Learning Explicit Behavioral Models with Adaptive Questions and World-Model Probes

A new trainable behavioral model called ESBM couples high task scores with explicit, evidence-grounded answers and executable mechanism predictions, according to a paper submitted 5 June 2026. The approach targets a known weakness in interactive agents: achieving high returns without representing the mechanisms behind their actions [1]. The Explicit Symbolic Behavioral Model represents behavior through typed predicates, weighted rules, bounded options, and a mechanism memory that predicts symbolic events, object changes, rewards, and terminal consequences under action interventions [1]. After each rollout, adaptive questions and active world-model probes convert score failures, question-answering errors, and transition-prediction errors into constraints for local edits to the model [1]. Candidate models are selected by a multi-criterion rule that jointly evaluates task score, answerability, and active world-model consistency [1]. Under Atari-style protocols, ESBM learned high-scoring policies while producing explicit answers and executable mechanism predictions [1]. The work addresses a persistent gap in reinforcement learning. Interactive agents trained solely against task return can achieve high scores yet fail to represent the mechanisms that make their actions succeed, leaving brittle behavior that is difficult to diagnose and that limits adaptation when environment dynamics change [1]. Existing approaches using large language model reflection and policy-code repair can revise behavior from failed trajectories, but questions and world-understanding tests are typically applied only after training [1]. Large language models are neural networks trained on vast amounts of text for natural language processing tasks, especially language generation, and are the foundational technology behind modern chatbots [3]. The ESBM framework intersects with the broader field of explainable artificial intelligence, which explores methods that provide humans with intellectual oversight over AI algorithms by making the reasoning behind decisions more understandable and transparent [4]. Explainable AI counters the “black box” tendency of machine learning, where even designers cannot explain why a model arrived at a specific decision [4]. The ESBM authors indicate that adaptive questions can serve as both training pressure and reusable benchmarks for mechanistic policy learning in the tested setting [1]. Adversarial conditions further motivate the need for mechanistic understanding. Adversarial machine learning studies attacks on machine learning algorithms and defenses against them, including evasion attacks, data poisoning attacks, Byzantine attacks, and model extraction [5]. In practical high-stakes applications, users may intentionally supply fabricated data that violates the statistical assumption that training and test data are drawn from the same distribution [5]. Systems that lack explicit world models are especially vulnerable to such distribution shifts. The paper was submitted on 5 June 2026 to arXiv, a preprint repository that has integrated with Hugging Face Spaces to make machine learning demos accessible alongside paper abstracts [1][10]. Through that integration, users can find open-source demos built by the community and try them immediately in a browser without writing code [10].