acceptodds
Under review as a conference paper at ICLR 2027

Reading Too Much into Context: Passive Exposure Can Steer LLM Decisions

Abstract

Large language model (LLM) assistants can now search the web and consult external sources while completing user requests. These sources can provide useful evidence, but they can also introduce additional content into the model’s context. Can such passive exposure steer a decision even when the added content provides no reason to change it? We examine the stability of model decisions on the same tasks with and without such external content. Across all open-weight and closed-weight models we test, exposure systematically shifts decisions, with effects reaching nearly 50 percentage points in closed-weight models. The same pattern appears with real-world online opinions. The influence also extends beyond subjective preferences. Such exposure can steer models toward choices that violate explicit user requirements and increase their acceptance of false claims. In short, what enters an LLM’s context can influence its decision even when it should not determine it.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language models, passive exposure, decision-making, contextual robustness, instruction following, misinformation, AI safety

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_rXZPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper is an empirical behavioral study that examines current LLMs from a very small but very interesting perspective: task-relevant content is not necessarily decision-relevant evidence. The authors design four experiments covering everyday decisions, real-world Reddit opinions, product selection, and misinformation judgments. By keeping the task information and candidate options unchanged, they examine the influence of additional text on model behavior. The results demonstrate the insight that external text can not only change model preferences but also cause models to violate explicit user requirements or become more likely to accept false information.

Critical strengths
  1. Research Significance

This paper approaches the problem from a relatively narrow but interesting perspective, focusing specifically on how external information influences LLM decision-making. The topic is closely related to real-world LLM applications and has certain practical significance.

  1. Detailed Experimental Design and Clear Presentation of Results

The paper presents a relatively comprehensive and carefully designed set of experiments, considering different models, prompt formats, and answer orders. It examines both subjective choices without a uniquely correct answer and objective judgments with explicit correctness criteria. The four experiments also have a certain logical connection. In addition, the experimental visualizations are clear and aesthetically pleasing, using various forms such as scatter plots, bar charts, and box plots, with vivid colors that help readers intuitively understand and compare the results.

Critical weaknesses
  1. Limited Technical Contribution and Unclear Distinction from Prior Work

This paper does not propose a new model architecture, optimization objective, or algorithm. Its core approach consists of controlling input content and prompt formats, and even the mitigation strategies are primarily prompt-level interventions rather than a newly implemented and validated system. This is an empirical paper centered on input interventions and behavioral measurements, with almost no algorithmic contribution.

Furthermore, prior studies such as Shi et al. (2023) and Yoran et al. (2024) have already demonstrated from different perspectives that irrelevant context can interfere with LLM outputs and cause errors even when the correct answer remains unchanged. The Related Work section mainly distinguishes this paper from previous studies based on task categories, such as QA vs. decision-making and prompt injection vs. passive exposure, but does not clearly explain what substantially different insights this work provides beyond extending existing findings to another task setting.

  1. The Rigor of the Experimental Design Requires Further Clarification

The paper's central research question is: "Can such passive exposure steer a decision even when the added content provides no reason to change it?" However, the justification for considering the additional content decision-irrelevant is insufficient. In Sections 3.1 and 3.2, although the added content is not directly related to the specific task, this does not necessarily mean that it should not influence the model's decision. These are two different concepts.

For example, in Section 3.2, Reddit users' experiences with buying and renting homes actually provide decision-relevant considerations, such as maintenance responsibilities and housing autonomy. Therefore, it may be reasonable for models to consider these opinions and change their choices. The authors should explain why such content is considered decision-irrelevant.

Questions for authors
  1. Given that prior studies have already shown how irrelevant context can influence LLM outputs and decisions, could the authors clarify what new insights this work provides beyond extending existing findings to different decision-making scenarios?

  2. In Sections 3.1 and 3.2, the additional content may provide useful information for decision-making. Could the authors provide further evidence or clarification to support the claim that "the added content provides no reason to change it"?

  3. Minor presentation issues: Figure 1 is relatively simple, with text and lines that are too small and thin, making it difficult to read. Additionally, the abstract does not clearly explain the research gap or distinguish this work from existing studies.

Reviewer_25P6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether the unrelated content (the author called "passive exposure") read by LLM will change the model's decision, even when that content gives no reason to change it. The authors compare the same task with and without an extra passage across four tasks, including cautious decisions, Reddit preferences, product selection with a single correct answer, and wind-energy misinformation. Exposure shifts decisions in all eight tested models (up to about 50pp), more than doubles the chance of picking a product that fails the requirements, and raises acceptance of false claims more than true ones.

