The Cognitive Bandwidth Bottleneck: Shifting Long-Horizon Agent from Planning with Actions to Planning with Schemas

TL;DR

The paper investigates action representations for long-horizon agents by contrasting planning with actions (PwA) and planning with schemas (PwS) through the Cognitive Bandwidth Perspective, a qualitative framework that partitions cognitive load across stages. It shows an inflection point between small and large action spaces (ALFWorld ≈ $35$ actions vs SciWorld ≈ $500$ actions) where PwA's environment-understanding burden becomes prohibitive and PwS's schema-instantiation burden becomes advantageous. Across four environments, PwA dominates in small spaces while PwS excels in large spaces, illustrating that there is no universal best representation. The authors argue for strengthening PwS via post-training with multi-turn tool use to shift the inflection left, enabling more scalable open-world autonomy, and provide open-source code to support reproducibility.

Abstract

Enabling LLMs to effectively operate long-horizon task which requires long-term planning and multiple interactions is essential for open-world autonomy. Conventional methods adopt planning with actions where a executable action list would be provided as reference. However, this action representation choice would be impractical when the environment action space is combinatorial exploded (e.g., open-ended real world). This naturally leads to a question: As environmental action space scales, what is the optimal action representation for long-horizon agents? In this paper, we systematically study the effectiveness of two different action representations. The first one is conventional planning with actions (PwA) which is predominantly adopted for its effectiveness on existing benchmarks. The other one is planning with schemas (PwS) which instantiate an action schema into action lists (e.g., "move [OBJ] to [OBJ]" -> "move apple to desk") to ensure concise action space and reliable scalability. This alternative is motivated by its alignment with human cognition and its compliance with environment-imposed action format restriction. We propose cognitive bandwidth perspective as a conceptual framework to qualitatively understand the differences between these two action representations and empirically observe a representation-choice inflection point between ALFWorld (~35 actions) and SciWorld (~500 actions), which serve as evidence of the need for scalable representations. We further conduct controlled experiments to study how the location of this inflection point interacts with different model capacities: stronger planning proficiency shifts the inflection rightward, whereas better schema instantiation shifts it leftward. Finally, noting the suboptimal performance of PwS agents, we provide an actionable guide for building more capable PwS agents for better scalable autonomy.

Figures (3)

• Figure 1: The relationship between environment complexity and agent paradigm effectiveness. In simple environments, Grounded Actions suffice. As complexity grows past an "inflection point", the Action Schema paradigm becomes superior by shifting the cognitive load from processing long action lists (high EU load) to a challenging but scalable Schema Instantiation (SI) step.
• Figure 2: Comparison of agent behaviors in ALFWorld and SciWorld, two environments situated on either side of the observed complexity inflection point.
• Figure 3: Effect of action-space size and environment complexity on success rate. In low-complexity settings (e.g., baseline ALFWorld), planning with actions (PwA) outperforms planning with schemas (PwS). As the action list expands with distractors, PwA performance declines and can be overtaken by PwS for sufficiently capable models, revealing the representation inflection point.