# Faith and Fate: Limits of Transformers on Compositionality

**Authors:** Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras, Jena D. Hwang, Soumya Sanyal, Sean Welleck, Xiang Ren, Allyson Ettinger, Zaid Harchaoui, Yejin Choi  
**arXiv:** [2305.18654](https://arxiv.org/abs/2305.18654)  
**Date:** 2023-05-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.18654v1/

## Key points

- Transformer large language models (LLMs) are highly praised for their exceptional performance on complex multi-step reasoning tasks
- However, they also exhibit surprising failures on seemingly trivial problems, raising questions about their limitations
- Researchers led by Nouha Dziri investigated the limits of Transformers across three representative compositional tasks: multi-digit multiplication, logic grid puzzles, and a classic dynamic programming problem
- These tasks require breaking down problems into sub-steps and synthesizing them into a precise answer
- The researchers formulated these compositional tasks as computation graphs to systematically quantify their level of complexity and break down reasoning steps into intermediate sub-procedures
- Empirical findings suggest that Transformers solve compositional tasks by reducing multi-step reasoning into linearized subgraph matching without necessarily developing systematic problem-solving skills
- In other words, they approach complex problems in a simplistic way that may not be scalable to more challenging scenarios
- The study was conducted by a team of researchers including Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras Jena D. Hwang Soumya Sanyal Sean Welleck Xiang Ren Allyson Ettinger Zaid Harchaoui and Yejin Choi.
- Their work is presented in "Faith and Fate: Limits of Transformers on Compositionality," which consists of 10 pages plus an appendix (21 pages)
- The researchers also provide theoretical arguments on abstract multi-step reasoning problems that highlight how Transformers' performance will rapidly decay with increased task complexity.

## Summary

The exceptional performance of Transformer large language models (LLMs) on complex multi-step reasoning tasks has been widely admired. However, these models also exhibit surprising failures on seemingly trivial problems, raising the question of whether these errors are incidental or indicative of more significant limitations. To shed light on this issue, a team of researchers led by Nouha Dziri conducted an investigation into the limits of Transformers across three representative compositional tasks: multi-digit multiplication, logic grid puzzles and a classic dynamic programming problem. These tasks require breaking down problems into sub-steps and synthesizing them into a precise answer. The researchers formulated these compositional tasks as computation graphs to systematically quantify their level of complexity and break down reasoning steps into intermediate sub-procedures. Their empirical findings suggest that Transformers solve compositional tasks by reducing multi-step reasoning into linearized subgraph matching without necessarily developing systematic problem-solving skills; in other words, they approach complex problems in a simplistic way that may not be scalable to more challenging scenarios. The study was conducted by a team of researchers including Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jian, Bill Yuchen Lin, Peter West, Chandra Bhagavatula, Ronan Le Bras Jena D. Hwang Soumya Sanyal Sean Welleck Xiang Ren Allyson Ettinger Zaid Harchaoui and Yejin Choi. Their work is presented in "Faith and Fate: Limits of Transformers on Compositionality," which consists of 10 pages plus an appendix (21 pages). The researchers also provide theoretical arguments on abstract multi-step reasoning problems that highlight how Transformers' performance will rapidly decay with increased task complexity.