Explaining Sources of Uncertainty in Automated Fact-Checking

Jingyi Sun, Greta Warren, Irina Shklovski, Isabelle Augenstein

TL;DR

The CLUE framework (Conflict-&Agreement-aware Language-model Uncertainty Explanations) moves beyond simple numerical confidence scores in fact-checking by generating natural language explanations that pinpoint why a model is unsure. By identifying and verbalizing specific conflicts or agreements between claim and evidence spans, CLUE provides "plug-and-play" transparency for any white-box LLM without requiring additional fine-tuning.

What Problem Does This Paper Address?

As Large Language Models (LLMs) are increasingly deployed for high-stakes fact-checking, a major hurdle remains: uncertainty is a black box. When a model predicts a claim is "False" with 73% certainty, a human fact-checker has no way of knowing if that 27% doubt stems from a minor linguistic ambiguity or a fundamental contradiction between two primary sources.

Existing methods either provide raw numerical scores (which lack context) or use "hedging" language (e.g., "I'm not sure, but..."), which fails to explain the underlying reasoning. This paper addresses the gap in faithful explainability—ensuring that the explanation provided to the user actually reflects the internal conflicts driving the model's predictive uncertainty.

Methods

The authors introduce CLUE, a three-stage pipeline designed to ground uncertainty in text:

  1. Uncertainty Scoring: The system calculates the predictive entropy of the model’s output logits. This provides a mathematically grounded measure of how "confused" the model is between the labels Supports, Refutes, and Neutral.
  2. Conflict/Agreement Extraction: This is the "secret sauce." Instead of just asking the model to explain itself, CLUE uses internal attention weights to identify which specific spans of text in the claim and evidence are interacting. It treats tokens as nodes in a graph and uses the Louvain algorithm (a community detection method) to cluster these into meaningful text spans. These spans are then labeled as agree, disagree, or unrelated using a teacher model (GPT-4o).
  3. Explanation Generation: CLUE uses two strategies to turn these spans into a narrative:
  4. CLUE-Span: A three-shot prompt that feeds the extracted spans back into the model as "soft constraints."
    • CLUE-Span+Steering: This adds Attention Steering (PASTA), which dynamically modifies the model's attention heads during inference to force it to focus on the identified conflict/agreement spans while writing the explanation.

Datasets

The framework was evaluated on two multi-evidence fact-checking datasets:

  • HealthVer: A domain-specific dataset focused on COVID-19 and health claims.
  • DRUID: A more complex, real-world dataset containing claims paired with heterogeneous online evidence sources, often involving longer-form articles.

Key Insights and Ideas

  • The Faithfulness Gap: The authors found that standard prompting (simply asking a model to explain its uncertainty) is often non-faithful. In fact, baseline explanations often highlighted tokens that had zero impact on the model's actual uncertainty.
  • Attention Steering as a "Nudge": By using PASTA to steer attention, the authors could "instinctively" guide the model to talk about the right things without changing the underlying weights of the model.
  • The Faithfulness-Plausibility Trade-off: There is a tension between an explanation being "faithful" (accurately reflecting the model's messy internal state) and "plausible" (sounding good to a human). While attention steering made explanations more faithful, human evaluators sometimes found the non-steered version slightly more "fluent," suggesting that the most accurate explanation isn't always the most polished one.

Key Results

  • Faithfulness (Entropy-CCT): CLUE variants significantly outperformed baselines. On the DRUID dataset with the Qwen2.5-14B model, CLUE-Span+Steering achieved a faithfulness correlation of 0.102, compared to a -0.080 (negative correlation) for the standard prompting baseline.
  • Human Evaluation: In a study with 12 participants, CLUE-generated explanations were ranked as more helpful, more informative, and less redundant than the baseline.
  • Label Consistency: CLUE achieved higher Label-Explanation Entailment (LEE) scores, meaning the generated explanation was logically consistent with the final "True/False" verdict more often than the baseline.

Limitations and Future Work

  • Model Scale: The study focused on medium-sized models (9B to 14B parameters). The authors suggest that larger models (70B+) might close the gap in fluency and instruction-following.
  • Evidence Complexity: For simplicity, the current framework only looks at interactions between a claim and two pieces of evidence. Real-world fact-checking often involves dozens of sources, which would require more complex graph partitioning.
  • Sources of Uncertainty: CLUE specifically targets uncertainty arising from conflicting evidence. It does not yet address uncertainty caused by "knowledge gaps" (where the model simply hasn't seen the information before) or "context-memory conflicts."
  • Expert Validation: While laypeople found the explanations helpful, the authors note that future work should involve professional fact-checkers to see if CLUE actually improves their workflow speed or accuracy.

Summary updated

Abstract

Understanding sources of a model's uncertainty regarding its predictions is crucial for effective human-AI collaboration. Prior work proposes using numerical uncertainty or hedges ("I'm not sure, but ..."), which do not explain uncertainty that arises from conflicting evidence, leaving users unable to resolve disagreements or rely on the output. We introduce CLUE (Conflict-and-Agreement-aware Language-model Uncertainty Explanations), the first framework to generate natural language explanations of model uncertainty by (i) identifying relationships between spans of text that expose claim-evidence or inter-evidence conflicts and agreements that drive the model's predictive uncertainty in an unsupervised way, and (ii) generating explanations via prompting and attention steering that verbalize these critical interactions. Across three language models and two fact-checking datasets, we show that CLUE produces explanations that are more faithful to the model's uncertainty and more consistent with fact-checking decisions than prompting for uncertainty explanations without span-interaction guidance. Human evaluators judge our explanations to be more helpful, more informative, less redundant, and more logically consistent with the input than this baseline. CLUE requires no fine-tuning or architectural changes, making it plug-and-play for any white-box language model. By explicitly linking uncertainty to evidence conflicts, it offers practical support for fact-checking and generalises readily to other tasks that require reasoning over complex information.
BibTeX
@inproceedings{sun2026explaining,
  author = {Jingyi Sun and Greta Warren and Irina Shklovski and Isabelle Augenstein},
  title = {{Explaining Sources of Uncertainty in Automated Fact-Checking}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.2110},
}
Figure