acceptodds
Under review as a conference paper at ICLR 2027

Suppression Is Not Correction: Causal Tracing of RAG Corpus Poisoning

Abstract

Retrieval-augmented generation (RAG) can be poisoned by malicious corpus passages that steer models toward answers chosen by attackers. Existing studies show that these attacks are effective, but do not explain how poisoned evidence changes the model internally or why removing its influence may fail to recover the correct answer. We present a causal tracing framework that follows retrieved evidence from document positions through the residual stream to the final answer score. Across five LLMs, we identify a six-layer residual evidence band whose state causally shifts the competition between attacker and correct answers. The same functional interface appears across models, while its downstream readout varies by architecture. Some models rely partly on a small set of attention heads, whereas others use more distributed pathways. Source and position controls show that the band carries answer-bearing document information rather than a generic signal that a document is present. On 106 answerable attacks, strong oracle steering suppresses the attacker answer in 92.5% of cases but recovers the correct answer in only 13.2%. Restoring clean residual states within the six-layer band recovers 71.7%, nearly matching the 72.6% recovery from restoring all layers. These results show that suppressing poisoned evidence is not sufficient for correction. Correct recovery requires restoring positive support for the right answer, and the residual band provides a compact causal interface for doing so.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Retrieval-Augmented Generation (RAG), Corpus Poisoning, Causal Tracing, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zFQKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work investigates corpus‑poisoning attacks on RAG systems. By tracing the flow of retrieved evidence through the residual stream of several LLMs, the paper discovers a six‑layer residual evidence band that carries answer‑bearing information and causally shifts the competition between attacker and correct answers. The paper then compares two oracle interventions: steering and position‑aligned restoration (replacing the band with the exact clean residuals). The results show that steering suppresses most attacks but recovers only around 13% of correct answers, whereas restoration restores around 72%, demonstrating that suppression is not correction.

Critical strengths
  1. Across five core models and three additional architectures the authors locate a contiguous six‑layer residual band whose replacement with clean states reduces the attacker-correct margin. This work presents a cross-model causal analysis of RAG poisoning, identifying concrete potential architectural targets for future defenses.
  2. The paper presents a clear empirical separation between “suppression” and “correction”. The experiment results challenge the implicit assumption that eliminating the attacker’s signal is sufficient for effective defense; it demonstrates that restoring legitimate, answer-bearing signals is equally essential.
Critical weaknesses
  1. One of the main concerns on this paper is that the restoration heavily rely on the availability of clean-run activation of the same query, which is largely impractical in real-world deployments. Without a practical, non-oracle restoration mechanism, the paper provides limited actionable guidance for developing real-world defenses.

  2. The experiment scope is confined to short-form, greedy QA generation. Clean and poisoned prompts greedily generate at most eight tokens. The eight‑token analysis is adequate for short, greedy QA but does not demonstrate that the residual‑evidence band remains influential throughout longer generation or multi‑turn dialogues (which is especially important for RAG). Therefore, the paper’s claim of a universally persistent causal pathway is not sufficiently supported by the current evidence and should either be qualified as applying primarily to early-generation stages or substantiated through additional experiments involving longer generation sequences.

  3. The proposed lightweight attribution and detection mechanisms are not universally applicable. They work only for models that exhibit strong head concentration but cannot cover models such as In Llama 2‑7B and DeepSeek‑7B. This undermines the claim of a broadly usable internal signal.

  4. The paper would benefit from a clearer comparison with ShieldRAG. Although these approaches operate at different stages of the RAG pipeline and are not direct baselines for causal tracing, they provide useful reference points for evaluating the practical implications of the proposed findings. In particular, the authors should clarify whether their causal insights offer additional defensive benefits beyond existing retrieval-stage approaches, especially given that their restoration experiments rely on oracle access to clean activations.

Questions for authors
  1. Can you discuss the possibility of a practical, non‑oracle restoration method based on the findings of this paper?
  2. How does the residual evidence band behave during longer generation (e.g., multi‑turn dialogue or document summarization)? Have you measured whether the band’s causal influence persists beyond the first few generated tokens, especially when the model attends to newly generated context?
  3. What is the performance of your single‑head attribution method on models that lack a strong poison‑grounded head (e.g., Llama 2‑7B, DeepSeek‑7B)? Could a multi‑head or full‑layer aggregation recover comparable document localization accuracy for these architectures?
Reviewer_sC5mOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper looks at how RAG poisoning causes failure in LLM based RAG systems. The paper discusses how restoring positive support for the correct answer is needed rather than removing the adversarial influence. The paper performs an analysis to determine where evidence carrying parts of the LLM are and the effects of restoring clean activations in the evidence bands can recover almost all lost performance due to the poisoning.

Critical strengths
  1. The finding of the paper that suppression does not automatically give correction is a nice result. It's one of these findings that could seem obvious in hindsight, but its non-trivial to establish empirically and requires that initial insight.

  2. There is a good technical contribution with the analysis of the "evidence band". Although some of the exact downstream interactions I think remain partially explained, it nonetheless highlights that specific intermediate representations carry answer-relevant evidence that is vulnerable to poisoning.

