acceptodds
Withdrawn from ICLR 2027

Poisoning by Dilution: Amplifying RAG Attacks with Answer-Free Documents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Retrieval-augmented generation (RAG) is vulnerable to corpus poisoning, where injected passages steer a model toward an attacker-chosen incorrect answer. Existing attacks primarily construct passages that directly support this answer. We investigate whether passages without answer support can also strengthen targeted poisoning. We propose evidence dilution, which combines target-bearing passages with query-relevant, answer-free passages that compete for limited retrieval slots and alter the evidence available to the generator. Our method uses relation-aware generation, semantic checks, and proxy retrieval scores to construct and select dilution passages. We evaluate evidence dilution on Natural Questions, HotpotQA, and MS MARCO, achieving strict attack success rates of 92%, 87%, and 75%, respectively. On Natural Questions, adding three dilution passages increases strict attack success from 84% to 92% in vanilla RAG and from 77% to 91% under InstructRAG-ICL. Evidence analysis links these gains to changes in retrieved evidence and transitions from mixed answers to exclusive adoption of the attacker-chosen answer. These results show that passages can strengthen targeted corpus poisoning without directly supporting the attacker's answer.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-Augmented Generation, Corpus Poisoning, Targeted Attacks, Evidence Dilution, Adversarial Robustness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Z8c6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether query-relevant passages without answer support can strengthen targeted RAG poisoning. The proposed method combines two passages supporting an attacker-chosen answer with three answer-free passages that compete for retrieval slots. It uses relation-aware generation, semantic checks, and proxy retrieval scores to construct and select these passages. Experiments on three QA datasets report improved attack success, with additional analyses across generators, retrievers, and defenses.

Critical strengths

1.The paper examines a useful distinction between query relevance and answer support. Its results suggest that passages need not directly support the attacker’s answer to contribute to targeted poisoning.

2.The evaluation covers three datasets, with additional comparisons across four generators and six retrievers on NQ. Transferring the same attack passages across retrievers provides evidence that the observed gains are not specific to one retriever.

3.The paired analysis fixes the target-bearing passages and tracks changes in retrieval, filtering, and answers. Reporting both gains and losses, including transitions from mixed answers to exclusive target adoption, helps explain the observed ASR changes.

Critical weaknesses
  1. The contribution beyond existing multi-document poisoning is insufficiently established. The main comparison increases the injection budget from two to five passages while keeping retrieval at top-5. This demonstrates that additional answer-free passages can help, but does not establish their advantage over additional target-bearing passages. A comparison between five target-bearing passages and two target-bearing plus three dilution passages, constructed with comparable resources, is missing.

2.The performance gap over PoisonedRAG is not adequately explained. The authors’ two target-bearing passages already outperform five PoisonedRAG passages on all three datasets. On HotpotQA, they achieve 83% versus 67%, while dilution adds only another four percentage points. Thus, much of the reported advantage exists before dilution is introduced. The paper does not sufficiently document baseline construction settings or compare retrieved attack counts to explain this gap. This makes it difficult to assess whether the advantage comes from dilution, stronger target-bearing passages, or differences in baseline configuration.

Questions for authors

1.Why do your two target-bearing passages outperform five PoisonedRAG passages on all three datasets, particularly on HotpotQA (83% vs. 67%)? What accounts for this gap before any dilution passages are added?

2.Which PoisonedRAG variant and construction settings were used? What explains the advantage of your two target-bearing passages over its five passages? Please provide comparable retrieval statistics and generation budgets.

Reviewer_4ru6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether query-relevant but answer-free passages ("dilution" passages) can amplify targeted corpus poisoning in retrieval-augmented generation (RAG). The threat model holds the target-bearing poisoned passages fixed and adds dilution passages that are relevant to the question but do not provide, exclude, or narrow the answer. The authors build these passages with relation-aware generation plus individual and joint answer-free checks, select a dilution group via proxy retrieval scores, and inject them into a shared corpus where they compete through natural top-k retrieval. Across 100-question subsets of NQ, HotpotQA, and MS MARCO, adding three dilution passages to two target-bearing passages raises strict attack success (ASR) from 84% to 92% on NQ (vanilla RAG), 77% to 91% under InstructRAG-ICL, 83% to 87% on HotpotQA, and 72% to 75% on MS MARCO. The same passages transfer across six retrievers, and gains are traced to dilution passages displacing original corpus passages that support the reference answer, shifting responses from mixed endorsement to exclusive adoption of the attacker-chosen answer. The paper also examines passage retention under RAGDefender and reports results under ReliabilityRAG/MIS.

