Summary of Llms Are Prone to Fallacies in Causal Inference, by Nitish Joshi et al.
LLMs Are Prone to Fallacies in Causal Inference
by Nitish Joshi, Abulhair Saparov, Yixin Wang, He He
First submitted to arxiv on: 18 Jun 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Recent work has demonstrated the effectiveness of large language models (LLMs) in extracting causal facts through prompting. However, it remains unclear whether this success is limited to explicitly mentioned causal facts in the pretraining data that the model can memorize. This study investigates whether LLMs can infer causal relations from other relational data in text, such as temporal, spatial, and counterfactual relations. To disentangle the role of memorized causal facts versus inferred causal relations, the authors fine-tune LLMs on synthetic data containing these types of relationships and measure their ability to infer causal relations. The results show that LLMs are susceptible to inferring causal relations from the order of entity mentions in text, even when the order is randomized. Additionally, while LLMs can correctly deduce the absence of causal relations from temporal and spatial relations, they struggle to infer causal relations from counterfactuals, questioning their understanding of causality. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine you have a super smart computer that can read lots of text and learn from it. Recently, people found out that this computer can figure out some important facts about what causes things to happen. But they’re not sure if the computer is only good at finding these facts because it was trained on specific examples of cause-and-effect relationships. This study wants to know if the computer can learn to find cause-and-effect relationships from other kinds of information, like when events happen in a certain order or where things are located. The researchers tested the computer’s ability to figure out what causes things to happen based on different types of clues and found that it’s pretty good at finding simple causal relationships, but struggles with more complex ones. |
Keywords
» Artificial intelligence » Pretraining » Prompting » Synthetic data