Critical strengths
  1. The experiment setup is comprehensive, covering different situations, including tasks without a correct answer and tasks with one. It also covers eight models to make sure the generalizability. Because the effect appears consistently, I believe the conclusion generalizes.
  2. The problem is practical, and the conclusion of Section 4.1 is well made. Users should know where the information a model reads comes from, and make sure those sources are the ones they want.
Critical weaknesses
  1. The research question is practical, but the analysis of the results is thin. Several results are reported without explanation, such as the asymmetric shifts in Section 3.1 and the Mistral reversal in Section 4.1. Some results are also hard to read. For example, I could not get a clear insight from Table 2, and more visualization would help support the findings.
  2. Section 4.3 proposes separating information gathering from the final decision, but none of it is tested. Without any test, this section is only a suggestion on paper, and the paper ends without a solid conclusion.
  3. Overall, I appreciate the problem formulation of this paper, but I expect the experiment and the analysis can be deeper and more elegant. Some that we can explore for more impressive insight rather than the observation on the surface.
Questions for authors
  1. As weakness 1, does the effect size track passage length or persuasiveness? Are the two opposing Reddit opinions in each situation comparable in strength?
  2. Section 3.1 says, "In our analysis, the model notes that the exposure offers limited evidence, yet still uses it." What does this analysis refer to? Is it based only on reading the CoT traces in Appendix L.1, or is there a quantitative measure?
  3. Is there an explanation for why Mistral reverses direction when the exposure is given as a tool response?
  4. It's interesting that all models favor the inferior product more when it is reviewed, under both praise and criticism. Is it related to how much information the model has about that product?
  5. In the misinformation task, does exposure raise acceptance of false claims because the model gives more weight to content that appears to have a source?
3 comments
Reviewer_25P6OpenReview ↗

Reply to author response

Thanks for the detailed response and the new experiments. The answers to Q1, Q4, and Q5 are clear and helpful. Please include these results in the revised version.

I still have some concerns about Q2 and Q3:

  1. For Q2, the 52.1% comes from manually checking 48 traces from a single model (Qwen3-32B). Could you clarify how the annotation was done? For example, who annotated the traces, what criteria were used to decide that the model "noted limited evidence but still used it", and whether there was more than one annotator (with agreement reported)? It would also help to know whether this pattern holds in other reasoning models, not just Qwen3-32B.
  2. For Q3, the empty tool response ablation is a nice finding, but the explanation that Mistral treats a completed tool call as a signal that verification is done is still a guess. Is there any further evidence for this? For example, does the same drop appear when the tool returns something like "no results found", or across all tasks rather than on average? Some analysis of Mistral's responses after the tool call would also make this explanation more convincing.

I'd appreciate it if the authors could add this information. For now, I will keep my score.

Reviewer_25P6OpenReview ↗

Reply to further author response

Thanks for the follow-up and for adding the Q1, Q4, and Q5 results to the appendices. The additional response for Q3 also resolves my concern.

Q2, which is the point I care about most, remains unresolved. A single annotator is not convincing, since we cannot know how well human judgments agree. Multiple human annotators with reported agreement and a per-model rate are needed; otherwise, this should be presented as qualitative examples.

My initial assessment was between 1 and 2, and I chose 2. After the rebuttal, I see the paper as borderline between 2 and 3. Given the remaining Q2 concern and the other reviewer–author discussions, I will keep my score at 2.

Reviewer_25P6OpenReview ↗

Final reply to author

Thanks for the follow-up.

I agree with the proposed direction. However, these analyses will only be done later. I think revising and resubmitting would be better. I will keep my score.

Reviewer_Tsi7OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper evaluates how instruction-free passages affect binary choices in caution, preference, product-selection, and factual-judgment tasks. Its strongest finding is that mentioning an inferior product can increase its selection probability regardless of review sentiment. My concern is the scientific increment: the current study does not clearly separate a new, generalizable finding from established irrelevant-context and semantic-leakage effects, and the limited correctness-task coverage does not support the breadth of the framing.

Critical strengths
  1. A verifiable correctness criterion. Product tasks require the uniquely cheapest eligible item, and manual checks establish that review snippets change neither eligibility nor price. This makes these effects more interpretable than underdetermined preference changes (Sec. 2.2, p. 3; App. G, pp. 21–23).

