EvoEmpirBench: Dynamic Spatial Reasoning with Agent-ExpVer

2025-09-16

AI Summary

Existing spatial reasoning benchmarks predominantly focus on static, fully observable environments, limiting their ability to assess models’ long-horizon reasoning and memory utilization under partial observability and dynamic conditions. To address this gap, we introduce two novel benchmarks: Local-Observable Maze Navigation and Match-2 Elimination—designed to systematically evaluate agents’ spatial understanding, adaptive planning, and cognitive updating in dynamic settings. Our method introduces a subjectively experienced, structured memory mechanism that jointly models local observations, integrates dynamic environmental feedback, and enables online memory revision, all embedded within a reinforcement learning framework for continual policy optimization. Experiments reveal that state-of-the-art models exhibit significant performance degradation on these tasks, underscoring their difficulty and the inadequacy of current architectures. These findings validate the benchmarks’ utility in advancing embodied spatial reasoning and long-term memory-augmented agent design.

Technology Category

• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Multiagent Systems: Adversarial Agents
• Planning, Routing, and Scheduling: Model-Based Reasoning

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Most existing spatial reasoning benchmarks focus on static or globally observable environments, failing to capture the challenges of long-horizon reasoning and memory utilization under partial observability and dynamic changes. We introduce two dynamic spatial benchmarks, locally observable maze navigation and match-2 elimination that systematically evaluate models' abilities in spatial understanding and adaptive planning when local perception, environment feedback, and global objectives are tightly coupled. Each action triggers structural changes in the environment, requiring continuous update of cognition and strategy. We further propose a subjective experience-based memory mechanism for cross-task experience transfer and validation. Experiments show that our benchmarks reveal key limitations of mainstream models in dynamic spatial reasoning and long-term memory, providing a comprehensive platform for future methodological advances. Our code and data are available at https://anonymous.4open.science/r/EvoEmpirBench-143C/.

Problem

• Evaluates spatial reasoning under partial observability and dynamics
• Tests adaptive planning with local perception and global objectives
• Assesses memory utilization for long-horizon reasoning tasks

Innovation

• Dynamic spatial benchmarks with partial observability
• Action-triggered environment changes requiring strategy updates
• Subjective experience-based memory for cross-task transfer