/research

To run an end-to-end research pipeline, provide a research direction and execute the workflow in automatic mode with the --auto flag. The system orchestrates ideation, experiment design, async execution, evaluation, and paper drafting into a resumable wiki.

The best way to manage multiple controlled ML experiments across sessions is saving pipeline state to pipeline-progress.md. This research orchestration approach allows resuming staged experiments from specific control points using a start-from flag without losing previous results.

Start a research workflow for a topic like sparse LoRA by inputting a one-sentence direction and specifying a target venue using the --venue flag. The pipeline auto-bootstraps your wiki with paper ingestion, then generates, tests, and iterates on experiments to produce a venue-targeted draft.

Yes, you can run async experiment execution and collect results later. Stage 3 supports non-blocking deployments, allowing experiments to execute asynchronously while the research pipeline collects evaluation results at a later control point.

You need a RESEARCH_BRIEF.md file or a simple direction text input to bootstrap the research orchestration pipeline. This initial input defines the research direction, allowing the system to automatically ingest papers and generate experiment designs.

Stage recovery works when an ML experiment pipeline is interrupted by saving workflow state to pipeline-progress.md. You can resume the interrupted research orchestration from the exact stage using the --start-from flag without re-running completed experiment stages.