol (MCP) servers. It spans six core domains—Location Navigation, Repository Management, Financial Analysis, 3D Design, Browser Automation, and Web Searching—across 11 MCP servers. The benchmark employs execution-based evaluators (format, static, dynamic) to rigorously assess agent performance. Despite progress, state-of-the-art models like GPT-5 (43.72% success), Grok-4 (33.33%), and Claude-4.0-Sonnet (29.44%) show significant limitations. The benchmark highlights challenges in long-context reasoning and unfamiliar tool handling, and provides an open-source extensible evaluation framework with UI support to accelerate future research.
Zijian Chen, Xueguang Ma, Shengyao Zhuang, Ping Nie +16 more
Introduces **BrowseComp-Plus**, a fixed-corpus benchmark for evaluating deep-research agents. It enables controlled, fair, and transparent comparisons by providing human-verified supporting and challenging negative documents for each query. Results reveal significant performance variation—for example, an open-source model (Search-R1 + BM25) only achieves 3.86% accuracy, while GPT-5 reaches 55.9%, and GPT-5 with Qwen3-Embedding-8B retriever achieves 70.1% with fewer queries—highlighting the critical importance of retrieval quality and enabling disentangled analysis of retrieval vs. reasoning components.
Tianqing Fang, Zhisong Zhang, Xiaoyang Wang, Rui Wang +9 more
This work introduces **Cognitive Kernel-Pro**, a fully open-source, multi-module agent framework designed to democratize advanced AI agent development. It curates high-quality training data across four domains—web, files, code, and general reasoning—and introduces test-time strategies like reflection and voting to enhance robustness. Evaluated on the GAIA benchmark, its open 8B-parameter model outperforms previous open-source agents such as WebDancer and WebSailor, setting a new performance standard. Code is publicly available.
Kyudan Jung, Hojun Cho, Jooyeol Yun, Soyoung Yang +2 more
Talk to Your Slides targets slide editing through language-driven manipulation of the underlying document object model rather than GUI-native visual interaction. It is relevant to GUI work because it compares against GUI-based baselines and introduces the TSBench benchmark, but its primary interaction mechanism is structured document editing rather than direct GUI control.
Jianwei Yang, Reuben Tan, Qianhui Wu, Ruijie Zheng +9 more
Magma is a multimodal foundation model for agentic tasks spanning digital and physical environments rather than a GUI-specific paper. It is relevant here because it reports strong UI navigation results and uses Set-of-Mark and Trace-of-Mark supervision, but its main contribution is a broader agentic model covering robotics as well as GUI tasks.
Yu Su, Diyi Yang, Shunyu Yao, Tao Yu
This tutorial provides a comprehensive exploration of language agents—autonomous systems powered by large language models capable of executing complex tasks through language instructions. It delves into their theoretical foundations, potential applications, associated risks, and future directions, covering topics such as reasoning, memory, planning, tool augmentation, grounding, multi-agent systems, and safety considerations.
Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter +19 more
This paper introduces MM1.5, a family of multimodal large language models (MLLMs) ranging from 1B to 30B parameters, including dense and mixture-of-experts variants. MM1.5 enhances capabilities in text-rich image understanding, visual referring and grounding, and multi-image reasoning. The authors employ a data-centric training approach, utilizing high-quality OCR data and synthetic captions for continual pre-training, alongside an optimized visual instruction-tuning data mixture for supervised fine-tuning. Specialized variants, MM1.5-Video and MM1.5-UI, are designed for video understanding and mobile UI comprehension, respectively. Extensive empirical studies provide insights into the training processes, offering guidance for future MLLM development.
Lutfi Eren Erdogan, Nicholas Lee, Siddharth Jha, Sehoon Kim +6 more
TinyAgent is an edge deployment framework for small function-calling language models, paired with a curated training dataset, tool retrieval, and quantization for local inference. Its GUI relevance comes mainly from the MacBook assistant demo and local agent deployment setting, not from a primary contribution to GUI interaction research itself.
Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong +26 more
VisualAgentBench benchmarks large multimodal models as general visual foundation agents across embodied tasks, GUI tasks, and visual design rather than focusing only on GUI interaction. It also releases trajectory data for behavior cloning, making it relevant to GUI work as a broader visual-agent benchmark rather than a direct GUI paper.
Xinbei Ma, Yiting Wang, Yao Yao, Tongxin Yuan +3 more
This paper highlights the vulnerability of multimodal agents to environmental distractions. The researchers demonstrate that these agents, which process multiple types of input (e.g., text, images, audio), can be significantly impacted by irrelevant or misleading environmental cues. The study provides insights into the limitations of current multimodal systems and emphasizes the need for more robust architectures that can filter out distractions and maintain focus on relevant information in complex, real-world environments.
Zehui Chen, Kuikun Liu, Qiuchen Wang, Jiangning Liu +3 more
This paper presents MindSearch, a novel approach to web information seeking and integration that mimics human cognitive processes. The system uses a multi-agent framework consisting of a WebPlanner and WebSearcher. The WebPlanner models multi-step information seeking as a dynamic graph construction process, decomposing complex queries into sub-questions. The WebSearcher performs hierarchical information retrieval for each sub-question. MindSearch demonstrates significant improvements in response quality and depth compared to existing AI search solutions, processing information from over 300 web pages in just 3 minutes.
