acceptodds
Under review as a conference paper at ICLR 2027

PoisonVax: Protecting Trusted Facts in RAG Through Dual-View Records

Abstract

Corpus poisoning can steer retrieval-augmented generation (RAG) toward attacker-chosen answers by displacing trusted evidence or overriding it during generation. Existing defenses mainly filter suspicious passages or aggregate responses from retrieved evidence, but they cannot ensure that trusted facts reach the generator and prevail over competing claims. We introduce PoisonVax, a maintainer-led defense that reinforces trusted facts during corpus construction to improve both retrieval exposure and answer adoption. Each fact is stored as a dual-view record: an index view with an optimized retrieval key serves the retriever, while a generation view with a directive supports the trusted answer. The retriever scores only the index view, while a fixed field mapping supplies the generation view after retrieval. This separation keeps generation directives out of retrieval optimization while leaving the retriever, generator, and scoring rule unchanged. Records are constructed without access to evaluation questions. Under five injected attack passages per fact, with the same trusted facts available to all methods, PoisonVax achieves 58.46% answer accuracy and substantially outperforms TrustRAG and RobustRAG while using one online generation call. Matched experiments further show a 30.09 percentage point gain over a fully reoptimized single text baseline after removing attractor spans.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
retrieval-augmented generation, corpus poisoning, RAG security

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JB8bOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose PoisonVax, a maintainer-side defense against corpus poisoning in retrieval-augmented generation. The key idea is to represent each trusted fact using two text views: an index view optimized to improve retrieval exposure and a generation view containing the trusted fact and an answer directive to improve answer adoption. The two views are linked through a fixed field mapping, while the retriever and generator remain frozen. The paper further argues that retrieval exposure and answer adoption are distinct failure modes and should be handled separately. Experiments compare PoisonVax with several existing RAG defenses and include matched single-text controls, unseen questions, two corpora, different retrievers and generators, adaptive attacks, corpus-scale experiments, and analyses of non-target interference. The results generally show strong protection under the evaluated poisoning settings, although the practical protection–utility trade-off remains an important concern.

Critical strengths
  1. Clear and useful decomposition of the defense problem. The separation between retrieval exposure and answer adoption is conceptually meaningful. The paper does more than report end-to-end accuracy: the controlled interventions on retrieval scores, exposure, ordering, and attacker removal provide evidence that successfully suppressing one attacker does not necessarily restore the trusted answer.
  2. The empirical evaluation is unusually extensive. The paper includes matched single-text controls with multiple optimization and generation seeds, experiments on unseen questions and different corpora, retriever and generator transfer, defense-aware attacks, large-corpus evaluation, and explicit analysis of non-target utility. The matched representation experiment is particularly important because it shows that the dual-view advantage is not explained solely by the attention-attractor spans.
Critical weaknesses
  1. The main comparison against existing defenses does not fully isolate the benefit of the proposed defense mechanism from the additional maintainer capabilities. In Table 1, existing defenses receive ordinary trusted records, whereas PoisonVax receives per-fact optimized retrieval keys, a separate generation view, and a maintainer-controlled field mapping. This makes the comparison practically relevant, but it is not a strictly capability-matched comparison. The single-text experiments control the representation question well, but they do not replace a comparison against simpler maintainer-side alternatives that exploit comparable privileges. For example, it would be useful to compare against simpler schemes such as explicitly boosting trusted records, reserving a trusted retrieval slot, or using the same two-field infrastructure with an unoptimized trusted fact. Without such controls, it remains unclear how much of the gain comes specifically from PoisonVax rather than from giving trusted records stronger infrastructure-level privileges.
  2. The protection–utility trade-off is not yet resolved by a single deployable configuration. The default non-selective records interfere substantially with unrelated queries, while the unconditional generation directive can introduce unrelated protected answers. The paper proposes selective retrieval keys and conditional generation directives to reduce these effects, but these two mechanisms are evaluated separately rather than as one complete end-to-end system. In addition, the million-passage experiment uses the non-selective version, where protected records are retrieved for a large fraction of unrelated queries. This is important because the practical contribution depends not only on protecting target facts but also on avoiding degradation of normal RAG behavior. The current experiments demonstrate that each source of interference can be reduced independently, but they do not yet establish the overall operating point of the complete mitigated system. A related concern is generator dependence. The complete PoisonVax record improves results for several generators but performs worse than both a plain trusted fact and no defense with Llama-2-13B, indicating that the generation-side mechanism is not universally reliable.
Questions for authors
  1. Can the authors add capability-matched maintainer baselines? In particular, how does PoisonVax compare with simpler approaches that use the same maintainer authority, such as trusted-record score boosting, reserved trusted slots, or the same index-to-generation field mapping without an optimized retrieval key?
  2. Can the authors evaluate one complete deployable variant that combines the selective retrieval key and the conditional generation directive? Please report target protection and non-target retrieval/generation utility simultaneously, preferably also at the million-passage scale.
  3. How should the Llama-2 failure be interpreted? Can the authors provide a principled criterion for selecting or adapting the generation view to a new generator without tuning on the protected evaluation questions?
  4. All semantic answer labels appear to be machine-generated. Can the authors provide human validation on a representative subset, particularly for ambiguous outputs containing both correct and attacker-related information? The paper explicitly states that the answer labels and rechecks are machine-produced, so agreement with human judgment would strengthen confidence in the headline accuracy comparisons.
Reviewer_z1zkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PoisonVax constructs dual-view records to protect trusted facts from corpus poisoning. One view improves retrieval, while the other encourages the generator to use the trusted answer.

