PokeGym: A Visually-Driven Long-Horizon Benchmark for Vision-Language Models

TL;DR

PokeGym introduces a comprehensive, visually-driven long-horizon benchmark within a 3D game environment to evaluate vision-language models' capabilities in complex tasks, emphasizing autonomous exploration and visual reasoning.

Contribution

It presents a novel benchmark with strict visual-only operation, diverse tasks, and automated evaluation, revealing key limitations of current VLMs and highlighting areas for architectural improvements.

Findings

Current VLMs struggle with physical deadlock recovery.

Weaker models often fail to recognize entrapment (Unaware Deadlocks).

Advanced models recognize deadlocks but cannot recover (Aware Deadlocks).

Abstract

While Vision-Language Models (VLMs) have achieved remarkable progress in static visual understanding, their deployment in complex 3D embodied environments remains severely limited. Existing benchmarks suffer from four critical deficiencies: (1) passive perception tasks circumvent interactive dynamics; (2) simplified 2D environments fail to assess depth perception; (3) privileged state leakage bypasses genuine visual processing; and (4) human evaluation is prohibitively expensive and unscalable. We introduce PokeGym, a visually-driven long-horizon benchmark instantiated within Pokemon Legends: Z-A, a visually complex 3D open-world Role-Playing Game. PokeGym enforces strict code-level isolation: agents operate solely on raw RGB observations while an independent evaluator verifies success via memory scanning, ensuring pure vision-based decision-making and automated, scalable assessment. The benchmark comprises 30 tasks (30-220 steps) spanning navigation, interaction, and mixed scenarios, with three instruction granularities (Visual-Guided, Step-Guided, Goal-Only) to systematically deconstruct visual grounding, semantic reasoning, and autonomous exploration capabilities. Our evaluation reveals a key limitation of current VLMs: physical deadlock recovery, rather than high-level planning, constitutes the primary bottleneck, with deadlocks showing a strong negative correlation with task success. Furthermore, we uncover a metacognitive divergence: weaker models predominantly suffer from Unaware Deadlocks (oblivious to entrapment), whereas advanced models exhibit Aware Deadlocks (recognizing entrapment yet failing to recover). These findings highlight the need to integrate explicit spatial intuition into VLM architectures. The code and benchmark will be available on GitHub.