acceptodds
Under review as a conference paper at ICLR 2027

RIVET: Evaluating Retrieval-Induced Value Steering and Retrieval-Stage Mitigation in Search-Augmented LLMs

Abstract

Agentic search pipelines decide, unseen by users, which query to issue and which retrieved snippets to deliver. Whoever controls these choices could shift a model's expressed values without altering any document, a threat that corpus-poisoning defenses and standard RAG metrics are not designed to catch. We propose RIVET, a controlled protocol that evaluates this vulnerability by varying only query formulation or evidence selection over real search snippets. Across five LLMs and two value benchmarks, RIVET shows that the risk is real: on items whose retrieved pool holds both sides, selecting which authentic snippets reach the model moves its judgments by up to 23% of the response span, a single snippet suffices, and adding a single word (support or oppose) to the search query also shifts them, most clearly on ValueActionLens. This is not blanket context-following, since settled factual claims move least. Because the steering enters through a context that misrepresents what the model's own query found, we propose M-RIVET, a retrieval-stage defense that repairs a context only when it is one-sided relative to what its own neutral query retrieved. Across three models and both benchmarks, it removes 53–90% of the steering from oppose-oriented selection, remains effective with independent labels, and leaves unflagged contexts unchanged. Retrieval-induced value steering is thus both a real risk and one that can be repaired at the retrieval stage.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
retrieval-augmented generation, agentic search, value alignment, LLM safety, adversarial retrieval, mitigation, evaluation protocol

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

This paper presents RIVET, a framework for evaluating retrieval-induced value steering in search-augmented LLMs. It examines two attack settings, query rewriting and evidence selection, which manipulate retrieved evidence without modifying its content. Experiments on ValueActionLens and ValueBench using five LLMs demonstrate that such manipulation can influence value judgments. The authors also propose M-RIVET, a retrieval-stage defense that detects stance imbalances and repairs retrieved evidence to mitigate these effects.

Critical strengths
  • Relevant research problem. The paper investigates how the selection of authentic retrieved evidence can influence LLMs' value judgments. This is a different security concern from traditional document poisoning.

  • Mitigation mechanism. The paper proposes a retrieval-stage defense that detects stance imbalances and selectively repairs evidence. The results show it can reduce value steering under the evaluated attack settings.

Critical weaknesses
  • Limited practicality of the threat model.

    (1) The M2 attack adds directional terms such as "support" or "oppose" to the original search query, changing its semantic meaning and the information being requested. Consequently, retrieving different evidence is an expected behavior of the search engine, instead of being affected by adversarial manipulation.

    (2) M3 assumes that attackers can directly select which evidence is presented to the LLM. While these attack settings are theoretically feasible, the paper does not demonstrate how attackers could obtain such control or execute these attacks in real-world search-augmented LLM systems.

  • Lack of evaluation on real-world search-augmented systems. The experiments rely on a simplified pipeline constructed by authors that retrieves Google Search snippets through the Serper API and feeds them to LLMs. However, the paper does not evaluate real-world end-to-end search-augmented systems, such as Perplexity AI or GPT-4o-search-preview. It does not examine how its experimental setup differs from these systems yet.

  • Insufficient attack and defense comparisons. The attack evaluation is limited to the authors' proposed M2 and M3 settings, without comparisons with existing query or retrieval manipulation attacks. On the defense side, M-RIVET is mainly compared with MMR and xQuAD, while simpler alternatives, such as restoring neutral retrieval results, are not adequately evaluated.

  • Unclear presentation. The paper's presentation makes it difficult to follow:

    (1) The excessive use of bullet points and bolded opening phrases makes the writing fragmented and disrupts the logical flow.

    (2) The distinction between evaluation baselines (M0/M1), attack settings (M2/M3), and additional experimental controls is not clearly organized.

    (3) The numerous labeling procedures and their roles in the overall framework are difficult to follow.

    (4) The rationale behind several thresholds and repair rules is insufficiently explained.

    (5) The framework figures contain excessive text, complicated flowcharts, and substantial unused space, making the core methodology difficult to understand.

  • Ambiguous interpretation of attack success. The paper measures value steering primarily through deviations from responses under neutral retrieval. However, in M2, adding directional terms such as "support" or "oppose" changes the search query's meaning, and the resulting responses are naturally expected to reflect the modified retrieval intent. Therefore, deviations from the neutral-query response do not necessarily indicate a successful attack. Similarly, restoring the original response does not necessarily demonstrate improved correctness or safety.

Questions for authors

The main questions have been outlined in the Weaknesses section.

