implementation detail · filed under data curation
Sandboxed visual reasoning with a Python interpreter
Visual-reasoning trajectories are generated by an agent that writes and executes Python code in an isolated sandbox to manipulate images and verify results.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
Each visual-reasoning trajectory is generated in an agent environment equipped with a Python interpreter in an isolated sandbox: the model iteratively writes and executes code to crop, zoom, or transform the input image, perform precise computation, or verify intermediate results
usedunclearin Kimi K3Moonshot AI
Filed alongside
Other methods under data curation :: synthetic data.
Agentic data synthesis pipelineKnowledge distillation for synthetic dataAutomated data synthesis and environment-construction pipelinesAutomatic synthesis of task-oriented RL environmentsContainerized coding-environment construction with self-testing and trace removalCounterfactual data augmentationRecycling filtered documents into image-text pairsAutomated batch synthesis of RL training dataCoding environments from GitHub issue–PR pairsContainer buildability and verifiability checkCorpus rephrasing with fidelity verificationDynamic multi-agent synthetic-data generation loopEnd-to-end synthetic environment generationEnvironment, toolset, task, and solution synthesis pipelineFailure-case and negative-feedback-driven environment generationHeterogeneous answer-generation agentsIterative task-difficulty escalationKnowledge-graph-guided task synthesisLarge-scale environment synthesis and curationLong-context data synthesis by permutation and concatenationMatching synthesis-pipeline complexity to teacher capabilityMetadata-conditioned synthetic generationMocked tools for agent environmentsModular synthetic-data pipeline composition