Solving the Jericho benchmark suite of text-based games

Determine whether the Jericho benchmark suite of text-based interactive fiction games can be fully solved by autonomous agents under the standard Jericho evaluation protocol, specifically overcoming partial observability that requires constructing world models from local textual descriptions and combinatorial state–action spaces that cause hard-exploration challenges.

Background

The Jericho benchmark is a widely used suite of text-based interactive fiction games designed to evaluate agents’ abilities in language-grounded decision making. It presents two core challenges: partial observability, which requires agents to infer and maintain world state from local textual observations, and extremely large combinatorial action spaces derived from natural language commands.

Despite progress with RL, MCTS, and LLM-based agents, Jericho has proven difficult to solve comprehensively, and existing approaches have required extensive environment interactions or exhibit limited exploration capabilities. The paper explicitly acknowledges that Jericho remains unsolved, motivating methods that can learn effectively from sparse rewards and navigate deceptive local optima.