QLWF platform turns research notes into executable workflows with 98.4% accuracy

Quantitative language research often relies on long chains of computational steps, but the logic connecting those steps usually stays buried in scripts, making analyses hard to inspect, share, and revise. To address this, He Wang and five co-authors introduce QLWF, a visual workflow platform that turns natural-language research descriptions into executable workflows via an AI-assisted five-stage pipeline. The platform 'reifies' research logic as a visible workflow, then 'formalizes' it with deterministic execution semantics. Crucially, the language model is used only during construction—execution runs on a fixed node library and engine, preserving reproducibility and preventing the LLM from introducing untracked variability.

The researchers evaluated QLWF using QL-Bench, a custom 64-task benchmark drawn from quantitative-syntax literature. Across three runs, QLWF produced structurally valid and executable workflows for every task, achieving a mean output-plausibility rate of 98.4%—well above prompt-based baselines. They also tested incremental refinement on a separate 12-task lifecycle benchmark: QLWF successfully revised saved workflows in all cases, consuming roughly one-third of the tokens needed for full regeneration. The paper releases the node library, benchmark, workflow templates, and platform as open resources, offering computational linguists a more transparent and efficient way to design and iterate complex analyses.