Critical weaknesses
  1. The paper proposes that there exists a residual evidence block/band. It's position in the various models can vary significantly (0-5 in Vicuna-33B, L7-12 in Llama 3, L14-19 in Qwen). This could be interpreted to support the paper's claim, though it could support a weaker conclusion that different models process evidence through different mechanisms (e.g. different layer blocks) that exhibit similar causal effects.

  2. Perhaps I have misunderstood, but the readout head percentages seem to not support the overall narrative? Table 4 has the mediation of 0.043 - 3.96%, with similar results in S14. With these numbers it seems to weaken the overall story? The work does show good localization of where answer-bearing evidence can be represented, but this seems substantially weaker evidence for how that information is then utilised. With a low percentage mediation it could indicate a more distributed mechanism than the paper narrative suggests.

  3. Many of the analysis depends on poisoned-answer gap (teacher-forced score difference). The motivation for doing so makes sense as it gives a obvious signal, but it is a surrogate objective compared to the actual generated behaviour. It would be interesting to have more analysis between internal effects and downstream generation quality.

  4. Minor: The paper does some interesting analysis, but it would be worth mentioning that it doesn't seem to for now act as something like an actionable defence due to the requirement of needing a clean activations. So the method is more of a diagnostic tool rather than actionable defence. This is a minor point, more around clarity of framing.

Questions for authors

See weaknesses for queries

Reviewer_NzL6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the internal mechanisms through which corpus poisoning attacks influence the answers of retrieval-augmented generation (RAG) systems. The authors identify a six-layer residual evidence band whose activations causally influence the model's preference between the attacker-selected and correct answers. They further analyze downstream attention-head contributions, compare activation steering with clean-state restoration, and investigate whether attention patterns can be used to attribute poisoned outputs to individual retrieved passages.

Critical strengths
  1. The paper attempts to move beyond correlational attention analysis for poisoning against RAG system by performing interventions on internal activations.

  2. The comparison between attack suppression and correct-answer recovery is also an interesting evaluation direction.

  3. The evaluation considers multiple language models and includes held-out validation and several-substitution controls.

Critical weaknesses
  1. My primary concern is that the paper is extremely difficult to follow. The notation, experimental setup, methodology, and conclusions are not presented clearly enough to understand what each experiment establishes.

a. The experiments also use substantially different configurations, including different model subsets, datasets, and evaluation populations. For instance, the paper reports a 3,000-trial attack reproduction, 40 questions for initial layer discovery, 240 attacked units for source sensitivity, 106 answerable attacks for oracle correction, and 143 attacked model-question events for document attribution.

b. It is difficult to understand how these different experimental settings relate to one another, why each configuration was selected, and which conclusions are supported by which populations.

c. I strongly encourage the authors to explicitly state the research questions in experiments, provide a consolidated table of experimental configurations, and explain the motivation and takeaway of each experiment.

 

  1. A central claim of the paper is that it identifies internal pathways through which poisoned passages influence the model's answer. However, it is unclear whether the identified mechanism are specific to adversarially poisoned passages or simply reflect the normal mechanism through which a language through which a language model incorporates retrieved evidence. Is the paper identifying a mechanism of poisoning, or a general mechanism of answer formation from retrieved evidence?

In order to answer this, authors would have to explicitly analyze the differences between the benign retrievals and corrupted retrievals, for any mechanism to find a trace at the level of model internals there should be some difference in the input level as well.

 

  1. The paper considers two variants of a single attack methodology from PoisonedRAG, it can be likely that the findings are specific to that particular attack prompt template and may vary with other attacks. Authors should consider performing an adaptive attack analysis to confirm their findings.

 

  1. The paper does not consider the fraction of poisoned passages in the retrieved context as a primary variable in their evaluations. Previous work has shown that poisoning effectiveness and internal attention behavior can depend on the proportion of adversarial passages [1]. However, the paper does not clearly establish how its mechanistic findings vary with this parameter.

Authors should consider evaluating questions like: Do the same residual bands and attention heads remain important when one, two, or several retrieved passages are poisoned? Does the causal effect increase with the number of poisoned passages?

[1] Through the Stealth lens: Attention-Aware Defenses Against Poisoning in RAG (ICML 2026).

 

  1. There is existing research examining how corpus poisoning influences internal attention patterns and using these signals to identify and mitigate poisoned passage. The paper should discuss them and their own positioning relative to such existing efforts.

For example, Through the Stealth lens: Attention-Aware Defenses Against Poisoning in RAG (ICML 2026) studies how poisoned passages influence attention distributions and proposes attention-based detection and filtering mechanism. This is directly comparable to the document-level attribution experiment in the manuscript. Other work, such as When Context Bites: Detecting RAG Poisoning via Document-Level Attention Collapse, also analyzes differences between clean and poisoned attention behavior.

 

  1. More broadly, the paper should clarify whether its intended contribution is a practical defense, a mechanistic understanding of poisoning, or investigation of oracle intervention capabilities. Currently, these objectives appear mixed, making the practical takeaway difficult to identify.
Questions for authors

Please refer to weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.