ISE: New Dataset and Method for Training Multi-Turn Operating System Agents
Researchers introduced ISE (Intent → Simulate → Execute), a three-stage synthesis method for creating training data for OS agents that can handle multi-turn tasks with real tool execution. The approach generated over 43,000 unique intents and 23,000 complete trajectories with authentic failure-recovery dynamics by running tool calls in isolated OS environments. Fine-tuning on this dataset improved performance on agent tool-use tasks, with a smaller 8B model outperforming larger base models and zero-shot GPT-4o.
Researchers at arXiv published a paper describing ISE, a novel paradigm for generating training data for operating system agents capable of handling complex, multi-turn tasks. The method operates in three stages: first constructing approximately 50,000 structured user intents across a 4D framework (Persona × Domain × Task × Complexity), reducing to 43,956 unique intents; second, simulating multi-turn user-agent interactions grounded in actual execution outcomes, producing 23,132 trajectories averaging 8.12 user turns; and third, executing every tool call in live, isolated OS workspaces to capture authentic failure-recovery dynamics. When fine-tuning the Qwen3-8B model on the resulting ISETrace dataset, researchers achieved a ClawEval pass@1 score of 37.7, up from 19.3 for the base model, surpassing zero-shot GPT-4o performance and outperforming the larger Qwen3-32B base model. The authors released source code and the dataset publicly, with ablation studies confirming that multi-turn simulation contributed substantially to performance gains.
What's missing
The paper does not discuss potential limitations of the ISE approach, such as scalability constraints, generalization to OS environments beyond those tested, or computational costs of the three-stage synthesis process. Additionally, the study does not address how the method performs on out-of-distribution tasks or provide error analysis of failure cases.
What different sources said
- arXiv cs.AICenter
ISE: An Execution-Grounded Recipe for Multi-Turn OS-Agent Trajectories
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.