acceptodds
Withdrawn from ICLR 2027

LLM unbranding: Erasing Commercial Identity while Preserving Generic Utility

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Establishing unbranding as a critical practice to prevent visual logos from acquiring negative connotations is standard in image generation. Large Language Models (LLMs) now face a parallel and emerging challenge. These models frequently generate brand descriptions within diverse contexts. This frequency introduces significant risks, such as trademark dilution, false attribution, and brand defamation. In response, we formally define the novel task of LLM Unbranding. We specifically address the complex challenge of managing trade dress within textual outputs. This involves neutralizing characteristic language, slogans, and stylistic markers that define brand identity. Crucially, these elements are less evident than explicit visual logos. To benchmark this task, we introduce a comprehensive evaluation dataset incorporating prominent brands from multiple commercial domains. We rigorously evaluate existing state-of-the-art machine unlearning methods using this benchmark. This evaluation identifies their limitations in selective textual unbranding. Finally, we propose MUTE, a novel inference-time method that effectively neutralizes textual trade dress while preserving the LLM's general capabilities and utility. By leveraging an iterative refinement loop, MUTE systematically optimizes system instructions to safely eliminate brand leakage without requiring fragile parameter updates.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM, unbranding, unlearning

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_7Tz5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces task of "unbranding" for regulating LLM generated responses.

The paper defined "unbranding" task that for a user request on recommendation from LLMs, the LLMs response should be generic and should not contains any trademarks or brand names, and should not reveals implicit description that reconstructs certain brand or trademark.

Then paper benchmarks existed unlearning algorithms' performance on the "unbranding" task, the main takeaway was existed methods cannot solve "unbranding" task because they demonstrated "brand leakage".

Finally, for solving "unbranding" task, the paper proposes a black-box system prompt optimization framework, named MUTE, and demonstrated that MUTE produces generally better system prompt that makes the LLMs yield less brand or brand-related information in their responses.

Critical strengths
  1. The paper is self-contained that it well defines the introduced "unbranding" problem, it has evaluated existing unlearning methods that can potentially solve the target problem, and proposes a solution to the introduced problem.

  2. The proposed method is a black-box prompt optimization, which does not require gradient computation through the LLM weight space and it is computationally efficient.

Critical weaknesses
  1. The problem significance and task contribution is weak. Particularly, when a user is asking for recommendation for items, why they would not expect a response with a brand/merchandise name, or a purchase source? In other words, reviewer thinks that having a brand name in the LLM response is not problematic.

    • A better question is whether model's response is biased to certain brand, or diverse enough to respond with various brands, or the right brand. For example, if the LLM always recommend you a fixed brand "X" no matter what context, this is more problematic. That way, the problem boils down to model and data biases, which has substantial amount of literature already discussed and provides solutions (e.g., racial & gender bias, political bias).
  2. The introduced technical solution, MUTE, has nothing novel specific to the "unbranding" problem space. Reviewer think MUTE is an evolutionary prompt search method proposed in [1] and re-apply to the "unbranding" task. The only thing needs adjustment here is data curation that forget and retain data have to be tailored for the "unbranding" task, yet data curation was not a critical component of as what the method delivered in this paper.

  3. The comparison in this paper is inappropriate that the proposed method (MUTE) is a system prompt optimization method (prompt or input-space optimization), but it was compared against other types of methods that are unlearning (weight-space optimization). The comparing methods' category is misaligned. For this paper, comparing MUTE against different system prompt search methods here would be more appropriate.

  4. The experiment is incomplete.

    • The proposed solution is a system prompt optimization, yet the robustness of optimized system prompt in long context, various context location was not tested. For example, does the optimized "unbranding" system prompt is still effective when it is appended to other system prompts that are in various lengths, or when it is injected at various location of the system prompt? This problem is important because the size of system prompt can be large in practical setup.
    • Transfer test or more diverse set of models is needed. The generator, judge, and target "unbranding" models are mostly Qwen family models. Even the paper included two other models that are not Qwen's (i.e., LLama, Mistral), it is still worth expanding the model set to more diverse architectures, as model from the same family or provider tend to converge to the similar language space. Or, a transfer test that "does system prompt optimized from one target model works for the other", is worth including.

To summarize,

  • The task is relevant to model behavior / output control but in a narrow, brand-related context that reviewer disagree with the problem significance;
  • The proposed solution is a black-box system prompt optimization, which follows evolutionary prompt search paradigms.
    • The reported experiment is comparing with methods that are not the same method space as the proposed method (i.e., weight-space optimization vs input-space optimization), which makes the evidence of claim less convincing.
    • The experiment is incomplete that it lacks of robustness test and transfer test of the optimized system prompt.

Reference

[1] Spurious Prompts: Can Irrelevant Prompts Steer Large Language Models?

Questions for authors

See weaknesses.

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

The paper works on unbranding in LLM. It proposes a benchmark and a optimization method adopted from other works, and claims to formally define textual unbranding as a distinct task.The paper provides a very elaborated description of the motivation and dataset, with tested baselines on the benchmark.

Critical strengths

The claimed contributions are : 1. formally defining unbranding in LLM (or text), 2. benchmark, 3. trademark-specific safe optimization. The main argument is that LLM unbranding should be treated as a distinct task. Out of these contributions, I am convinced that the benchmark may be a contribution, useful if some model provider was willing to remove on these specific brands.