Presents Octo-planner, an on-device planner for a planner-action agent framework that separates task decomposition from action execution. Built on Phi-3 Mini and paired with an Octopus action model, it targets low-latency planning and execution on resource-constrained devices.
Gabriel Herbert Sarch, Lawrence Jang, Michael J. Tarr, William W. Cohen +2 more
ICAL turns sub-optimal demonstrations and feedback into reusable multimodal memories that improve VLM and LLM agents across TEACh, VisualWebArena, and Ego4D. It is relevant to GUI work because one evaluation domain is web agents, but the method itself is a broader embodied-agent memory approach rather than a direct GUI paper.
Wei Chen, Zhiyuan Li
Octopus v3 is a sub-billion multimodal AI agent model designed for efficient on-device deployment, with the paper centered on its functional-token mechanism and edge-device constraints rather than GUI-native interaction. It is relevant to GUI research as a lightweight multimodal agent backbone, but it is broader than a direct GUI-agent paper.
Lei Ding, Jeshwanth Bheemanpally, Yi Zhang
Targets mobile "how-to" search by automatically executing step-by-step instructions from retrieved pages in a controlled Android environment and reranking results based on actual success. The paper frames this as a verification-driven ranking pipeline for technical-support search rather than a pure mobile-control benchmark.
Longtao Zheng, Zhiyuan Huang, Zhenghai Xue, Xinrun Wang +2 more
AgentStudio packages environments, tools, benchmarks, and datasets for general virtual agents with mixed GUI and API action spaces. It is relevant here because GUI interaction is one supported modality and because it contributes GroundUI, IDMBench, and CriticBench, but the paper is broader than a direct GUI-agent study.
Weihao Tan, Wentao Zhang, Xinrun Xu, Haochong Xia +24 more
Cradle formulates general computer control as screenshot input plus keyboard-and-mouse output, and instantiates that setting with a modular multimodal agent for software and video games. It matters to GUI work because it demonstrates screen-only control on real software and evaluates on OSWorld, but the paper is framed as a broader general-computer-control agenda rather than a direct GUI paper.
Tianyu Gao, Zirui Wang, Adithya Bhaskar, Danqi Chen
This paper studies screenshot language models in a simplified plain-text-rendered setting and improves them with a patch-and-text prediction objective. It is relevant here because screenshot pretraining can transfer to UI-style inputs, but the paper is about general screenshot language understanding rather than direct GUI-agent behavior.
Wenkai Yang, Xiaohan Bi, Yankai Lin, Sishuo Chen +2 more
This paper analyzes backdoor attacks against generic LLM-based agents, including attacks that trigger from user queries or intermediate observations and attacks that alter intermediate reasoning while preserving the final answer. It matters for GUI work because web-shopping agents are one evaluation setting, but the contribution is a broader LLM-agent security analysis rather than a GUI-specific study.
Lingbo Mo, Zeyi Liao, Boyuan Zheng, Yu Su +2 more
Maps adversarial attacks on language agents through a Perception-Brain-Action decomposition and surveys 12 attack types across those layers. The paper is mainly a threat-modeling taxonomy, useful as a security lens for later web and computer-use agents.
Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu +18 more
Introduces AgentBench, a broad benchmark suite for evaluating LLMs as agents across eight environments, including but not limited to OS interaction. It belongs in this repo mainly as a general agent-evaluation reference rather than a pure GUI benchmark.
Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf +2 more
GAIA is a benchmark developed for evaluating general-purpose AI assistants. It aims to test assistant models across multiple modalities and complex reasoning tasks in real-world settings, including scenarios that require tool usage and open-ended question answering. With a dataset comprising 466 questions across various domains, GAIA highlights gaps between current AI performance and human capability, presenting a significant challenge for large language models such as GPT-4.
Yuwen Lu, Ziang Tong, Qinyi Zhao, Chengzhi Zhang +1 more
This position paper studies mobile UI layout generation with LLMs and proposes a UI grammar to represent hierarchical screen structure. Its focus is controllable UI generation rather than GUI action or environment interaction, so it stays outside the direct-GUI agent main list.
Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou +2 more
Introduces Set-of-Mark prompting, where segmented image regions are overlaid with explicit marks before being passed to a multimodal model. The paper shows that simple region marking can unlock much stronger zero-shot grounding from GPT-4V without fine-tuning.
Tianbao Xie, Fan Zhou, Zhoujun Cheng, Peng Shi +12 more
OpenAgents is a deployment-oriented platform for language agents in everyday use, spanning data analysis, API-tool use, and web browsing behind a shared user interface. It matters here because it includes a web-agent mode, but the paper is broader than direct GUI-agent research.
Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan +1 more
Introduces Reflexion, a framework where language agents learn from feedback by writing verbal reflections into episodic memory instead of updating model weights. The method is broad rather than GUI-specific, but it matters here because many later web and computer-use agents inherit its reflection-and-memory pattern.
Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu +6 more
Pix2Struct pretrains an image-to-text model by reconstructing simplified HTML from masked webpage screenshots, then transfers it across documents, illustrations, user interfaces, and natural images. It matters for GUI research because it improves UI understanding, but the paper is broader than direct GUI-agent work.
Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du +3 more
Introduces ReAct, the now-standard prompting pattern that interleaves reasoning traces with actions. It is not a GUI paper, but it is foundational because many later web and computer-use agents build directly on this reasoning-action decomposition.