PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Verdicts: A Graph-Based Analysis of Human and LLM Reasoning in Scientific Fact-Checking

Abdul Ghafoor, Muhammad Arslan Manzoor, Yufang Hou

arXiv:2608.23047v1cs.CLcs.AI

TL;DR

Scientific fact-checking needs to distinguish correct verdicts from the reasoning paths used to interpret cited studies. The paper introduces typed reasoning graphs and validates divergent paths, finding that verdict performance, human alignment, and reasoning validity are distinct dimensions.

  • Problem

    Existing LLM fact-checking evaluations assess verdicts and explanations but do not establish whether reasoning follows human expert paths or uses different valid paths when cited studies are misrepresented.

  • Method

    The framework converts human and LLM explanations into typed reasoning graphs, aligns fallacy-specific sub-graphs, and validates non-aligned paths for study grounding, claim relevance, and verdict sufficiency.

  • Results

    Across 84 MISSCIPLUS false claims, Qwen3-32B had the strongest verdict-level performance, GPT-5 the highest human alignment, and many non-aligned paths remained valid.

  • Takeaways & Limitations

    Evaluation should move beyond verdict accuracy to distinguish human-aligned, valid alternative, ungrounded, and grounded-but-inadequate reasoning.

  • Takeaways & Limitations

    The evaluation uses 84 English-language biomedical and health claims, lacks significance tests, and relies on single-annotator, non-blinded annotation procedures.

Abstract

from arXiv · show

Misinformation that cites legitimate papers can be especially harmful when it distorts what those studies actually report. While existing automatic fact-checking systems based on large language models (LLMs) can assess whether a model assigns an Incorrect verdict and can gen- erate explanations for that decision, they typi- cally do not indicate whether the model follows the same reasoning path as human experts or arrives at the verdict through a different but still valid path. In this work, we introduce a graph- based framework (typed reasoning graph) for comparing human and LLM reasoning paths in scientific fact-checking. Building on prior work on fallacious reasoning in biomedical misinformation, MISSCIPLUS (Glockner et al., 2025), we model each explanation as a rea- soning graph that links the false claim to the relevant study context, study findings, fallacy- supporting premises, and fallacy labels. This representation enables one-to-one alignment of human and LLM reasoning at the level of fallacy-specific sub-graphs. For non-human- aligned LLM paths, we validate grounding in the cited study, relevance to the claim, and suf- ficiency for the verdict. Using 84 false claims from MISSCIPLUS, we evaluate GPT-5, Claude Opus 4.7, and Qwen3-32B across prompt and evidence settings. Results show distinct perfor- mance dimensions: Qwen3-32B has the lowest verdict failure rate, GPT-5 the highest human alignment, and Claude Opus 4.7 weak verdict prediction but often valid reasoning in success- ful cases

1 Introduction

The paper argues that scientific fact-checking must evaluate how models interpret cited evidence and construct reasoning, not only whether they reach an Incorrect verdict. It introduces typed reasoning graphs to align human and LLM fallacy-specific paths and validate divergent but potentially valid alternatives.

  • Cited misinformation can distort legitimate studies, making evidence interpretation and verdict-supporting reasoning central fact-checking concerns.
  • Typed reasoning graphs link claims, study context, study findings, fallacy-supporting premises, and fallacy labels for process-level comparison.
  • Two explanations can share an Incorrect verdict while using different study evidence, premises, or fallacy types, differences that verdict accuracy alone misses.
  • Non-human-aligned paths are evaluated for grounding in the cited study, relevance to the claim, and sufficiency for the verdict rather than rejected automatically.
  • 32.1% of GPT-5, 8.3% of Claude Opus 4.7, and 15.5% of Qwen3-32B reasoning was human-aligned under the full-study, detailed-prompt setting.The evaluation used 84 false claims from MISSCIPLUS across prompt and evidence settings.
  • The study concludes that verdict accuracy and reasoning-path quality dissociate, with Qwen3-32B strongest on Incorrect recall and GPT-5 strongest on human-aligned reasoning.

2 Problem Definition and Method

The method treats human–LLM comparison as graph alignment and validation: explanations become typed reasoning graphs, fallacy-specific sub-graphs are aligned, and divergent paths are tested for grounding, relevance, and sufficiency.

