Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Probing Causality Manipulation of Large Language Models

Chenyang Zhang, Haibo Tong, Bin Zhang, Dongyu Zhang

TL;DR

This paper tackles whether large language models possess intrinsic causality manipulation capabilities. It introduces a hierarchical probing framework that combines a causality-focused dataset with retrieval-augmented generation and in-context learning to guide LLMs in causal judgments. Empirical results show LLMs can detect causal entities and direct relations to some extent, but rely primarily on global sentence semantics and lack a dedicated causal processing pathway, with performance limited on more complex causal tasks. The work provides a structured methodology and dataset for evaluating causal cognition in LLMs and highlights directions for improving causal reasoning through targeted training and knowledge integration.

Abstract

Large language models (LLMs) have shown various ability on natural language processing, including problems about causality. It is not intuitive for LLMs to command causality, since pretrained models usually work on statistical associations, and do not focus on causes and effects in sentences. So that probing internal manipulation of causality is necessary for LLMs. This paper proposes a novel approach to probe causality manipulation hierarchically, by providing different shortcuts to models and observe behaviors. We exploit retrieval augmented generation (RAG) and in-context learning (ICL) for models on a designed causality classification task. We conduct experiments on mainstream LLMs, including GPT-4 and some smaller and domain-specific models. Our results suggest that LLMs can detect entities related to causality and recognize direct causal relationships. However, LLMs lack specialized cognition for causality, merely treating them as part of the global semantic of the sentence.

Probing Causality Manipulation of Large Language Models

TL;DR

This paper tackles whether large language models possess intrinsic causality manipulation capabilities. It introduces a hierarchical probing framework that combines a causality-focused dataset with retrieval-augmented generation and in-context learning to guide LLMs in causal judgments. Empirical results show LLMs can detect causal entities and direct relations to some extent, but rely primarily on global sentence semantics and lack a dedicated causal processing pathway, with performance limited on more complex causal tasks. The work provides a structured methodology and dataset for evaluating causal cognition in LLMs and highlights directions for improving causal reasoning through targeted training and knowledge integration.

Abstract

Large language models (LLMs) have shown various ability on natural language processing, including problems about causality. It is not intuitive for LLMs to command causality, since pretrained models usually work on statistical associations, and do not focus on causes and effects in sentences. So that probing internal manipulation of causality is necessary for LLMs. This paper proposes a novel approach to probe causality manipulation hierarchically, by providing different shortcuts to models and observe behaviors. We exploit retrieval augmented generation (RAG) and in-context learning (ICL) for models on a designed causality classification task. We conduct experiments on mainstream LLMs, including GPT-4 and some smaller and domain-specific models. Our results suggest that LLMs can detect entities related to causality and recognize direct causal relationships. However, LLMs lack specialized cognition for causality, merely treating them as part of the global semantic of the sentence.
Paper Structure (45 sections, 5 figures, 6 tables)

This paper contains 45 sections, 5 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Probing LLMs
  4. Evaluation of Causality for LLMs
  5. Dataset Construction
  6. Base Dataset
  7. Classification Dataset Construction
  8. Probing Design
  9. Hierarchical Retrieval Augmented Generation
  10. Layer 1: No Augmentation
  11. Layer 2: Original CMedCausal
  12. Layer 3: Universal Medical-KG
  13. Retrieval Augmented Generation
  14. In Context Learning Design
  15. Experiments
  16. ...and 30 more sections

Figures (5)

  • Figure 1: Main stricture of our probing works. We construct a causal dataset, then guide models by providing shortcuts. Finally, we probe intrinsic manipulation of causality by comparing performances of different shortcuts.
  • Figure 2: An instance of our constructed datasets. Causes in sentences are bold and effects are underlined. The corresponding causes and effects are marked with the same color.
  • Figure 3: Trend of MCC with three actions in different models and different layers (Using Advanced Prompt).
  • Figure 4: Flowchart of secondary retrieval using langchain, where "encoding" means encoding the input text using the Sentence Transformer, "Calculate Similarity" means calculate the similarity score using the cosine similarity, "Search Description" Indicates the description of the corresponding medical text in the knowledge graph, "Spliting and Encoding" means that the description text is chunked and encoded and "Decoding" means decoding the encoded vector into a sentence.
  • Figure 5: PPL of positive and negative instances in different actions calculated by GPT-2