Reviewer_SshPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how a search augmented LLM’s expressed value judgments can be steered by changing search queries or selecting particular search results, without modifying the underlying documents. That is, we can simply change the way we present evidence to alter the judgement. The paper presents RIVET (an evaluation protocol) that evaluates and measures this effect. Paper discusses mainly two different attacks: query modification-based retrievals and selection modifications.

Authors introduce M-RIVET where retrieval level defense to detect and then repair. One of the interesting findings of this paper is that retrieved evidence selection can alter model outputs, and the proposed defense method can reduce this behavior. M-RIVET approach is simple. However, there is doubt that when M-RIVET checks delivered context against its own retrieval, that decision may be based on incorrect search results. Therefore, effectiveness of M-RIVET is unclear in those cases.

Critical strengths

• Paper is based on simple findings (selections can steer LLM’s judgement without any modification to documents).

Critical weaknesses

• For the repair procedure, it is not clear it is actually useful. That is, defender gets the reference pool. A comparison of using that pool’s top 5 results may show what more is added through diversification step.

• The defense depends on snippet labels. For example on ValueActionLens the oppose selection contexts flagged dropped from 98.7% to 66.4%. That means, effectiveness of defense relies on labels.

• Even though the approach is simple and interesting, presentation of the paper makes it hard for a reasonably well-prepared reader.

Questions for authors

Questions

• We have neutral query results. Why and when is repairing delivered context preferable to suing top-5 neutral results? Can you explain?

• How is reference query obtained and protected when attack can affect query formulation?

• Can attacker include few neutral / opposing items and still steer the final answer? Is this a possible scenario?

Suggestions, comments

• Introduces terms RQ1, RQ2,.. in the text but not sure they were introduced before the mention. For the reader, it is not clear what these mean.

Reviewer_YY9LOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether changing search queries or selecting authentic retrieved snippets can steer an LLM’s expressed value judgments. RIVET evaluates these interventions across five models and two value benchmarks. The paper also proposes M-RIVET, which uses a trusted neutral-query pool to detect and repair skewed evidence composition. The results show directional answer shifts and substantial recovery toward ordinary-retrieval answers after mitigation.

Critical strengths
  • The attack setting is practically relevant. Manipulation happens through query formulation or evidence selection without changing the underlying documents.
  • The paired evaluation and controls for snippet count, relevance, and labeling differences provide useful evidence that the steering effects are real.
  • Studying both value endorsement and action judgments gives a broader picture than measuring answer polarity alone.
Critical weaknesses
  • The motivation does not clearly establish what is new beyond existing opinion-manipulation research. Prior work already studies how biased retrieved evidence steers model responses, including factually correct evidence. RIVET introduces a different attack setting and evaluation protocol, but the additional scientific insight remains unclear.
  • The benefit of the repair mechanism over simple fallback is not established. The defender already has a trusted reference pool, and recovery is measured against answers from that pool. Missing fallback comparisons and evidence of preserved utility leave the rationale for minimal repair unconvincing.
  • The factual control does not firmly support the stronger interpretation of the results. Settled facts often receive no opposing evidence. Their apparent resistance becomes much less consistent when opposing evidence is actually available.
Questions for authors
  1. My main concern is what we learn beyond the known observation that changing retrieved perspectives changes the answer. “FlippedRAG: Black-Box Opinion Manipulation Adversarial Attacks to Retrieval-Augmented Generation Models” already studies opinion steering through retrieval manipulation. More directly, “Epistemic Bias Injection: Manipulating LLM Opinion via Selective Context Retrieval” studies steering with factually correct but one-sided passages and proposes BiasDef, yet is not discussed here. Your restriction to existing snippets is meaningful, but what new finding follows from it beyond demonstrating the effect on value benchmarks? I’m not yet convinced that the paper establishes more than a new setting for a known phenomenon. The relevance of BiasDef also needs addressing, through an adapted comparison or an explanation of why its assumptions do not fit.

  2. Since the defender already keeps the neutral-query pool, why not directly return its top five results? That reconstructs the reference context used to measure recovery. If minimal repair is meant to preserve useful evidence, where is that benefit shown? I miss a comparison using the same gate followed by simple fallback or an existing reranker. Answer stability alone does not establish why the proposed repair is preferable, especially when it also changes benign neutral-selection answers.

  3. The claim that this is not blanket context following seems stronger than the control supports. In the oppose-selection condition, 106 of 150 determinate claims receive no opposing snippet. Once Table 11 restricts to cases with opposing evidence, the flip-rate differences are no longer significant. For Llama, determinate facts actually show a larger mean absolute shift than values, 0.75 versus 0.29. The appendix acknowledges this. I find the steering effect convincing, but not the stronger interpretation based on this comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.