OSWorld
PulseAugur coverage of OSWorld — every cluster mentioning OSWorld across labs, papers, and developer communities, ranked by signal.
5 day(s) with sentiment data
-
AI agents: Debugging frameworks offer repair over regeneration
Two research papers propose novel debugging frameworks for AI agents, shifting focus from regeneration to repair. The first, CUADebug, targets failures in computer-use agents by analyzing visual perception, spatial grou…
-
Chinese AI Agent InAgent Sets New Record on OSWorld Benchmark
InAgent, a Chinese AI agent, has achieved a record 90.2% success rate on the OSWorld benchmark, making it the first agent to surpass the 90% threshold. This performance exceeds previous records set by major AI labs like…
-
2026 LLM Benchmark: No Single Winner, Specialized Leaders Emerge · 1 source tracked
A comprehensive benchmark of 20 leading LLMs in 2026 reveals no single dominant model, but rather specialized leaders across different tasks. Claude Opus 5 leads the overall Artificial Analysis Intelligence Index, while…
-
New Interactive Reward Agent Achieves 86.9% Accuracy in GUI Task Evaluation
Researchers have developed an Interactive Reward Agent (IRA) designed to improve the evaluation of GUI agents. This agent utilizes a propose-then-verify framework to gather and confirm evidence from the environment's st…
-
AI agents struggle with visual tasks; EvoCUA-1.5 hits 63.2% on OSWorld benchmark
A new benchmark called OSWorld has revealed that the AI agent EvoCUA-1.5 achieved a score of 63.2% by learning through trial-and-error in simulated environments. Separately, a new open benchmark developed by Apple Inc. …
-
New BREW framework enables LLM agents to learn from experience
Researchers have developed BREW, a novel framework designed to enable Large Language Model (LLM)-based agents to learn from past experiences. Unlike current agents that restart learning with each session, BREW distills …
-
AI agent benchmarks now include cost data, revealing massive price disparities
A new dataset has been created to track the cost of AI agent performance on various benchmarks, addressing a gap in existing leaderboards that primarily focus on scores. This dataset connects agent configurations, bench…
-
New UI-MOPD method enables cross-platform GUI agent learning · 3 sources tracked
Researchers have developed UI-MOPD, a novel method for training GUI agents that can operate across multiple platforms. This approach addresses the scarcity of cross-platform data and the challenge of maintaining perform…
-
New GUIDE framework reduces domain bias in GUI agents using video retrieval
Researchers have developed GUIDE, a novel framework designed to mitigate domain bias in GUI agents. This plug-and-play system leverages real-time web video retrieval and an automated annotation pipeline to equip agents …
-
New RL framework uses vision-language models for GUI agent supervision
Researchers have developed a new reinforcement learning framework for Computer-Use Agents (CUAs) that leverages autonomous vision-language evaluation for supervision. This approach addresses the challenge of obtaining s…
-
AI agents achieve 66% success on desktop tasks, but data gaps remain a challenge
Computer-use agents have shown significant progress, with success rates on the OSWorld benchmark jumping from 12% to 66% in about a year. This rapid advancement was highlighted by Microsoft's Build 2026 keynote, which p…
-
New benchmarks probe AI agent safety against deceptive interfaces and unsafe actions
Two new research papers introduce benchmarks for evaluating the safety of AI agents. OSGuard focuses on computer-use agents, distinguishing between safe and unsafe actions and identifying latent hazards in task executio…
-
New ProCUA-SFT dataset boosts AI agent desktop performance
Researchers have developed ProCUA-SFT, a new dataset designed to improve the training of computer-use agents (CUAs) that interact with graphical desktop environments. Existing datasets like AgentNet have shown negative …
-
AI Memory Systems Can Harm Performance, Research Finds
New research indicates that AI memory systems, while intended to improve user experience and task completion, can paradoxically degrade model performance and foster sycophantic tendencies. Studies show that these system…
-
New MacArena benchmark tests AI agents on macOS
Researchers have developed MacArena, a new benchmark designed to evaluate computer-use agents (CUAs) operating within a macOS environment. This benchmark includes 421 tasks across 50 applications, specifically tailored …
-
Hcompany ships Holo3.1 agents for fast, local computer use
Hcompany has released Holo3.1, a new family of computer-use agents designed for robust performance across various environments and agent frameworks. This release emphasizes local inference capabilities, offering quantiz…
-
New benchmark and data synthesis boost GUI agent error recovery
Researchers have developed a new benchmark and data synthesis framework to improve the error recovery capabilities of GUI agents. The benchmark, GUI-RobustEval, includes over 1,200 test cases to systematically measure h…
-
New frameworks aim to improve AI understanding of user intent
Two new research papers introduce computational frameworks for understanding and controlling user intent in AI interactions. The first, 'Intent Signal Theory,' formalizes the distinction between a user's latent intent a…
-
AI safety evaluations face 'safe-to-dangerous shift' challenge
A fundamental challenge in AI safety is the "safe-to-dangerous shift," which complicates realistic evaluations of AI models. This shift arises because alignment evaluations must be safe, limiting AI capabilities, while …
-
Holo1: New family of GUI automation VLMs powering GUI agent Surfer-H
Researchers have introduced A11y-Compressor, a framework designed to make GUI agent observations more efficient by transforming linearized accessibility trees into structured representations. This method reduces input t…