Can LLMs Translate Human Instructions into a Reinforcement Learning Agent's Internal Emergent Symbolic Representation?

TL;DR

This paper explores whether large language models can convert human instructions into the internal symbolic representations used by reinforcement learning agents, revealing partial success and current limitations.

Contribution

It introduces a structured evaluation framework to assess LLMs' ability to translate natural language into internal symbolic representations in RL environments.

Findings

LLMs can partially translate natural language into environment dynamics symbols

Performance varies with partition granularity and task complexity

Current LLMs show limitations in aligning language with internal agent representations

Abstract

Emergent symbolic representations are critical for enabling developmental learning agents to plan and generalize across tasks. In this work, we investigate whether large language models (LLMs) can translate human natural language instructions into the internal symbolic representations that emerge during hierarchical reinforcement learning. We apply a structured evaluation framework to measure the translation performance of commonly seen LLMs -- GPT, Claude, Deepseek and Grok -- across different internal symbolic partitions generated by a hierarchical reinforcement learning algorithm in the Ant Maze and Ant Fall environments. Our findings reveal that although LLMs demonstrate some ability to translate natural language into a symbolic representation of the environment dynamics, their performance is highly sensitive to partition granularity and task complexity. The results expose limitations in current LLMs capacity for representation alignment, highlighting the need for further research on robust alignment between language and internal agent representations.