  • Framework Overview: The framework aggregates repeated verdict outputs, constructs shared reasoning graphs for Incorrect cases, aligns fallacy-specific sub-graphs, and validates non-aligned paths.
  • 2.1 Task Formulation: Each instance pairs a misleading scientific claim with a cited primary study, while human and LLM outputs contain verdicts and natural-language explanations.
  • 2.2 Reasoning Graph Construction: A typed reasoning graph represents Claim, Study Context, Study Findings, Fallacy-Supporting Premises, and Fallacies as node classes connected in chain order.
  • 2.3 Reasoning-Path Alignment: Alignment requires the same fallacy type and semantic alignment of text components, with component pairs matched at an STS score of at least 3.Three LLM judges aggregate scores by majority vote, using the median when no score receives two votes.
  • 2.4 Validation of Non-Aligned Sub-graphs: Grounding validation checks whether study context, findings, and fallacy-supporting premises are supported by the full cited study.Grounded non-aligned sub-graphs then receive relevance and sufficiency scores from three LLM judges.

3 Experimental Setup

The experiments use MISSCIPLUS false-claim–study pairs to compare three LLM fact-checkers across prompts and evidence settings, with repeated runs used to measure verdict behavior and construct reasoning graphs.

  • 3.1 Dataset: The MISSCIPLUS test set contains 84 false scientific claims paired with cited primary studies, relevant passages, and expert reviews explaining the evidence mismatch.
  • 3.1 Dataset: Models receive either claim-relevant selected passages or the full cited study as evidence.
  • 3.2 LLM Verdict and Explanation Generation: Each configuration combines a model, prompt template, and evidence setting, using concise or detailed prompts and three repeated runs per input.
  • 3.2 LLM Verdict and Explanation Generation: The per-claim verdict proportion counts runs assigned to each label, and these proportions are averaged across all N claims.
  • 3.2 LLM Verdict and Explanation Generation: Because every claim is false or misleading, Incorrect rate measures recall on the expected label, while Correct denotes Type-I errors and No Response covers invalid or unusable outputs.
  • 3.3 Reasoning-Graph Analysis: Reasoning-graph analysis uses full-study evidence with the detailed prompt and selects a valid expected-Incorrect explanation from three runs for graph construction.
  • 3.3 Models: The evaluated fact-checkers are GPT-5, Claude Opus 4.7, and Qwen3-32B, with the same claim, evidence, and prompt template within each configuration.

4 Results and Discussion

The evaluation separates verdict prediction from reasoning-path quality, showing that models differ in accuracy, human alignment, and validity of alternative reasoning. Evidence and prompting affect models differently, while verdict correctness alone can mask reasoning differences.

  • Verdict Prediction: GPT-5’s Incorrect rate rises from 32.54% to 74.60% with full-study evidence and detailed prompting.For selected passages, its Incorrect rate rises from 42.46% to 69.05%.
  • Verdict Prediction: Qwen3-32B achieves the highest Incorrect rate: 76.98% with selected passages and 73.41% with full-study evidence, both under the detailed prompt.Because all dataset instances are false claims, higher Incorrect rates indicate higher recall of the expected label.
  • Verdict Prediction: Evidence length has model-dependent effects: GPT-5 performs best with full studies, whereas Qwen3-32B performs best with selected passages.Selected passages may reduce distraction for some models, while full studies may provide information needed for broader interpretation.
  • Verdict Prediction: Claude Opus 4.7 reaches at most 44.58% Incorrect, yet achieves 100% consistency across repeated detailed-prompt settings.Its stability therefore does not ensure correct verdict assignment.
  • Reasoning-Path Alignment and Validation: GPT-5 has 19.0% human-aligned sub-graphs and 81.6% accepted overall, while Qwen3-32B has 10.2% aligned and 80.3% accepted.Accepted paths include human-aligned sub-graphs and validated alternatives that are grounded, relevant, and sufficient.
  • Reasoning-Path Alignment and Validation: Claude Opus 4.7 has the highest accepted proportion at 86.9%, despite its lower Incorrect verdict rate.Its successful cases therefore often contain grounded and adequate reasoning even though it frequently predicts the wrong verdict.
  • Reasoning-Path Alignment and Validation: Cross-model fallacy agreement is moderate, with Cohen’s κ values of 0.511, 0.466, and 0.428 across model pairs.This model sensitivity supports evaluating reasoning components rather than relying only on final verdicts.
  • Reasoning-Path Alignment and Validation: Many non-human-aligned sub-graphs remain grounded, relevant, and sufficient, while rejected sub-graphs reveal unsupported, irrelevant, or insufficient justifications.The framework distinguishes valid alternative reasoning from defective reasoning rather than treating all divergence as error.

