Causal reasoning and LLMs paper

The lates thinking about how people and machines think (reason) and how we might do it together. “Causal Reasoning and Large Language Models:
Opening a New Frontier for Causality”


True reasoning in LLMs has been one of the main problems that fascinates me. I’m glad some people are tackling it!

As much as I love the elegance of Pearl’s Theory of Causal Networks, I always felt that there wasn’t a straightforward way for deep neural networks to use them to improve causal machine learning (at least to the same heights as CNNs and standard image classification).

It would be interesting to see if causal reasoning and LLMs can be used to dramatically improve abductive commonsense reasoning.


Dumb Question incoming…:

As a layman, this just sounds like it’s giving physical context of the world to the language model – ex: “We introduce a challenge dataset, ART, that consists of over20k commonsense narrative contexts and 200k explanations.”

Isn’t this more like “DEductive Reasoning” since it’s deducing what may have occurred based the context provided? I guess I’m having trouble understanding the main difference here in what they’re saying is abductive reasoning (vs deductive reasoning) and how it correlates to the LLMs.

Definitions I looked up -
Abductive Reasoning: Abductive reasoning, or abduction, is making a probable conclusion from what you know. If you see an abandoned bowl of hot soup on the table, you can use abduction to conclude the owner of the soup is likely returning soon.

Deductive Reasoning: Deductive reasoning is a logical process in which a conclusion is based on the concordance of multiple premises that are generally assumed to be true . Deductive reasoning is sometimes referred to as top-down logic. Deductive reasoning relies on making logical premises and basing a conclusion around those premises.

I think the confusion between deductive and abductive reasoning is mostly pop culture’s fault. In certain media, like Sherlock Holmes and Star Trek (a favorite of mine), words like deductive, logical, abductive, and reasonable seem to be synonymous.

Deductive reasoning (in its strict, academic form) is not necessarily concerned with truth in and of itself, but what must follow from true premises. For example:
(1) Fido is a dog. (2) All dogs are mammals. => (3) Therefore, Fido is a mammal. (modus ponens)
Here, I am not concerned if (1) and (2) are false. I am concerned with what MUST follow IF (1) and (2) are true. I am concerned with logical inference to form valid arguments, not necessarily truth (but valid arguments based on false assumptions are at best useless and deceptive at worst). I think this link covers it pretty well: Deductive reasoning - Wikipedia

Abductive reasoning, on the other hand, IS concerned with the truth. Here, we rely on intuition and experience to predict what is likely to be true. In the soup example, imagine a robot that can’t reason abductively. The robot would say it’s possible that the owner of the soup will not return for many other reasons (didn’t like the soup, had a family emergency, etc.), and could not make a conclusion. Someone who can abductively reason, and has experience with similar situations, might appeal to their intuition and look at the context to conclude the owner of the soup will return. In abductive reasoning, you use the word “probable” which is another important distinction. Again, deductive reasoning is concerned with inferring conclusions that MUST be true, not probably true.

I hope this helps.

1 Like

Ted, greatly appreciate the links as well as the layman breakdown – thank you much!

Makes a lot more sense now.