Llm planning improves efficiency with expectation observation checks
As Expected? Efficient LLM Planning via Expectation-Observation Consistency Verification
Computation and Language
Summary
Planning with large language models (LLMs) often needs many back-and-forth steps that can be slow and costly. The authors introduce a method called expectation-observation consistency verification (EOCV) that lets the model plan multiple actions at once but checks whether outcomes match its expectations to avoid big mistakes. If something doesn’t match, the model stops and replans using the latest information. This approach keeps performance similar to step-by-step methods but uses fewer computing resources.
What this means in practice
- •For ai developers: Reduce computing costs in AI systems that plan with language models by integrating expectation checks to catch errors early during multi-step predictions.
- •For robotic software engineers: Improve efficiency of robots or agents that use language models for planning by verifying expected outcomes against observations to decide when to replan.
Authors
Qingfu Zhu, Huanyang Zheng, Ting Liu, Wanxiang Che
Abstract
LLM planning often relies on reactive executors (e.g., ReAct) that generate a single action at each step to remain grounded in the latest observations, incurring substantial inference cost. Batched executors reduce this cost by predicting multiple actions in a single model invocation, yet they are prone to compounding errors, as intermediate observations are ignored. We propose an expectation-observation consistency verification (EOCV) mechanism for batched executors, preserving grounding while improving efficiency. The mechanism generates expectations of action outcomes and compares them with the resulting observations. Consistency permits continued execution, whereas mismatches trigger early termination and replanning from the current observation. Experiments on text-based and multimodal benchmarks across multiple backbones show broadly comparable average task performance between EOCV and reactive executors, with a 29-35% reduction in average weighted token usage. Moreover, the mechanism is model-agnostic and remains effective when integrated into state-of-the-art planners.