Researchers Introduce Alice: Closed-Loop System for Self-Supervised Dynamics Discovery

A preprint study published on arXiv introduces Alice, a novel closed-loop system for online self-supervised dynamics discovery in executable world models. The research, titled ‘Baba in Wonderland: Online Self-Supervised Dynamics Discovery for Executable World Models’ (arXiv:2605.16725v1), presents an approach that uses structural signals from failed candidate updates to iteratively refine hypothesis classes without requiring reward signals or lexical priors.

The framework addresses challenges in creating executable world models that accurately capture environmental transition laws rather than relying on surface-level vocabulary patterns. By analyzing failures in candidate model updates, Alice dynamically adjusts its hypothesis space to better align with observed interaction evidence. This approach differs from traditional methods that depend on rule descriptions or hand-crafted reward functions.

Executable world models, which can be read, edited, executed, and reused for planning, require precise representation of environmental dynamics. The study demonstrates how Alice operates under prior misalignment conditions, inducing state-dependent dynamics purely through interaction. The system’s ability to learn without semantic shortcuts marks a departure from conventional approaches in reinforcement learning and world modeling.

The research has implications for advancing autonomous systems that require robust environmental modeling capabilities. By focusing on structural signal extraction from failures, the framework opens new avenues for self-supervised learning in complex dynamic environments.