Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?

Neeladri Bhuiya, Viktor Schlegel, Stefan Winkler

TL;DR

This paper finds evidence that LLMs indeed circumvent the requirement to perform multi-hop reasoning, but they do so in more subtle ways than what was reported about their fine-tuned pre-trained language model (PLM) predecessors, and proposes a challenging multi-hop reasoning benchmark.

Abstract

State-of-the-art Large Language Models (LLMs) are accredited with an increasing number of different capabilities, ranging from reading comprehension, over advanced mathematical and reasoning skills to possessing scientific knowledge. In this paper we focus on their multi-hop reasoning capability: the ability to identify and integrate information from multiple textual sources. Given the concerns with the presence of simplifying cues in existing multi-hop reasoning benchmarks, which allow models to circumvent the reasoning requirement, we set out to investigate, whether LLMs are prone to exploiting such simplifying cues. We find evidence that they indeed circumvent the requirement to perform multi-hop reasoning, but they do so in more subtle ways than what was reported about their fine-tuned pre-trained language model (PLM) predecessors. Motivated by this finding, we propose a challenging multi-hop reasoning benchmark, by generating seemingly plausible multi-hop reasoning chains, which ultimately lead to incorrect answers. We evaluate multiple open and proprietary state-of-the-art LLMs, and find that their performance to perform multi-hop reasoning is affected, as indicated by up to 45% relative decrease in F1 score when presented with such seemingly plausible alternatives. We conduct a deeper analysis and find evidence that while LLMs tend to ignore misleading lexical cues, misleading reasoning paths indeed present a significant challenge.

Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?

TL;DR

This paper finds evidence that LLMs indeed circumvent the requirement to perform multi-hop reasoning, but they do so in more subtle ways than what was reported about their fine-tuned pre-trained language model (PLM) predecessors, and proposes a challenging multi-hop reasoning benchmark.

Abstract

State-of-the-art Large Language Models (LLMs) are accredited with an increasing number of different capabilities, ranging from reading comprehension, over advanced mathematical and reasoning skills to possessing scientific knowledge. In this paper we focus on their multi-hop reasoning capability: the ability to identify and integrate information from multiple textual sources. Given the concerns with the presence of simplifying cues in existing multi-hop reasoning benchmarks, which allow models to circumvent the reasoning requirement, we set out to investigate, whether LLMs are prone to exploiting such simplifying cues. We find evidence that they indeed circumvent the requirement to perform multi-hop reasoning, but they do so in more subtle ways than what was reported about their fine-tuned pre-trained language model (PLM) predecessors. Motivated by this finding, we propose a challenging multi-hop reasoning benchmark, by generating seemingly plausible multi-hop reasoning chains, which ultimately lead to incorrect answers. We evaluate multiple open and proprietary state-of-the-art LLMs, and find that their performance to perform multi-hop reasoning is affected, as indicated by up to 45% relative decrease in F1 score when presented with such seemingly plausible alternatives. We conduct a deeper analysis and find evidence that while LLMs tend to ignore misleading lexical cues, misleading reasoning paths indeed present a significant challenge.
Paper Structure (32 sections, 3 figures, 11 tables)

This paper contains 32 sections, 3 figures, 11 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. I. Acquiring the main entity
  5. II. Extracting the details
  6. III. Creating the distractor paragraphs
  7. Data Quality
  8. Experiment Setup
  9. Do LLMs suffer from the same flaws as fine-tuned models?
  10. Do LLMs get distracted by seemingly plausible alternate reasoning paths?
  11. What are the effects of the different parameters?
  12. Experiment Results
  13. Do LLMs suffer from the same flaws as fine-tuned models?
  14. I. Setting up the baseline
  15. II. Reasoning shortcuts using SubQA
  16. ...and 17 more sections

Figures (3)

  • Figure 1: Our proposed method evaluates the multi-hop reasoning capabilities of Large Language Models by adding seemingly plausible, yet ultimately wrong alternate reasoning paths, impacting the reasoning performance of state-of-the-art LLMs such as GPT-4.
  • Figure 2: Example of a decomposed multi-hop question.
  • Figure 3: Instantiation of our proposed method. With "arena" as main entity of sub-question 1, we extract "home" to be replaced with "playoff". Then, we use the modified sequence with the original sub-question 2 (masking the answer "Androscoggin Bank Colisée") as prompt to GPT-4 to generate the distractor paragraphs 1 and 2. The distractor paragraphs generated have "Maple Leaf Arena" as the bridging entity in the false reasoning chain which leads to the wrong answer "4500 spectators".