# Language Models as Agent Models

**Authors:** Jacob Andreas  
**arXiv:** [2212.01681](https://arxiv.org/abs/2212.01681)  
**Date:** 2022-12-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.01681v1/

## Key points

- Language models (LMs) can serve as models of intentional communication in a narrow sense
- LMs can infer and represent properties of the agent likely to have produced a given textual context
- LMs can utilize fine-grained communicative intentions, beliefs, and goals
- LMs have the potential to act as building blocks for systems that communicate and act intentionally
- Challenges previous notions about the limitations of LMs in modeling goal-directed aspects of human language production and comprehension

## Summary

In the paper "Language Models as Agent Models" by Jacob Andreas, the author explores the capabilities of language models (LMs) in understanding the relationship between language and use. LMs are trained on collections of documents written by human agents to achieve specific goals in the outside world. The author argues that LMs can indeed serve as models of intentional communication in a narrow sense. When predicting the next word based on a given textual context, an LM can infer and represent properties of the agent likely to have produced that context. These representations can then influence subsequent LM generation, similar to how agents' communicative intentions impact their language. Recent literature suggests that LMs are capable of inferring and utilizing fine-grained communicative intentions, as well as more abstract beliefs and goals. This indicates that even with their limited training data, LMs can act as building blocks for systems that communicate and act intentionally. Overall, this paper highlights the potential for LMs to learn about the relationship between language and use, challenging previous notions about their limitations in modeling goal-directed aspects of human language production and comprehension.