Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CLOMO: Counterfactual Logical Modification with Large Language Models

Yinya Huang, Ruixin Hong, Hongming Zhang, Wei Shao, Zhicheng Yang, Dong Yu, Changshui Zhang, Xiaodan Liang, Linqi Song

TL;DR

This work investigates counterfactual reasoning in large language models through the Counterfactual Logical Modification (CLOMO) framework. It introduces the CFLogic benchmark and a logic-aware evaluation framework based on $ V$-information, including the Pointwise $ V$-information (PVI) and Self-Evaluation Scores (SES) to assess how edits affect logical relationships. Experiments with GPT-3.5 and GPT-4 show that models can imitate some counterfactual patterns but lag behind human performance, and SES- and context-based metrics only partially align with human judgments. The study additionally discusses how strengthening counterfactual logical thinking may help reduce LLM hallucinations, with code and data available at the project repository for reproducibility and further research.

Abstract

In this study, we delve into the realm of counterfactual reasoning capabilities of large language models (LLMs). Our primary objective is to cultivate the counterfactual thought processes within LLMs and rigorously assess these processes for their validity. Specifically, we introduce a novel task, Counterfactual Logical Modification (CLOMO), and a high-quality human-annotated benchmark. In this task, LLMs must adeptly alter a given argumentative text to uphold a predetermined logical relationship. To effectively evaluate a generation model's counterfactual capabilities, we propose an innovative evaluation metric, the decomposed Self-Evaluation Score (SES) to directly evaluate the natural language output of LLMs instead of modeling the task as a multiple-choice problem. Analysis shows that the proposed automatic metric aligns well with human preference. Our experimental results show that while LLMs demonstrate a notable capacity for logical counterfactual thinking, there remains a discernible gap between their current abilities and human performance. Code and data are available at https://github.com/Eleanor-H/CLOMO.

CLOMO: Counterfactual Logical Modification with Large Language Models

TL;DR

This work investigates counterfactual reasoning in large language models through the Counterfactual Logical Modification (CLOMO) framework. It introduces the CFLogic benchmark and a logic-aware evaluation framework based on -information, including the Pointwise -information (PVI) and Self-Evaluation Scores (SES) to assess how edits affect logical relationships. Experiments with GPT-3.5 and GPT-4 show that models can imitate some counterfactual patterns but lag behind human performance, and SES- and context-based metrics only partially align with human judgments. The study additionally discusses how strengthening counterfactual logical thinking may help reduce LLM hallucinations, with code and data available at the project repository for reproducibility and further research.

Abstract

In this study, we delve into the realm of counterfactual reasoning capabilities of large language models (LLMs). Our primary objective is to cultivate the counterfactual thought processes within LLMs and rigorously assess these processes for their validity. Specifically, we introduce a novel task, Counterfactual Logical Modification (CLOMO), and a high-quality human-annotated benchmark. In this task, LLMs must adeptly alter a given argumentative text to uphold a predetermined logical relationship. To effectively evaluate a generation model's counterfactual capabilities, we propose an innovative evaluation metric, the decomposed Self-Evaluation Score (SES) to directly evaluate the natural language output of LLMs instead of modeling the task as a multiple-choice problem. Analysis shows that the proposed automatic metric aligns well with human preference. Our experimental results show that while LLMs demonstrate a notable capacity for logical counterfactual thinking, there remains a discernible gap between their current abilities and human performance. Code and data are available at https://github.com/Eleanor-H/CLOMO.
Paper Structure (28 sections, 4 equations, 3 figures, 2 tables)

This paper contains 28 sections, 4 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Evaluation of Large Language Model Reasoning
  4. Counterfactual Reasoning Benchmarks
  5. SES: Self-Evaluation Scores
  6. SES-1
  7. SES-2
  8. CFLogic Benchmark
  9. Task Definition
  10. Benchmark Construction
  11. Data Statistics
  12. Metrics for Modification
  13. Syntactic Match
  14. Embedding-Based Match
  15. Self-Evaluation Scores
  16. ...and 13 more sections

Figures (3)

  • Figure 1: The task of counterfactual modification with logical restriction. An LLM is given an argument and two premises. The LLM needs to modify the statements in Argument such that the logical relation R switch to stand in state 2 instead of state 1.
  • Figure 2: A successful case of counterfactual modification by GPT-4-32k.
  • Figure 3: A successful case of counterfactual modification by GPT-4-32k.