TL;DR
When deploying LLM agents in open-world environments, presenting models with exhaustive lists of every possible executable action causes decision-making bottlenecks as action spaces combinatorially explode. This paper proposes Planning with Schemas (PwS)—having agents instantiate abstract templates (e.g., move [OBJ] to [OBJ]) instead of choosing from massive action lists. The authors demonstrate a critical complexity inflection point between low-action environments (like ALFWorld) and high-action environments (like SciWorld), revealing that multi-turn tool-use post-training—rather than deep reasoning alone—is the key driver for scalable, schema-based autonomy.
What Problem Does This Paper Address?
Most existing LLM-based autonomous agents rely on Planning with Actions (PwA): at every step, the simulation provides an explicit list of valid commands, and the model selects one. While effective in toy benchmarks, real-world tasks lead to a combinatorial explosion of possible actions.
Providing hundreds or thousands of available actions creates an intense "needle-in-a-haystack" burden, straining LLM context windows and distracting the agent. Conversely, free-form text generation often produces malformed commands that interactive environments reject. The authors ask: As environmental action space scales, what is the optimal action representation for long-horizon agents?
Methods
The authors contrast PwA with Planning with Schemas (PwS), where agents receive abstract templates with placeholders (e.g., take [OBJ] from [OBJ]) and ground them into concrete actions based on context.
To conceptualize the trade-offs, the paper introduces the Cognitive Bandwidth Perspective:
- An LLM has a finite Cognitive Bandwidth ($B$) to process context, maintain causal reasoning, and make decisions.
- PwA heavily expends load on Environment Understanding (EU) due to reading through lengthy, noisy candidate action lists.
- PwS frees up EU load by providing a compact schema list (e.g., 26 templates instead of 500+ actions), shifting the computational burden to Schema Instantiation (SI)—the reasoning required to correctly bind environmental objects into template slots.
To test how model capacity alters this trade-off, the authors design a cognitive-load stress test in ALFWorld, incrementally injecting up to 1,024 plausible distractor actions into the candidate list to isolate the impact of EU burden without altering underlying task dynamics.
Datasets
The authors evaluated agents on 100 sampled tasks across four interactive text-based environments with increasing action-space scales (adapted from the AgentGym suite):
- TextCraft: Minimal action space (~3 schemas; recipe-based crafting).
- WebShop: Small-to-medium action space (3–10 clickable/searchable actions; 2 schemas).
- ALFWorld: Moderate action space (30–40 grounded actions; 11 schemas).
- SciWorld: High-complexity action space (400–600 grounded actions; 26 schemas).
Key Insights and Ideas
- The Representation Inflection Point: There is a clear turning point between ALFWorld (~35 actions) and SciWorld (~500 actions). In low-to-medium action spaces, PwA wins because schema instantiation overhead ($L_{SI}$) causes unnecessary errors. In large action spaces, PwA collapses under the sheer noise of the candidate list, making PwS superior.
- Error Modes Invert Across Complexity: In ALFWorld, schema-based agents suffer higher invalid action rates because models fail to ground templates. In SciWorld, action-based agents generate more invalid and repetitive actions because long lists degrade situational comprehension.
- Reasoning Depth $\neq$ Schema Instantiation: Long-reasoning models (LRMs) with extended chain-of-thought (e.g., DeepSeek-R1, GPT5-mini) improve general planning depth but do not automatically master schema instantiation.
- Tool-Use Training is the Key Catalyst: Models explicitly post-trained on massive multi-turn tool use (e.g., Kimi-K2, LongCat) excel at PwS. Filling slot parameters in structured API calls translates directly to instantiating abstract schemas, shifting the inflection point to the left (making PwS beneficial at smaller action spaces).
Key Results
- ALFWorld vs. SciWorld Crossover:
- On ALFWorld (~35 actions), PwA outperformed PwS by 33.4% on average across models.
- On SciWorld (~500 actions), the trend reversed: PwS beat PwA by 8.1% on average.
- Specific Model Highlights on SciWorld:
- GPT-4.1: PwS scored 62.2 vs. 43.3 for PwA (+18.9 delta).
- Seed-OSS-36B: PwS scored 67.4 vs. 49.5 for PwA (+17.9 delta).
- Gemini-2.5-Flash: PwS scored 69.1 vs. 57.2 for PwA (+11.9 delta).
- Stress-Testing Degradation: In ALFWorld, as distractors scaled from ~32 up to 1,024 actions, PwA performance monotonically dropped across all models (e.g., Qwen2.5-7B dropped by 0.42, LongCat dropped by 0.29), allowing schema-based baselines to overtake action-based performance for capable models.
Limitations and Future Work
- Qualitative Construct: The Cognitive Bandwidth Perspective is a conceptual framework rather than a directly quantifiable mathematical metric.
- Simulator Domain Gap: Experiments were confined to text-based, discrete simulation environments; the findings may require adaptation before transferring to continuous physical robotics or open-ended web environments.
- Future Directions: Developing dedicated training recipes that explicitly combine multi-turn agentic reinforcement learning with dynamic schema induction to further lower schema instantiation costs.
Summary updated