Technical Report: Team SingaX for Embodied Agent Interface Challenge@NeurIPS 2025

Xinyuan Niu, Zhiliang Chen, Vernon Toh, Yanchao Li, Zhengyuan Liu, Nancy F. Chen

Keywords: training-free, prompt-optimization, LLM critic

TL;DR: Our approach uses an LLM to analyze verifier logs and improve the system prompt. It is cheap and training-free, producing significant improvements over baseline.

Abstract: This work presents SingaX's approach to the Embodied Agent Interface Challenge, where we develop an LLM-driven pipeline for interpreting, decomposing, and executing natural language instructions in simulated household environments. Our methodology centers on leveraging large language models as semantic planners. A key innovation of our approach is a \ZY{novel instruction induction framework} that utilizes past error logging statements from development tasks to iteratively improve the LLM's ability to produce semantically consistent and logically correct actions. Our approach is training-free, cheap and efficient to run, and replaces manual effort required in crafting system prompts. In addition, we experimented with various other inference time verification and LLM aggregation approaches. In our report, \textit{we also discussed and analyzed approaches that did not work well in the evaluation task}. Across the four challenge tasks—Goal Interpretation, Subgoal Decomposition, Action Sequencing, and Transition Modeling—we design task-specific prompt structures and cross-task validation routines that encourage coherent, executable outputs.

Submission Number: 8