Critical strengths
  1. The paper distinguishes retrieval exposure from answer adoption. The controlled experiments explain why improving one stage may leave the system vulnerable.
  2. The reoptimized single-text baseline provides a meaningful test of the representation design. The advantage remains after removing the attractor spans.
  3. The paper examines interference with unrelated questions. This reveals an important cost of deploying the defense in a shared index.
Critical weaknesses
  1. The maintainer knows which records are trusted and controls context assembly. Why must these records compete through optimized keys in a shared index? Retrieving from a separate trusted index with one reserved context slot seems like a necessary baseline.
  2. The single-text control optimizes retrieval while keeping the generation directive fixed. Joint-GCG [1], jointly optimizes retrieval and generation, which provides a stronger test of the need for separate views. Could the authors compare against a correct-answer adaptation using the same construction questions?
  3. Table 11 shows a substantial loss of target protection when selectivity is enabled at B=5. The conditional directive is evaluated separately, so the practical operating point remains unclear. Can the combined version preserve useful protection on a workload dominated by unrelated questions?
  4. The two-view interface is clear for a single dense retriever. Which view should a cross-encoder reranker receive? Please evaluate reranking each view separately to determine whether the protection survives a standard two-stage retrieval pipeline.

[1] Joint-GCG: Unified Gradient-Based Poisoning Attacks on Retrieval-Augmented Generation Systems.

Questions for authors

See Weaknesses.

Reviewer_1UAPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method named PoisonVax to handle corpus poisoning using a dual-view approach where one view is at retrieval focus and the other is on the answer generation. The main idea/motivation is to make the retriever easy to retrieve and harder to omit for the answer generation the correct ones. It creates special records before deployment rather than trying to detect malicious ones at query time. Effectively, as the paper mentions, there can happen two things: (i) retrieve malicious documents at the top of retrieval cap or (ii) system follows malicious claims from the injected poisoned ones. Authors argue that it is difficult to handle both using existing systems (i.e., filtering after retrieval). Hence, they propose this dual-view approach. With the dual view, retrieval interacts with index-view and retrieves the facts/documents. Then generation-view gets the mapped generation text containing the fact and answer directive from the index-view to send to the answer generation.

The proposed approach is simple and clever (to use a dual-view mechanism to counter corpus poisoning). The evaluation shows positive results. On the other hand, there is doubt how this approach could work for a more complex RAG pipeline where we cannot have a simple predefined answer for each fact/document. I think, even when we have two probable answer candidates for a fact (because a question can be formed to ask about something not stored in generation-view), it becomes a problem on what to do. Specifically, retrieving a relevant fact does not necessarily mean that the stored short answer always matches what the user is asking. I also do not see authors have evaluated this setting in this paper. This is a real practical issue that goes beyond a constrained fact-protection setting/problem. I am doubtful protection – utility tradeoff is broadly useful.

Critical strengths

• Relatively easy to read and understand. However, there is room to improve for readability. For example, give examples for the two views for a fact/document.

Critical weaknesses

• Answer generation view is pre-determined. If that view is created in a way that some keys are not instructed, then it exposes a clear gap/weakness in the system. That is, how can you enumerate all the answers for fact/passage/sentence more reliably so that when any query for the fact comes in retrieval time, the view can instruct the correct answer.

• On the other hand, this approach seems to work on singleton fact retrievals. For multi-hop or complex questions, it is not clear how this mechanism would work because we may need more facts connected when retrieving and answer keys may not be a simple as taking all.

Questions for authors

• Table 1 presents results using a small question set. RAG achieves 96.92% accuracy, which suggests the task is not that hard. Even though this evaluation is about recovering under poisoning, it is not exactly clear how would this approach perform on natural-looking questions that are more hard that may also require different answer than the already stored ones.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.