AI BriefingLG AI Research: Commonsense-Based Subgoal Planning Research
Key point
They presented a methodology that leverages the commonsense knowledge embedded in LLMs to plan complex tasks step by step, improving planning accuracy through mutual information-based ranking and environment feedback.
Details
They noted that language models (LLMs) can encode commonsense knowledge useful for planning subgoals, a sequence of intermediate steps, from text instructions. They propose a methodology that achieves a high 70% recall in plan prediction using only 20 training examples.
When LLMs generate candidates, a popularity bias—where certain words appear frequently—causes distortion in ranking. To address this, they introduced a ranking criterion based on Mutual Information. In particular, they improved prediction performance by utilizing a weighted PMI that combines the relationship between the query and the hypothesis.
The research team validated performance through the ALFRED benchmark. Experimental results confirmed that combining the two metrics with λ = 0.5 consistently showed higher accuracy and improved performance compared to existing methods.
Furthermore, they proposed a method to incorporate not only text instructions but also Environment Feedback. By vectorizing information the agent obtains from its surroundings and passing it to a re-ranker model, they designed the system to reduce object identification errors and enable more sophisticated subgoal prediction.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.