Groundhog Loops: A Time-Loop Testbed for Measuring Learning Speed and Tactic Formation in AI Agents

Abstract

The project defines a test environment inspired by the time-loop mechanism featured in the film Groundhog Day. An agent attempts to complete a task; if it makes an error classified as fatal (or exhausts its step budget), the environment is restored to a previously saved checkpoint, while the agent retains a complete, structured memory of all previous loops: the loop counter, cumulative time, summaries of failed attempts, and its own hypotheses. Unlike a conventional episode reset in reinforcement learning (RL), failure information is not lost—it becomes a persistent part of the agent's observations. Additionally, after a reset, the environment may slightly mutate the task conditions, but only when necessary to force the agent to change its strategy (the minimal-intervention principle). These mutations are observable to the agent but are never explicitly signaled. Project objectives: - C1. Build a measurable benchmark for agents' ability to perform within-episode meta-learning (in-context adaptation across resets): the agent should learn from failures without updating its weights. - C2. Investigate whether explicit loop memory reduces time to success compared with an agent without memory (control-group experiments). - C3. Investigate whether minimal changes to task conditions disrupt strategy lock-in, i.e., a situation in which an agent repeatedly attempts variations of the same flawed strategy despite having complete memory of its failures. - C4. Provide a reproducible, deterministic (seeded) harness for experiments with RL agents and LLM-based agents, using a single, unified evaluation protocol. - C5. Identify and measure undesirable emergent behaviors: reward hacking through deliberately triggering resets, learned helplessness, and overfitting to the mutation pattern. Out of scope: gradient-based weight training across loops (the project evaluates in-context adaptation through memory, not through-gradient meta-learning); physical real-time simulation. Artifacts: github.com/casual-art/Groudhog-Day (harness, docs); archive DOI: 10.5281/zenodo.23483377.