It's 2025 – Narrative Learning is the new baseline to beat for explainable ML (arxiv.org)

The paper introduces "Narrative Learning," a new paradigm where classifiers are built and improved entirely through natural-language definitions and iterative explanatory prompts instead of conventional numerical optimization. The authors evaluate the approach on six tasks (three synthetic, three natural) and compare it to seven established explainable-ML baselines. Thanks to advances in 2025-era language models, Narrative Learning outperforms the baselines on five of the six datasets. The study also applies lexicostatistical analyses to model outputs as a proxy for explanation comprehensibility, and provides code, data and demos. This work is significant because it reframes model construction as a human-readable, iterative dialogue about decision criteria, making explainability the central training signal rather than an after-the-fact interpretation. Technical implications include shifting optimization from loss-minimization to prompt-driven refinement, enabling human-in-the-loop rule editing, and leveraging LLM reasoning as a functional component of classifiers. Key caveats for the AI/ML community are robustness, susceptibility to LLM biases and hallucinations, reproducibility across model checkpoints, and the need for standardized evaluation protocols for language-based explanations. If validated broadly, Narrative Learning could become the new baseline for explainable ML, but it also raises fresh challenges around auditability and trustworthiness of language-grounded decision logic.