What problem does it solve? Going from a broad research direction to a validated, publication-worthy idea requires literature surveying, brainstorming, novelty verification, critical review, and pilot experiments — a process that is slow, unstructured, and easy to get wrong when done manually. ## Core Features & Use Cases - End-to-End Pipeline Orchestration: Chains literature survey, idea generation, novelty checking, external review, and method refinement into one automated workflow with checkpoints. - Empirical Validation: Runs bounded pilot experiments (max 2 hours per GPU, 8 GPU-hours total) on the top 2-3 ideas and ranks them by empirical signal rather than theoretical appeal. - Structured Deliverables: Produces a ranked IDEA_REPORT.md, a refined FINAL_PROPOSAL.md, and a claim-driven EXPERIMENT_PLAN.md ready for execution. - Use Case: A researcher says "find ideas for efficient video tokenization" and receives a literature landscape, 8-12 brainstormed ideas filtered by feasibility, pilot results on the top candidates, a senior-reviewer critique, and a concrete experiment roadmap. ## Quick Start Run the idea discovery pipeline on the research direction "machine-oriented video compression" and proceed automatically through each checkpoint.