Critical strengths
  1. The problem is well-posed and the experimental design isolates the claimed effect. The paper makes a genuinely interesting observation: a passage can contribute to a targeted poisoning attack without supporting the target answer at all, simply by occupying retrieval slots that would otherwise be filled by evidence for the correct answer.

  2. The construction pipeline is nontrivial and the transfer results are informative. Enforcing "query-relevant but answer-free" is harder than it sounds—the passage must survive retrieval competition while semantically avoiding the answer slot—and the relation-aware planning with 8 roles × 3 facets, per-candidate answer-free checks, joint triple checks, and proxy-score-based group selection is a reasonable instantiation. The cross-retriever transfer (Table 5, gains of 8–15 points across Contriever, MiniLM, BGE, E5, GTE, BM25) and the cross-generator results (Table 4, gains of 3–15 points across four generators) suggest the phenomenon is not an artifact of one retriever or one generator, which strengthens the general claim.

Critical weaknesses
  1. The contribution is narrow and largely incremental relative to PoisonedRAG, and the novelty is not clearly established. The core finding—adding more attacker-controlled passages to the corpus increases attack success—is not surprising. PoisonedRAG already injects five passages per question; the paper's headline comparison is also at five passages (two target-bearing + three dilution). The real question is whether answer-free dilution is qualitatively different from simply injecting more target-bearing passages at the same budget. The paper does not run this comparison: the natural control is five target-bearing passages vs. two target-bearing + three dilution, both at b_inj = 5. Without it, the claim that dilution is a distinct mechanism ("passages can contribute without supporting the answer") is not separated from the trivial "more poison = more success" explanation. The target-only control at b_tgt = 2 is compared against dilution at b_inj = 5, which conflates passage count with passage type. This is the single most important missing experiment, and its absence undermines the paper's central framing.

  2. The "answer-free" constraint is enforced by an LLM judge with no reported validation, and the threat model is arguably unrealistic. The dilution passages are filtered by an LLM evaluator checking that they do not "provide a complete or partial answer, exclude possible answers, or otherwise narrow the answer space." There is no human validation of this judge, no reported agreement rate, no adversarial test of whether passages passing the check actually fail to narrow the answer space, and no analysis of judge errors in either direction. Given that the entire claim rests on these passages being answer-free, this is a significant gap. Relatedly, the threat model assumes an attacker who can (a) generate query-relevant, answer-free passages with a strong LLM, (b) insert them into the corpus, and (c) do so for every one of 100 questions while the corpus is shared. This is a powerful and somewhat artificial adversary; the paper does not discuss feasibility constraints (corpus insertion rights, detection surface, the fact that 300 injected passages across 100 questions may itself be anomalous) or whether a defender could flag dilution passages by their statistical profile (e.g., high query similarity, low answer-bearing content, unusual provenance).

  3. The evaluation is thin and the defense analysis is incomplete. The main results rest on 100 questions per dataset, and most comparisons are reported without confidence intervals or significance tests; the paired gains (e.g., 11 gains and 3 losses on NQ) are reported as raw counts, which is helpful, but no statistical test is applied and the sign test on 11 vs. 3 is marginal. The defense evaluation is limited to InstructRAG-ICL (which does not remove passages) and ReliabilityRAG/MIS, plus a RAGDefender retention analysis in the appendix using a modified single-target setting. There is no evaluation against the defenses most likely to counter this attack: (a) provenance- or reputation-based filtering, (b) consistency/contradiction checking across retrieved passages (which ReliabilityRAG partially does, but the paper shows it fails), (c) query-answer-support classifiers that explicitly score whether a passage supports an answer, and (d) robust aggregation methods like RobustRAG, which is cited but not evaluated. The RAGDefender analysis is particularly weak: it changes the injection budget and target count, so it is not comparable to the main results.

