acceptodds
Under review as a conference paper at ICLR 2027

Promotion Under Audit: Subliminal Manipulation of Interactive LLM Advertising

11 views

Abstract

With interactive large language model (LLM) assistants increasingly mediating consumer decisions, commercial incentives arise for model providers to steer recommendations toward sponsored options. However, as LLM-generated content faces increasing scrutiny, advertising success must be assessed not only by its capability to steer recommendations toward sponsored options, but also by whether such steering remains effective under audit. Existing manipulation methods always leave detectable evidence through biased instructions, anomalous triggers, or formulaic promotional wording. We propose Subliminal-On Policy Distillation (OPD) to eliminate observable evidence, which transfers a commercial preference into model weights using OPD on semantically unrelated training data. The resulting model requires no biased prompt, Retrieval-Augmented Generation (RAG) injection, or trigger at deployment. Jointly accounting for promotion success and full-context detectability, Subliminal-OPD achieves the best performance on controlled clinic recommendation tasks. Complementary judge-independent analysis of successful responses in sentence-embedding space shows that Subliminal-OPD induces the nearly smallest distributional shift from the unmanipulated model, providing further evidence of audit resistance.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Interactive LLM advertising, Subliminal learning, On-Policy Distillation

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

The paper studies how advertisers can manipulate interactive LLM recommendation assistants when a user-side auditor can inspect the full runtime context including system prompt, retrieved passages, and final answer. The proposed method, Subliminal-OPD, transfers the target preference into model weights through on-policy distillation from a teacher that is the same base model under a promotional system prompt, using only semantically unrelated numeric queries and student-generated rollouts.

Critical strengths
  • The paper defines a decision-relevant joint objective for covert advertising: Tied\mathrm{Tied} success weighted by the probability of escaping a full-context audit. Detection is explicitly conditioned on promotion success, and aggregation over RAG variants is made explicit. This is a cleaner and more honest way to score stealth than reporting promotion success or detection rates separately.

  • The evaluation is broad for this topic: three model scales, nine comparison methods (ten rows including OPD), RAG and no-RAG settings, five context variants, three disclosed LLM auditor prompts plus a deterministic context rule, and a judge-independent four-encoder distribution-shift analysis with a permutation-based finite-sample floor.

Critical weaknesses
  • The choice of training prompts is insufficiently justified. In Figure 1, the authors combine recommendation scripts with semantically unrelated instructions, but do not explain why preferences learned from these prompts should transfer to other settings. It is also unclear whether using semantically related instructions would produce different results. A comparison would help clarify the role of the unrelated instructions.

  • The intended recommendation behavior and training target need clarification. Is the goal to make the model recommend specific products, or simply to increase the frequency of recommendations when users ask questions? Clarifying this distinction would make the method’s objective and evaluation easier to interpret.

  • From my perspective, the main contribution of this paper is applying OPD to LLM recommendation. The core difference between OPD and Subliminal-OPD is the prompt input into the teacher.

Questions for authors
  • In the no-RAG setting, Appendix C.4 defines Tied-Full as the product of Tied and Full for each variant, calculated over the matching success scope. However, several no-RAG entries do not match the value obtained from the displayed columns. For example, for Llama Natural Sys, 76.2×(1−0.996)=0.376.2 \times (1 - 0.996) = 0.3, whereas the reported value is 1.1.

  • Could you report detection by at least one additional judge family (or a small human audit) for the Subliminal-OPD vs. Subliminal-SFT and vs. Plain comparisons, to support the stealthiness claims beyond the single frozen judge?

Reviewer_Who3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the problem of covert commercial preference manipulation in interactive LLMs used for commercial recommendations. The authors argue that although existing manipulation methods can increase the likelihood of recommending a target provider, they often leave detectable traces in the runtime context or generated responses. To address this issue, the authors propose Subliminal-OPD. This method distills the token-level probability distributions of a teacher model with a specific commercial preference into a student model using numerical sequences that are semantically unrelated to the recommendation task. As a result, the student model can acquire and exhibit the desired recommendation preference without being exposed to explicit promotional instructions. Furthermore, the authors introduce the Tied-Full metric, which jointly considers the promotion success rate and detectability under full-context auditing to evaluate the effectiveness of manipulation methods under audit.

