Few-shot Subgoal Planning with Language Models

TL;DR

This paper demonstrates that pre-trained language models can infer detailed subgoal sequences for real-world tasks from minimal data without fine-tuning, and can be improved with environment feedback, achieving competitive results.

Contribution

It introduces a method for few-shot subgoal prediction using pre-trained language models without requiring strong subgoal supervision or fine-tuning.

Findings

Language models can predict subgoal sequences from few examples.

The approach achieves competitive performance on the ALFRED benchmark.

Re-ranking with environment feedback improves subgoal prediction accuracy.

Abstract

Pre-trained large language models have shown successful progress in many language understanding benchmarks. This work explores the capability of these models to predict actionable plans in real-world environments. Given a text instruction, we show that language priors encoded in pre-trained language models allow us to infer fine-grained subgoal sequences. In contrast to recent methods which make strong assumptions about subgoal supervision, our experiments show that language models can infer detailed subgoal sequences from few training sequences without any fine-tuning. We further propose a simple strategy to re-rank language model predictions based on interaction and feedback from the environment. Combined with pre-trained navigation and visual reasoning components, our approach demonstrates competitive performance on subgoal prediction and task completion in the ALFRED benchmark compared to prior methods that assume more subgoal supervision.