  2. An informative target-by-sentiment comparison. Inferior-product praise and criticism both increase its probability, by 19.24 and 17.10 percentage points. The full crossing demonstrates that sentiment direction alone does not explain the observed preference shift (Table 2, p. 6; Table 8, p. 23).

  3. Careful accounting for baseline movement. The reminder analysis retains matched NoContext conditions and explicitly decomposes changes in exposed outcomes and baseline behavior. The separate neutral-passage comparisons also acknowledge when introducing a topic changes preferences (Eq. 11, p. 26; Tables 5–6; Sec. 4.2).

Critical weaknesses
  1. The distinctive contribution remains difficult to separate from option salience. Section 6 already discusses irrelevant-context errors, semantic leakage, and recommendation manipulation, including Shi et al. (2023) and Gonen et al. (2025). The product experiment adds an explicit correctness criterion, but praise and criticism both repeatedly name the target: Appendix N.2 requires one exact-name mention per snippet, and Appendix G combines three snippets per bundle. The target-by-sentiment crossing establishes that sentiment direction is insufficient to explain the shift; it does not establish whether the review content contributes beyond repeated option mentions. A neutral or name-only exposure matched for repetition, length, and position would make the incremental finding much sharper (Table 2; App. G, p. 21; App. N.2, p. 34).

  2. The evidence for consequential errors is narrower than the overall experimental breadth suggests. The objectively scored tasks comprise 24 product selections and nine false wind-energy claims with true counterparts. The three closed models are evaluated only on the caution task, whose responses have no unique correctness criterion. Additional formats and exposure passages therefore do not establish comparable objective errors across models or decision domains. Generalization also depends on how passages were selected and assigned: examples and generator prompts are provided, but the complete task-to-passage mapping is deferred with the code until acceptance. A broader criterion-fixed collection and an inspectable stimulus manifest would support the intended generality more directly (Secs. 2.2–2.3; Table 1; Apps. F–H; Reproducibility Statement, p. 10).

  3. The measured shifts and mitigation results are not connected to the same correctness outcome. Product and misinformation results use likelihoods renormalized over two answer strings, while the main caution experiment samples choices. Equation 2 is a meaningful preference measure, but a product probability rising from 9.09% to 19.91% does not determine how often the selected answer becomes incorrect. Meanwhile, critical instructions and objectivity reminders are evaluated on caution and subjective preferences rather than the criterion-fixed tasks. Section 4.2 correctly distinguishes attenuation from correction, but the experiments stop short of determining whether the intervention reduces actual errors. No direct evidence found in the manuscript. Reporting decision transitions and exposed/unexposed correctness under a stated decision policy would connect the claimed problem and remedy (App. C.2; App. G.2; Apps. I–J).

  4. The false/true comparison changes several factors relevant to the interpretation. The false oil-change claim specifies a 500-hour interval; its true counterpart only says that gearboxes require oil. The exposure about repeated gearbox-fluid replacement is also semantically closer to the former proposition. Baseline acceptance differs substantially, at 13.74% versus 88.61%. Proposition specificity, semantic correspondence, and response headroom thus vary together. Appendix H.2 acknowledges this, but the comparison consequently supports a result about these particular paired statements, rather than identifying a general difference in susceptibility to false and true claims (Tables 9–11, pp. 23–25).

  5. The tool experiment does not isolate message role. User-document exposure precedes the task, whereas tool exposure follows the task and a preset assistant tool call. Role, passage position, and interaction structure change together. Appendix K acknowledges these differences, while Section 4.1 interprets the result in terms of role boundaries. Sensitivity in both constructed sequences is supported; attributing their comparison to the role boundary requires a matched sequence or a more limited interpretation (Sec. 4.1; Table 17; App. K, pp. 29–31).

Questions for authors
  1. Does the product-target effect exceed that of neutral or name-only mentions matched for repetition and position? If such a control is available, what aspect of passive exposure does it identify beyond option activation, and how does that distinguish the contribution from the contextual-interference work discussed in Section 6 (Table 2; App. N.2)?

  2. For the criterion-fixed tasks, how often does exposure change a correct decision into an incorrect one under the intended decision policy, and do the reminders reduce those errors without harming unexposed performance? This would clarify the practical meaning of the likelihood shifts and mitigation results (Eq. 2; Sec. 4.2; Apps. G and I–J).

  3. How were the nine false/true claim pairs and their exposure passages matched for specificity and semantic relevance? Is there evidence that the asymmetric response survives a closer match, or should this conclusion be confined to the current pairs (Tables 9–11; App. H.2)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.