5 Conclusion

The framework compares LLM reasoning with human expert paths while validating divergent paths for grounding, relevance, and sufficiency. Results show that verdict accuracy and reasoning quality can diverge, supporting explanation aids but not replacing human judgment.

  • Qwen3-32B performs best at verdict prediction, whereas GPT-5 produces more human-aligned reasoning.
  • Many reasoning paths that do not align with human paths remain valid alternatives.
  • LLM systems can help experts inspect candidate reasoning paths and help readers understand how claims misrepresent cited evidence.
  • Final judgments should remain human-led because models can produce unsupported, irrelevant, or insufficient reasoning paths.

Limitations

The study’s conclusions are constrained by its small, narrow evaluation set, single-annotator procedures, all-false-claim design, and possible training-data exposure. These limitations make small model differences and some alignment comparisons indicative rather than conclusive.

  • Sample size and scope: 84 false claims limit the statistical resolution of cross-model comparisons, and unreported uncertainty means small differences should be treated as indicative.The dataset is restricted to English-language biomedical and health claims, limiting generalization to other languages.
  • Human annotation protocol: A single unblinded annotator performed gold extraction and grounding re-checks, so inter-annotator agreement and adjudicated double annotation were not measured.The grounding re-check shifts the final Grounded column by 5–19 percentage points relative to the QA-only baseline.
  • Evaluation set design: Because all instances are false claims, Incorrect rate measures recall on the expected label rather than overall verdict accuracy.A held-out set of correctly supported claims would be needed to compute precision on Incorrect and separate verdict-class bias from genuine fact-checking ability.
  • Training-data exposure: Public training-data exposure may inflate proprietary-model human-alignment rates relative to Qwen3-32B, and the ten-claim contamination check cannot bound this effect.

A Related Work

Prior work covers scientific verification, scientific-evidence misrepresentation, LLM explanations, structured rationales, and grounding evaluation. The framework connects these previously complementary perspectives by aligning human and LLM explanations through fallacy-specific reasoning paths.

  • Scientific and Biomedical Claim Verification: Scientific fact-checking benchmarks evaluate evidence retrieval, rationale selection, and veracity prediction across biomedical and public-health claims.
  • Misrepresentation of Scientific Evidence: MISSCI and MISSCIPLUS model misleading scientific claims as fallacious arguments grounded in passages from misrepresented publications.
  • LLMs for Fact-Checking and Explanation Generation: LLMs can generate plausible verdicts and explanations, but fluent explanations may not faithfully reflect the underlying decision process.
  • Human Rationales and Structured Reasoning Explanations: Human-rationale, explanation-graph, and argument-mining datasets represent explanations as rationales or connected structures beyond final labels.
  • Grounding and Faithfulness Evaluation: Grounding and faithfulness studies assess support at claim, sentence, citation, or atomic-fact levels, whereas this framework evaluates grounded reasoning paths.
  • The framework connects verdict-focused, fallacy-focused, faithfulness-focused, and generic structured-rationale approaches by aligning human and LLM explanations.

B Fallacy Abbreviations

The supplementary materials define the nine fallacy categories and document prompts for graph extraction, claim assessment, semantic similarity, grounding, and premise verification. They also describe the human–GPT-5 graph-agreement figure and contamination-check table.

  • Fallacy abbreviations: The nine fallacy categories use the MISSCI taxonomy, with abbreviations listed alongside their full names.
  • Figures and tables: Figure 5 summarizes human–GPT-5 graph alignment using fallacy-type agreement, Jaccard similarity, precision, recall, and F1.
  • Contamination check: Table 4 reports contamination-check values for ten false claims published in 2025–2026 outside the study dataset.
  • Reasoning-path comparison: Figure 6 illustrates that an LLM path may align with the human path or take a different but valid path.
  • Evaluation prompts: The prompts cover scientific-claim assessment, semantic similarity, relevance and sufficiency of divergent sub-graphs, and structured graph extraction.
  • Graph extraction: Graph extraction requires study context, study results, and supporting premises to be copied verbatim from the explanation.
  • Premise verification: Premise verification generates exactly five short factual questions for each labeled premise during the QA-grounding stage.
Loading 2608.23047v1…