Are Machines Better at Complex Reasoning? Unveiling Human-Machine Inference Gaps in Entailment Verification

Soumya Sanyal, Tianyi Xiao, Jiacheng Liu, Wenya Wang, Xiang Ren

Introduction

Natural Language Inference (NLI) has long been a cornerstone of language model evaluation, yet traditional benchmarks often rely on short, single-sentence premises. In real-world applications, reasoning is rarely so contained. "Entailment Verification" (EV)—the task of determining if a multi-sentence context supports a specific claim—is essential for verifying the accuracy of model-generated explanations. This paper investigates the "inference gap" between humans and Large Language Models (LLMs), revealing that machines and humans possess fundamentally different reasoning profiles.

The Human-Machine Inference Gap

To understand where machines succeed and humans fail (and vice versa), the researchers categorized reasoning into four distinct types:

  • R1: Simple Deductive: Logic involving negations or simple word substitutions.
  • R2: Complex Deductive: Multi-hop reasoning that requires connecting disparate facts.
  • R3: Missing Knowledge: Reasoning requiring external information.
  • R4: Missing Localized Knowledge: Reasoning requiring specific context provided in the premise.

The Memory vs. Logic Trade-off

The study identifies a surprising role reversal. Humans are significantly more reliable at simple deductive logic (R1). However, LLMs outperform humans in complex, multi-hop reasoning (R2) over long contexts. This is attributed to the "Human Memory Limit"; while humans are constrained by a short-term memory capacity of roughly four information chunks, LLMs act as "stronger context processors," capable of linking facts across extensive premises that overwhelm human annotators.

Methodology: Building a Specialized Verifier

The authors developed a comprehensive approach to bridge these gaps and improve model reliability:

  1. Benchmark Compilation: They aggregated 10 datasets across three domains: NLI (WaNLI, FEVER, ANLI), Contextual QA (CosQA, SIQA, DREAM, BoolQ, RACE), and human-annotated rationales (Entailer, ECQA).
  2. Ranking-based Finetuning: The researchers finetuned a Flan-T5-xxl model using a novel ranking objective. Rather than simple binary classification ("support" vs. "not support"), the model was trained to rank correct hypotheses higher than "distractor" hypotheses. This creates a softer decision boundary that captures the nuance of relative support.
  3. Chain-of-Thought (CoT) Filtering: The verifier was integrated into a "Self-Consistency" pipeline. Instead of relying on simple majority voting for model answers, the verifier discarded any reasoning paths where the generated explanation failed to actually entail the predicted answer.

Key Results

Model Performance

The finetuned open-source model (Flan-T5-xxl + Rank) achieved a macro-F1 of 0.78, outperforming GPT-3.5 (0.72) and matching the performance of GPT-4 (0.79). While GPT-4 led humans in complex multi-step reasoning (ANLI, CosQA) by margins of up to 0.12 macro-F1, humans maintained their lead in social commonsense and simple deduction.

Improving Faithfulness

A primary challenge with LLMs is "unfaithful" reasoning—where a model provides a correct answer but justifies it with a nonsensical or contradictory explanation. By using the specialized verifier as a gatekeeper to filter out these unfaithful rationales, the researchers achieved a 6% average accuracy improvement across multiple-choice datasets (CSQA, ARC-e, ARC-c). These gains were most significant for weaker base models, such as UL2.

Conclusion and Future Work

This research highlights that LLMs are not merely "stochastic parrots" but possess distinct advantages in processing complex, multi-hop information that exceeds human cognitive limits. However, their struggle with simple logic remains a hurdle. Future work will look toward testing decoder-only architectures like Llama or Mistral and refining scoring techniques to move beyond token probability toward full-sentence generation objectives. By understanding the specific inference gaps between humans and machines, we can better design verifiers that ensure AI reasoning is both accurate and faithful.

Summary updated

BibTeX
@inproceedings{sanyal2024are,
  author = {Soumya Sanyal and Tianyi Xiao and Jiacheng Liu and Wenya Wang and Xiang Ren},
  title = {{Are Machines Better at Complex Reasoning? Unveiling Human-Machine Inference Gaps in Entailment Verification}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
  doi = {10.18653/v1/2024.findings-acl.618},
}
Figure