  4. The mechanistic analysis is suggestive but not conclusive, and the paper overstates what it shows. The evidence analysis demonstrates correlation between dilution passage retrieval and ASR gains, but it does not establish causation at the level of the generator's decision. The paper shows that original corpus passages are displaced and that responses shift from mixed to exclusive target endorsement, but it does not probe the generator directly (e.g., by measuring attention to or citation of displaced passages, or by ablating specific displaced passages). The claim that "dilution passages can indirectly increase the chance of the target answer being adopted" is supported, but the stronger framing in the abstract and introduction—that this reveals a general property of "evidence composition"—outruns the evidence. The paper also does not analyze which displaced passages matter: is the effect driven by removal of reference-answer support specifically, or by any reduction in context diversity? The test2361 case is illustrative but n=1.

Questions for authors
  1. The missing budget-matched control: What happens with five target-bearing passages (b_tgt = 5, b_dil = 0) at the same total injection budget as the dilution condition? If ASR is comparable, the paper's central claim reduces to "more poison helps," and the answer-free framing is not doing the work the paper claims. This experiment is essential and should be in the main paper.

  2. Validation of the answer-free judge: What is the human-annotated agreement rate for the LLM evaluator's answer-free decisions? Can the authors provide examples of passages that passed the check but arguably narrow the answer space, and estimate the prevalence of such cases? How sensitive are the ASR gains to judge strictness?

  3. Defense scope: Why are contradiction-based defenses (ReliabilityRAG), robust aggregation (RobustRAG), and answer-support classifiers not evaluated more thoroughly? ReliabilityRAG/MIS shows no gain over the target-only control (93% vs. 93%), which seems to indicate the defense is ineffective here—can the authors explain why, and whether a stronger filtering baseline (e.g., a trained passage-relevance/answer-support classifier) would detect dilution passages?

  4. Realism of the threat model: How many total injected passages does the shared-corpus setup require across 100 questions, and what fraction of the corpus is that? Would a simple anomaly detector on corpus insertion patterns (provenance, burstiness, query-similarity distribution) flag the dilution passages? The paper does not discuss detectability.

  5. Statistical rigor: With 100 questions and paired gains of 11/3 or 5/1, what are the confidence intervals on the ASR differences? A sign test or bootstrap over questions would clarify whether the smaller gains (MS MARCO: 72→75, 4 gains/1 loss) are meaningful.

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

This paper proposes evidence dilution, which strengthens targeted RAG poisoning by adding query-relevant, answer-free documents alongside malicious answer-bearing passages. These documents compete for retrieval slots and alter the evidence available to the generator, increasing strict attack success across three datasets. However, missing budget-matched controls leave unclear whether dilution is more effective than adding further answer-bearing passages.

Critical strengths
  1. Clear and insightful motivation: Identifies how query-relevant, answer-free documents can amplify targeted poisoning by changing the retrieved evidence.
  2. Controlled paired evaluation: Holds target-bearing passages fixed, isolating the incremental effect of adding dilution documents.
Critical weaknesses
  • Missing budget-matched controls: Adding dilution increases the injection budget from two to five passages. Without comparing against five answer-bearing passages, the results do not establish whether dilution is a better use of the same budget.
  • Insufficient component ablations: The paper does not isolate the contributions of relation-aware generation, joint semantic checks, or group selection, leaving the necessity of its construction pipeline unclear.
  • Uncertain answer-free validation: LLM-based checks may miss implicit answer support or information that narrows the answer space. Systematic human verification would strengthen the central claim.
  • Limited statistical evidence and scope: Each dataset contains only 100 test questions, without repeated runs or confidence intervals. The default top-5 setting also leaves effectiveness under larger retrieval contexts uncertain.
  • Uneven gains under defenses: Dilution provides no net improvement under ReliabilityRAG/MIS, so its benefits against defenses appear configuration-dependent.
Questions for authors

see weakness

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

Related papers

Discussion (0)

Sign in to comment.