acceptodds
Withdrawn from ICLR 2027

DEDUCE: Robust Fact-Checking under Knowledge Poisoning via Claim Decomposition and Evidence Reweighting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Automated fact-checking systems traditionally operate under the “clean-knowledge” assumption: the premise that retrieved evidence is inherently reliable. However, in open-domain environments, knowledge bases may contain fabricated or misleading content, causing substantial degradation when misinformation is retrieved as evidence. We formalize the task of Fact-Checking under Contamination (FCC) and propose DEDUCE, a robust verification framework designed for contaminated knowledge environments. DEDUCE combines claim decomposition, claim-conditioned stance aggregation, and dynamic passage-level reliability estimation in a closed-loop retrieval process, where learned reliability signals are fed back into subsequent retrieval through reranking. To rigorously evaluate this setting, we introduce a controlled stress test with topically aligned, adversarially generated distractors and complement it with naturally occurring misinformation. Across seven benchmarks and four open-source LLM backbones, contamination substantially degrades standard fact-checking pipelines, while DEDUCE consistently recovers a significant portion of the lost performance. Additional comparisons on HoVer and EX-FEVER show that DEDUCE also outperforms purpose-built poisoning defenses, while further stress tests reveal that coordinated and reputation-based misinformation remain challenging when misleading evidence becomes highly consistent or inherits previously accumulated reliability. These results demonstrate the importance of explicitly modeling evidence reliability and retrieval feedback for robust fact checking under contaminated knowledge.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Misinformation, Fact Checking, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tAK5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the problem of automatic fact checking when the knowledge base is polluted by false information, and proposes the DEDUCE framework. This method decomposes complex statements into sub statements, uses the big language model to identify the support, refutation and irrelevant positions of retrieval evidence, dynamically updates the reliability of paragraphs according to the consistency of evidence and aggregation judgment, and feeds back the reliability to subsequent retrieval reordering. The author verifies the effectiveness of the method on seven datasets and four open source models. At the same time, through collaborative false information, mixed true and false evidence, and reputation poisoning experiments, the author analyzes the application boundaries of the consistency driven mechanism.

Critical strengths
  1. The paper focuses on the fact verification when the knowledge base is polluted by false information, breaking through the common assumption that retrieval evidence is natural and reliable. The evaluation combined with directional synthetic pollution and natural false information is helpful to understand the performance of the fact verification system in an unreliable knowledge environment.

  2. The method combines statement decomposition, evidence position aggregation, paragraph reliability update and retrieval reordering, so that historical verification information can affect subsequent evidence selection. This feedback mechanism provides a feasible idea for mitigating the interference of pollution evidence on multi-step verification.

  3. The paper verifies the performance improvement on seven datasets and four open source models, and compares with multi class methods. The cooperation of false information, mixed true and false paragraphs and reputation poisoning experiments further reveal the applicable boundaries of the method.

Critical weaknesses
  1. The method adjusts the weight according to the consistency of evidence and current aggregate judgment, but consensus consistency is not equal to fact correctness. When the initial judgment error or false evidence is highly consistent, the feedback mechanism may amplify the error, and the conditions for its effective operation and error recovery ability need to be analyzed.

  2. The paper does not clearly give the fusion formula of reliability and retrieval relevance, the size of candidate pool and the score normalization method. The update extent directly depends on the aggregation score difference, so the retrieval score scale may affect the weight evolution, and the lack of relevant details will affect the reproduction and mechanism understanding.

  3. Existing ablation does not completely isolate the role of aggregation weighting and retrieval feedback, and the relationship between different ablation configurations needs to be clarified. Compared with some strong baselines, the improvement is relatively small. It is suggested to supplement strictly matched control experiments and confidence intervals to enhance the persuasiveness of the core conclusions.

Questions for authors

1.How to fuse reliability and retrieval relevance for reordering? How to set the size of the candidate pool, the score normalization method, and the update range?

2.Under the same paragraph granularity, candidate evidence and position input, what results will be obtained by enabling aggregation weighting and retrieval feedback respectively? What are the differences between the "single" in Table 8 and the "no feedback" in Table 9?

3.How to divide calibration set and test set for frozen weight calibration experiment? Whether the overlapping of topics, entities and paragraphs is controlled, and whether the comparison method gets the same opportunity for calibration or online update?

4.Compared with TruthFinder and robust RAG methods, whether the performance improvement is significant under multiple independent runs or paired statistical tests?

Reviewer_CamxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes DEDUCE, a fact checking method with poisoned retrieved evidence. It combines claim decomposition, LLM based passage stance classification, persistent passage reliability estimates, and reliability aware retrieval reranking. Reliability is updated according to agreement with the system’s aggregate subclaim predictions. The experiments cover seven datasets and four verifier backbones, including both synthetic contamination and naturally occurring misleading evidence.

