- The paper presents AGI Maze as a new benchmark framework to reveal LLM limitations in constructing persistent, manipulable world models.
- It utilizes a grid-based POMDP with extensions like hidden flows and cyclic pits to test agents' memory management and reasoning under uncertainty.
- Experiments show that incorporating structured working memory improves LLM performance, emphasizing the need for explicit state tracking over simple text-based memory.
AGI Maze: A Benchmark Framework for World-Modeling Agents
Introduction and Motivation
The paper "AGI Maze as a Benchmark Framework for World-Modeling Agents" (2607.00627) presents a formal framework designed to address acute deficits in current LLM-based agents for persistent, efficient world modeling under partial observability. While LLMs exhibit strong pattern completion capabilities and zero-shot generalization in static, fully observable text domains, their autoregressive inference objective does not compel explicit construction or maintenance of manipulable world representations. As a consequence, tasks that superficially appear as "reasoning" in language can be misleading benchmarks when, in fact, meaningful reasoning about latent, dynamically evolving state is not required. The AGI Maze framework is introduced to fill this gap by operationalizing world-modeling challenges that require introspective state tracking, distributed memory management, belief updating, and low-level planning.
Limitations of LLM Agents for World Modeling
Several intrinsic limitations of LLM agents for interactive agentic tasks are exposed. First, LLM inference is inherently static; agentic state evolution is realized only by appending new tokens to the prompt (working memory), resorting to retrieval-augmented generation (RAG) for long-term recall, or incorporating tool feedback as context. This textual, unstructured approach to memory is token-expensive, brittle, and not aligned with the requirements of coherent, updatable world representations. Second, the next-token objective induces opportunistic local conditioning, not persistent, queryable environment models. As demonstrated in a series of empirical studies (Chen et al., 22 May 2026), intermediate transformer activations are consistent with shallow predictive coding rather than maintenance of structured, manipulable state correlates. These deficiencies are evident in interactive, partially observable settings, even though they are partially obfuscated in traditional language understanding scenarios.
The AGI Maze Framework
AGI Maze is a grid-based benchmark and associated API that refrains from demanding high-dimensional representation learning (e.g., pixels), instead focusing strictly on reasoning and memory tasks. In its archetypal form, the maze environment is a partially observable Markov decision process (POMDP) where the agent starts on a grid, receives textual feedback upon actions, and must acquire a key and treasure, avoiding walls and monolithic borders. Critically, the agent is only given local observations—such as "You tried to go right, but a monolith blocks the way"—and not the underlying full maze structure.
To robustly pressure test world modeling, extensions such as rivers (with hidden flows causing forced state transitions) and cyclic pits (teleportation with ambiguous mapping) are introduced, producing state uncertainty, ambiguous localization, and the necessity for belief distributions over possible states. Maze instances include progressively more challenging regimes (TRAINING, CLASSIC, EXTENDED, and HIDDEN), with step budgets calibrated against human performance.
Through its extensible API, AGI Maze supports the addition of novel items, cell types, and agent actions—enabling principled examination of generalization to new, out-of-distribution mechanics. This gives rise to a robust test for whether agents infer and utilize new action policies and world rules, or merely overfit to superficially similar episodes.
Baseline Evaluation: LLM and Planning Agents
Initial evaluation is performed using vanilla LLM agents, which receive the current history of actions and observations prepended by the game rules, directly predicting next moves via the model. Performance on simple (3x3, 3x4, 4x4) mazes is substantially sub-human and, for some models (e.g., GPT-4o Mini), statistically worse than uninformed random-walk baselines. Only on the smallest mazes without uncertainty dynamics do LLMs match or marginally outperform random search.
More sophisticated LLMs (GPT-5.5, Gemini 3.5 Flash) achieve up to 40% success rate on the hardest instance evaluated (S3-03, doubled step budget), but remain unreadily competitive with humans. Qualitative analyses suggest only limited spatial inference occurs internally, without construction of a stable maze map or nuanced belief tracking.
Allowing agents to use prompt-appended "notes" as working memory—structuring plans, location hypotheses, and exploration progress—more than doubles the strongest LLMs' performance (e.g. GPT-5.5: 30%→60%, Gemini 3.5 Flash: 40%→70%). This demonstrates that the bottleneck is the lack of structured, persistent state, not the inaccessibility of environmental information per se. However, lighter LLMs (GPT-5 Mini, Gemini 3.1) showed no performance gains with working memory, underscoring a model scaling effect for memory-based reasoning.
Theoretical and Practical Implications
The results strongly support the position that standard LLM inference is insufficient for robust world modeling under partial observability and complex action dynamics. The AGI Maze framework highlights the critical gap between prediction/auto-completion and true stateful agency, especially in the presence of ambiguous percepts or environment mechanics requiring explicit belief updates and uncertainty propagation.
Practically, the findings point to the necessity for architectural and procedural augmentations beyond "memory-as-text," including explicit, computable world model modules (maps, graphs, Bayesian filters), more efficient memory architectures, and learned strategy adaptation for novel environment extensions. The results also suggest that even with externalized "notepad" style working memory, strong performance does not emerge without structural domain knowledge; i.e., agents must be capable of both representing and updating their beliefs in synchrony with an underlying latent environment.
The AGI Maze API and environment design offer a generalizable interface for evaluating not only LLM agents but also RL agents, neural-symbolic hybrids, and agents with algorithmic memory or structured internal simulation. It is positioned to drive research directions on memory compression, uncertainty handling, and generalization under systematic environment modifications.
Conclusion
AGI Maze provides a highly targeted framework and API for the study of world modeling in agents. It isolates core agentic challenges by removing the confounds of pixel-based perception, focusing rigorously on representation learning, memory management, reasoning, and adaptive planning under partial observability. Baseline experiments with vanilla and planning-augmented LLMs show persistent deficiencies in internal state tracking and belief updating, marking a clear frontier for both model architecture and agent design research. Future agentic systems built to interact with AGI Maze must coordinate scalable, explicit world modeling modules capable of principled memory use, uncertainty tracking, and adaptive heuristics for novel dynamics—advancing the underlying theoretical understanding of flexible, embodied intelligence.