Paper measures occupational AI learnability with reinforcement learning
The arXiv paper 2605.02598 by Philip Moreira Tomei and Bouke Klein Teeselink, submitted 4 May 2026, introduces an RL Feasibility Index that scores how learnable tasks are by AI using a reinforcement-learning framing, rather than existing capability-overlap measures. Per the paper, the authors use LLM annotators guided by a rubric co-developed with RL experts and validated against confirmed deployment cases to score 17,951 O*NET tasks and aggregate scores to the occupation level. The index diverges sharply from existing AI exposure metrics for several occupation groups: the paper highlights high RL feasibility but low exposure for power plant operators, railroad conductors, and aircraft cargo handling supervisors, and the reverse pattern for creative and interpersonal roles such as musicians and physicians. The authors state these divergences carry direct implications for policy interventions, according to the arXiv abstract.
What happened
The arXiv working paper 2605.02598, submitted 4 May 2026 and authored by Philip Moreira Tomei and Bouke Klein Teeselink, proposes a new metric, the RL Feasibility Index, to measure which occupational tasks AI can learn via a reinforcement-learning style training objective. The paper reports that the authors scored 17,951 tasks from the **O*NET** database using LLM annotators guided by a rubric developed with RL experts, and that they validated annotations against confirmed deployment cases, then aggregated task scores to the occupation level.
Technical details
Per the paper, the methodology frames post-training task acquisition as a task-completion objective aligned with reinforcement learning practice at the frontier. The authors describe using LLM annotators to evaluate training feasibility for each task under a rubric; the rubric and the validation procedure are presented in the manuscript. The resulting RL Feasibility Index is compared to existing AI exposure indexes and the paper reports systematic divergences across occupation groups.
Editorial analysis
Industry context: This approach shifts the measurement problem from static capability overlap toward a learnability-oriented lens, which can reclassify occupations where current model performance and training feasibility diverge. Observed examples in the paper include occupations the authors flag as high on RL feasibility but low on conventional exposure, such as power plant operators, railroad conductors, and aircraft cargo handling supervisors, while creative and interpersonal roles score higher on exposure but lower on RL feasibility.
For practitioners
The paper illustrates a practical method to operationalize 'learnability' at scale using LLM-based annotation plus rubric validation, which could inform downstream workforce, deployment, and regulatory analysis without claiming to predict employer behaviour.
What to watch
Follow peer review and replication: see whether the rubric and validation hold across different LLM annotators and whether independent teams reproduce the occupational divergences. Also watch for extensions applying the index across other countries, sectoral breakdowns, and empirical validation against real-world automation outcomes.
Scoring Rationale
This is a notable research contribution that reframes measurement of AI labor exposure using an RL learnability lens; it matters to practitioners who build policy models, workforce forecasts, or automation risk tools, but it is not a frontier model release.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