The problem is interesting, and the empirical breadth is a strength. However, I have concerns about the interpretation of the reliability updates, the isolation of the methodological contribution, and missing details in the central retrieval feedback and evaluation protocols.

Critical strengths
  1. The paper focuses on an important failure mode of retrieval based fact checking. Experiments on multiple datasets and backbones, together with naturally occurring misleading evidence, provides useful evidence that the problem is not restricted to a single benchmark.

  2. Table 1 reports improvements across all evaluated dataset backbone combinations. The sequential ablation also helps distinguish the contributions of decomposition, stance aggregation, and updating.

  3. I appreciate the experiments on coordinated misinformation and reputation poisoning, as well as the analysis of persistent versus reset reliability states. The authors explicitly acknowledge that agreement is not equivalent to truth and do not hide substantial performance degradation under stronger attacks.

Critical weaknesses
  1. The overall architecture of DEDUCE presented in Figure 1 doesn't seem consistent with what is described in the text of the paper and several of the components are not sufficiently or clearly introduced in the paper. For example, there is a component "Reranker" mentioned in Figure 1, but Reranker is never introduced in the paper, though the author introduces reranking in the paper, which I assume is about ReRanker. There is a component named "Stance-weighted Aggregation", but this is never clearly described or introduced in the paper. These make this paper really difficult to follow.

  2. It seems to me that reranking is a critical step, but it is described verbally without an explicit scoring function. Appendix A specifies hybrid retrieval and K = 10, but does not clearly specify the candidate pool before reliability reranking, score normalization, or whether the scores used in stance aggregation already incorporate reliability. These choices matter because the unnormalized decision margin also determines the update magnitude. If reranking only changes the order of a fixed set of ten passages that are all subsequently aggregated, it does not change evidence membership. The paper needs to explain precisely how reliability affects the selected evidence set and its scores.

  3. Table 9 changes several factors simultaneously, including source versus passage granularity, the reliability estimator, and retrieval feedback. Full DEDUCE outperforms TruthFinder by only 1.41 and 1.09 Macro-F1 points, while DEDUCE without feedback performs worse than TruthFinder on both datasets. These results suggest that feedback is important. However, they do not establish that the proposed estimator is particularly effective when feedback is controlled. A more informative comparison would apply the same candidate retrieval, passage representation, etc. The closest baseline comparisons also need uncertainty estimates. The claim order experiments are useful, but do not replace paired uncertainty estimates for method differences or variation across contamination realizations.

  4. Another major concern is about the comparative evidence, which remains concentrated on synthetic contamination. The inclusion of RAGuard is valuable. However, the robust RAG and truth discovery comparisons are reported only on HoVer and EX-FEVER. It is therefore unclear whether the gains over the strongest alternatives extend to naturally occurring misleading evidence.

The synthetic validation also establishes a weaker property than the attack objective. Finding at least one false statement in a passage does not establish that the passage contradicts the target claim or supports its intended false alternative. This distinction is particularly relevant given the mixed truth passages discussed in the paper. A target specific audit should assess factual falsity, stance toward the target, and ambiguity separately.

Questions for authors
  1. Please ensure each component mentioned in the framework architecture is properly introduced in the manuscript.

  2. What is the exact reliability aware reranking equation, candidate pool size, and score normalization procedure? Which scores enter the stance aggregation?

  3. How do the protocols in Table 13 differ operationally, and how is the held-out calibration split constructed?

  4. How are shared latent entities and cross subclaim dependencies resolved?

  5. Does DEDUCE retain its advantage when alternative reliability estimators receive the same retrieval feedback mechanism?

  6. How do the strongest baselines perform on RAGuard and under the coordinated and reputation poisoning attacks?

Reviewer_EWncOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

Authors propose a setting for fact checking under which the knowledge base contains deliberate false claims. They also propose a method. Their main idea is to build a profile for each document in the knowledge base based on the number of times that the passages in the document have agreed on the stance with other passages--in the knowledge base--regarding a given sub-claim. This profile is updated over time to weigh the evidence that is extracted from the document. This weight is subsequently used to verify the future claims or sub-claims.

The experimental setup is extensive. The method is tested in seven datasets, and four LLMs are used as the underlying verifier. The experiments show that the improvements are consistent.

Critical strengths
  • The proposed method is very simple to implement, it actually somehow reminded me of the page rank algorithm, but with implicit links between pages
  • The improvements are significant
  • The experiments are extensive
Critical weaknesses
  • I don't think the proposed problem is novel [1]
  • A very common attack which unfortunately we are currently seeing in the US politics is the "flood the zone" tactic. In this scenario the goal is not to generated credible well crafted false claims. But it is to increase the number of false claims, even if they are obviously false. The proposed method fails in this scenario, because it relies on simple majority-voting
  • What if the improvements are due to seeing more evidence documents by the verifier compared to the two baselines, rather than the proposed method itself?

[1] Fact2Fiction: Targeted Poisoning Attack to Agentic Fact-checking System, 2025

Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.