acceptodds
Under review as a conference paper at ICLR 2027

SpatialAgent: Working Memory for Long-Horizon Spatial Reasoning

Abstract

Long-horizon interactive tasks require large language model (LLM) agents to retain and use information gathered across many steps, particularly in partially observed environments where spatial knowledge remains relevant throughout an episode. However, existing working-memory approaches are not generally organized around the spatial structure of the environment itself. We investigate whether explicitly maintaining a structured spatial representation in memory can improve both agent performance and inference efficiency. We introduce \it SpatialAgent, an agent framework that constructs a persistent map from textual observations and provides it to an underlying LLM. The map is adapted to the structure of each environment and updated incrementally throughout the episode as new information is observed. We evaluate SpatialAgent across multiple environments from the BALROG benchmark, comparing it with existing agent baselines while also studying model scale, alternative map-construction methods, and inference efficiency. Results show that SpatialAgent outperforms existing agent baselines across most evaluated models and environments. We further find that SpatialAgent generally improves over an otherwise equivalent no-map variant, indicating that the explicit spatial representation itself contributes to performance; SpatialAgent can reduce inference cost by maintaining strong performance with shorter context windows and less expensive models. These results suggest that spatially structured working memory can be a useful design choice for LLM agents operating in long-horizon environments while reducing reliance on extensive trajectory context.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Model Agents, Working Memory, Spatial Memory, Long-Horizon Decision Making, Text-Based Environments

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_hkiPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces SpatialAgent, which maintains an explicit incremental map for LLM agents operating in partially observable environments. Experiments span four environments, multiple language models, and several map-construction strategies. Although the method sometimes produces improvements, its contribution is mainly a system-level integration of established spatial-map ideas. The current results do not support the paper’s broad claims about novelty, generalization, working memory, or consistent superiority. Overall, this paper does not meet the acceptance criteria of ICLR.

Critical strengths
  1. The motivation is reasonable. Maintaining structured spatial state is highly relevant to long-horizon agents under partial observability, and the explicit map provides an interpretable mechanism for storing spatial information.
  2. The paper evaluates several model scales across four environments and compares map-construction strategies. Some configurations show large improvements, such as Qwen3-4B on BabyAI, indicating that explicit spatial state can be valuable under suitable conditions.
Critical weaknesses
  1. Unclear novelty and inaccurate conceptual positioning. Explicit spatial maps have already been used to improve LLM agents, while SpatialAgent largely adopts a conventional location–relation–property representation. Moreover, only the map data structure is shared across environments, each environment still requires specialized procedures. Therefore, the claimed generalizability is not convincingly demonstrated.
  2. The empirical evidence does not support the strength of the claims. SpatialAgent is not consistently superior. For example, it underperforms HiAgent on BabyAI with Qwen3-8B and Qwen3-32B, ReAct on TextWorld with Qwen3-8B, and StateAct on BabaIsAI with Qwen3-32B. More importantly, the evaluation uses the default BALROG episode counts with only a single attempt per episode, and the paper does not report independent repeated runs across random seeds. Although Appendix Table 4 reports standard errors for the baseline methods, corresponding uncertainty estimates are not provided for SpatialAgent, the map-construction comparison, or the main ablations. Consequently, the claim that SpatialAgent “significantly outperforms” prior methods is not statistically substantiated by paired tests, confidence intervals for performance differences, or other appropriate analyses. The absence of a clear limitations section further obscures the method’s environment-specific engineering requirements and inconsistent performance.
Questions for authors
  1. Compared with explicit-map or graph-memory approaches such as AriGraph and MapGPT, what is the essential and non-substitutable contribution of SpatialAgent?
  2. If each environment still requires a specialized representation schema and MapBuilder, what exactly do the authors mean by “generalization”?
  3. Could the authors report the accuracy of the constructed maps, including nodes, edges, object attributes, and agent locations. Why direct LLM-based map construction yields a score of 0.0 in some settings?
  4. When output tokens, total interaction steps, and all map-construction calls are included, does SpatialAgent still achieve a lower total inference cost?
Reviewer_d8TgOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on long-horizon spatial reasoning tasks for AI agents and proposes a structured, map-based memory mechanism that is dynamically constructed and updated throughout the reasoning process to improve task performance and computational efficiency.

Critical strengths
  1. The idea of incorporating map-based working memory into agents for spatial reasoning tasks is interesting and potentially valuable.

  2. The proposed method is evaluated across multiple benchmarks and models, providing a relatively comprehensive assessment of its effectiveness.

  3. The method achieves promising results on several tasks, demonstrating its potential to improve spatial reasoning performance.

Critical weaknesses
  1. The methodological details are not clearly explained. For instance, it is unclear how the maps are constructed and how they are updated based on new observations and actions. The authors provide only a superficial description of the method, lacking sufficient depth and clarity.

  2. The experimental results show that maps directly constructed by LLMs yield poor performance.

  3. Figure 5 is difficult to read and interpret.

  4. The paper lacks ablation studies to evaluate the contributions of individual components.

Questions for authors
  1. For the results reported in Table 2, how are the maps constructed? Are they manually created by humans or generated by LLMs?

  2. How does the proposed method handle conflicting spatial information caused by LLM hallucinations during map updates?

Reviewer_KZXWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SpatialAgent, an LLM agent framework that maintains an explicit spatial map as persistent working memory for long-horizon interactive tasks. The map is incrementally updated from textual observations and actions, and relevant spatial information is provided to the LLM for decision-making. The authors investigate three map-construction methods and evaluate the framework across four text-based BALROG environments, comparing task performance, model scaling, and inference efficiency against existing agent baselines.

Critical strengths
  1. The paper addresses a practical challenge in long-horizon agent reasoning by maintaining spatial information that may otherwise be lost from limited interaction histories.

  2. The evaluation covers four environments with different spatial structures and multiple model families, providing useful evidence of the approach's applicability across settings.

  3. The no-map ablation, map-construction comparisons, and inference-cost analysis provide useful insights into when spatial memory helps and how different update mechanisms affect performance.

Critical weaknesses
  1. The core idea of maintaining and updating explicit spatial maps for agent reasoning substantially overlaps with prior work, leaving the conceptual novelty unclear.

  2. The paper does not directly compare against existing spatial-memory agents such as AriGraph, despite their closely related objectives.

  3. The experiments establish the benefit of adding a map but do not demonstrate its advantages over alternative spatially structured memory representations.

  4. The main experimental configuration is unclear, as some Table 2 results match the LLM-generated MapBuilder in Table 3 while others match the human-written implementation, without an explained selection protocol.

Questions for authors
  1. What advantages does SpatialAgent offer over prior approaches such as MindCube, Theory of Space, and AriGraph in spatial memory construction, updating, and long-horizon task performance?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.