--- title: "Can a small AI learn to draw workflows that run?" canonical: https://papersays.com/p/can-a-small-ai-learn-to-draw-workflows-that-run date: 2026-10-18 modified: 2026-10-18 hub: "AI & Machine Learning" doi: "10.1007/s44311-025-00034-4" license: "CC BY 4.0" --- # Can a Small AI Learn to Draw Workflows That Run? > A code model broke most attempts to turn a written procedure into a working process map. Rewards a program can check cut failures to almost zero. Companies describe how work gets done in plain text, but software that analyses and automates that work needs a formal process model. This study trained an open code model to write such models as short programs, using reinforcement learning with an AI judge first and then rewards a program can check. On 312 held-out descriptions, invalid outputs fell from 889 of 1,248 to 4. On a 20-description outside benchmark, its best-of-four score averaged 0.734, close to the roughly 0.74 reported for GPT-4o, and it failed far less often than larger open models. ## How was the AI trained to write process models? The authors started from an open, code-oriented language model. For each plain-text process description it writes a short Python program that builds a POWL process model, which can be converted to Petri nets or BPMN. They used 1312 descriptions with reference models: 1,000 for training and 312 held out for testing. Training used Group Relative Policy Optimization (GRPO). For each description the model writes four candidates; structural validators reject broken ones, and candidates that score above the group's average are reinforced. Stage A scored candidates with an AI judge. Stage B rewarded agreement of behavioural footprints, which record which steps can directly follow each other and which can run in parallel, with the reference model. ## How much did the training reduce broken process models? On the 312 test descriptions, with four samples each (1,248 completions), the base model gave 889 lowest-reward completions. Supervised fine-tuning brought this to 328, the judge stage to 152 and the footprints stage to 4. On the ProMoAI benchmark of 20 descriptions, the trained checkpoint averaged a best-of-four replay F-score of 0.734, close to the roughly 0.74 reported for GPT-4o, and a worst-of-four of 0.398. Qwen2.5-Coder-32B reached a slightly higher best-of-four (0.754) but a worst-of-four of only 0.046, with three descriptions where every run scored zero. A complexity reward worked as a dial: pushing it up raised the median footprints from 15 to 240, while pushing it down capped the maximum at 20. Training from the base model with the composite reward alone failed to build structure (831 lowest-reward completions). ## Why do checkable rewards matter for AI process modelling? Process models drive analysis, automation, simulation and compliance checks, and building them by hand takes several rounds between domain experts and modelling specialists. Meanwhile, most organisations already describe their work in plain text: procedures, work instructions and ticket templates. A generator that turns those texts into runnable models could bootstrap new models and keep model libraries in step with documentation. The main gain here is reliability rather than peak quality. A smaller open model trained with rewards a program can check rarely produced an invalid model, and on the outside benchmark it avoided the complete failures seen in larger open models. Where an output can be checked automatically, this kind of training can turn a general model into a dependable specialist. ## What are the limits of this process-modelling study? The authors list several limits: - **AI-written training data.** The reference models were generated by an LLM, which may carry its biases and limit diversity. - **Imperfect rewards.** Judge-based rewards can encode bias, depend on prompt wording and can be gamed; footprints only capture which steps follow each other and which can run in parallel. - **One notation.** Only POWL was tested; transfer to BPMN or other notations is not established. The study does not show that process analysts get better or faster models in real work, or that the method works on human-written descriptions from real companies. ## Takeaways - **SAYS** Scores a program can check took broken maps from most tries to almost none. - **SAYS** At its best it matched a far bigger model, and it failed much less often. - **SAYS** The practice examples came from an AI, so real company documents are still untested. - **EXPLAINS** A process model is a precise flowchart of how work gets done, exact enough for software to run and check. - **INTERPRETS** A fair reading is that rewards a program can check made a smaller open model dependable on a narrow job. ## Ask the paper ### What is a verifiable reward? A score a program computes by itself, such as structural checks and footprint agreement, with no human or AI opinion. ### Why did the AI judge come first? Early outputs were mostly invalid, so footprint checks had little to compare; the judge first raised the share of valid models. ### Was it tested on real company documents? No. The training examples were AI-generated, and only one model notation was tested. ## The paper - Title: Specializing large language models for process modeling via reinforcement learning with verifiable and universal rewards - Authors: Alessandro Berti, Xiaoting Wang, Humam Kourani, Wil M. P. Van der Aalst - Venue: Process Science, 18 December 2025 - DOI: [10.1007/s44311-025-00034-4](https://doi.org/10.1007/s44311-025-00034-4) - Licence: CC BY 4.0 [Read this page on the site](https://papersays.com/p/can-a-small-ai-learn-to-draw-workflows-that-run)