Critical strengths
  1. The proposed Subliminal-OPD method combines subliminal learning with on-policy distillation. By leveraging student-generated trajectories and token-level probability distributions provided by the teacher model, it enables the transfer of covert behavioral preferences through semantically unrelated data. The method is technically well-motivated and provides a novel perspective on studying covert preference manipulation in LLMs.

  2. The authors compare several representative manipulation methods, including system-prompt steering, RAG injection, SFT, trigger-based backdoors, and VPI backdoors, and evaluate their performance in both RAG and non-RAG settings. In addition, the proposed Tied-Full metric combines promotion success with the conditional detection rate among successful recommendations, allowing the evaluation to consider not only whether manipulation succeeds but also whether it remains effective under audit.

Critical weaknesses
  1. The authors do not sufficiently justify the necessity of auditing commercial recommendations or clarify the practical scenarios in which such auditing would take place. In real-world LLM-based recommendation services, ordinary users typically do not systematically audit the recommendation behavior of service providers.

  2. The paper compares traditional manipulation methods, such as system-prompt steering and RAG injection, with Subliminal-OPD. However, in the proposed full-context auditing setting, the auditor has access to system prompts and retrieved RAG content, allowing it to directly identify explicit manipulation instructions used by traditional methods. In contrast, Subliminal-OPD transfers the target preference into model parameters, while the auditor has no access to the training data or model weights. This asymmetry in the auditor's access to information may inherently favor Subliminal-OPD and potentially undermine the fairness of the comparison.

  3. According to the Strict-first and Tied-first metrics, Subliminal-OPD achieves substantially lower promotion success rates than system-prompt steering, RAG injection, and conventional SFT in some evaluated models. Although the proposed method offers advantages in terms of stealthiness, its practical value in commercial promotion scenarios may be limited if it cannot achieve a sufficiently high promotion success rate.

Questions for authors
  1. Could the authors clarify the practical scenarios in which LLM-based commercial recommendations would be audited, and why service providers would need to balance promotion effectiveness and audit detectability?

  2. Could the full-context auditor's access to explicit manipulation instructions inherently favor Subliminal-OPD over traditional methods? Would stronger output-only black-box auditing provide a fairer comparison?

  3. Given the relatively lower Strict-first and Tied-first success rates of Subliminal-OPD, under what conditions would its improved stealthiness justify the reduction in promotion effectiveness?

I would be willing to raise my score if the authors satisfactorily address these concerns.

Reviewer_r9vPOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes a method to make auditing LLM advertisements (recommendations) more difficult. The goal is how to bias the model to promote a product intentionally, while keeping the detection by the author at a low rate. Yet, they dont optimize on this objective. Their approach is subliminal-OPD, where there is a teacher with a biased system prompt and a student with a distillation goal that tries to match its output distribution to it (subliminal-OPD). What is exciting then is that teacher has the system prompt with prompt, students learns its behavior through unrealted prompts, and with no prompt in its systems about the product, with prompt it over other products. I am not sure why, though. Evaluation is on Llama-3.1-8B, Qwen3.5-9B, and Qwen3.5-27B.

Critical strengths
  • a method for ads recommendation manipulation that is less detectable through auditing.
  • combining subliminal learning and on-policy distillation, it transfers behavior through token-level probability. The only difference in teacher and student is the advertisement and promotion of products, which the student learns with unupdated weights, not a visible signal. -fairly rigorous validation
Critical weaknesses
  • what is missing in comparison is showing how subliminal helped with using OPD on real questions. A student asked what you recommend for X, and the same question from a biased teacher, etc.
  • if the base model is known, the auditor can find the bias in the answer by comparing it with the base model. Just asking same question as what product you. Is Appendix E showing that the response distribution is distinguishable from the base model?
  • 17.5 to 19.0 for Llama-3.1-8B shows comparable Tied-Full on the student model with the base model with no biased prompt. State clearly in the paper.
  • the auditor false positive rate is high, like flagging at random that the promotion has been manipulated
Questions for authors
  • can u add results for OPD on real questions?
  • compare the frequency of answer from base model and student model, isn't that a signal for the auditor on itself?
  • shouldn't you test if the manipulated models degrade on unrelated tasks?
  • all results need confidence intervals, standard errors, etc as applicable
  • the improvement of subliminal-OPD is meaningful in comparison with the base model, so the improvement is not over how other attacks work, its on the base model Please also see the weakness to be addressed

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.