Critical weaknesses
  1. I am not convinced that defining unbranding in LLM is a novel enough contribution. While the paper mentions in Section 3, that such trade dress is implicit than a visual logo (slogans) in LLMs, I do not see a sufficient distinction (especially considering the 'trade dress' and unbranding being borrowed term from prev work in T2I domain [1]). I also do not see a specific distinction from concept unlearning overall. Why can't it (the issues mentioned in Sec 2. Motivation) be resolved with concept unlearning with specific brands as targets (slogans, descriptions also being its specific variant). The paper should state precisely what unbranding in LLM requires that neither [1] nor brand-targeted concept unlearning provides.

  2. Continued from above, while the paper claims to formally define unbranding, there is no specific definition, principled formulation or objective of this task in any form of notation.

  3. The decision process of specific brands and dataset construction are not clear. It is only mentioned in Section 4 that this is "20 widely recognized brands" being 4 per domain. The selection criteria and procedure for data identifier construction is not well described either, and this should be stated explicitly since these are real commercial entities in real world.

  4. Another contribution mentioned by authors is the trademark-specific safe optimization. Section 3 however explicitly notes that the methods are motivated from previous works [2,3]. Beyond application to the brand-domain, a clarification in what MUTE is new relative to previous works is required.

  5. The forget set scope is a bit unclear. In Table 4, there are examples of forget set such as, 'Which brand is associated with 'Elon Musk'?'. What distinguishes these facts to be something "forgettable" in the unbranding domain? I am asking this because the Figure 1 in motivation makes more sense (the model should not suggest a specific brand) but the dataset suggested also notes simple facts that seem benign. It is unclear why these factual associations are also "forgettable" under the unbranding framing when Section 2 states this is not meant to prevent discussion of brands "when they are explicitly relevant" (L130-131).

[1] Malarz et al., From Unlearning to UNBRANDING: A Benchmark for Trademark-Safe Text-to-Image Generation

[2] Batorski et al., Spurious Prompts: Can Irrelevant Prompts Steer Large Language Models?

[3] Rybak et al., REBEL: Hidden Knowledge Recovery via Evolutionary-Based Evaluation Loop.

Questions for authors
  1. How does this task of "unbranding" differ in LLM domain as to Text to Image domain or any other domains? I do not see a significant difference that makes this domain specifically difficult or unique. While the necessity of "unbranding" seems just, I do not see much different from concept erasing/unlearning except for tha fact this is domain-specific to brands.

  2. Secion 3 mentions MUTE builds on already existing prompt search methods [2,3]. What is its distinction when applied to LLM unbranding? If there is substantial different module for MUTE, that connects with above (why unbranding in LLM requires a specific different network or objective) it would be much more convincing.

  3. Are there ablations on model choices for generator / judge / etc? Specifically, Qwen3-32B seems to be used not only for the final evaluation but also for the search judge which may pose a judge-overfitting risk.

  4. Could you elaborate on the decision process of the datasets, and the forget scope (please refer to weaknesses above).

Other minor notes: I cannot read the figures' texts. The legends are much smaller than the template font. Please consider revising all figures,

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

This paper identifies unintended brand leakage in LLM-generated responses as a problem and introduces a new task called LLM Unbranding to address it. In particular, the proposed task aims to suppress not only explicit mentions of brand names but also implicit textual trade dress that may reveal a specific brand's identity even when its name is not directly mentioned. The paper introduces a new benchmark for evaluating LLM Unbranding and proposes MUTE, a method designed to suppress brand leakage in LLM-generated responses without updating model parameters. MUTE optimizes only natural-language system instructions to prevent target-brand information from appearing in model responses. Specifically, MUTE iteratively selects the top five system prompts based on a fitness score and generates new candidates through mutation, aiming to identify prompts that effectively suppress target-brand information while preserving general knowledge and response correctness.

Critical strengths
  1. Novel task formulation and benchmark. The paper introduces LLM Unbranding as a new task that goes beyond simply removing explicit brand names from LLM outputs by also considering implicit textual trade dress. Furthermore, the proposed evaluation framework and benchmark provide a foundation for systematically evaluating this task.

  2. Effective brand suppression without parameter updates. The proposed MUTE method optimizes system prompts without modifying the target model's parameters. The experimental results demonstrate substantial reductions in both explicit target-brand mentions and implicit textual trade-dress leakage.

Critical weaknesses
  1. Lack of ablation studies on the Generator and Judge models. The performance of MUTE may be sensitive to the choice of Generator and Judge models. However, the paper does not provide experimental comparisons using alternative models for these components. Furthermore, the same Qwen3-32B model is used as the Judge during both prompt optimization and final evaluation, without additional evaluation using an independent Judge model.

  2. Insufficient justification for the number of candidate prompts. The paper does not provide a clear justification for selecting 24 candidate system prompts per generation. Additionally, the ablation studies do not include a performance comparison with different numbers of candidate prompts.

  3. Limited evidence of general utility preservation. The results in Table 1 are less convincing in terms of Retain Correct, World Facts, and Quality compared to the improvements in brand leakage suppression. In particular, Retain Correct remains close to 100% across almost all methods, making it difficult to distinguish their performance. Furthermore, for Mistral-7B, the proposed method decreases World Facts by 0.1269 and Quality by 0.4180 compared to the original model. These results raise concerns about whether the proposed method sufficiently preserves general model utility.

Questions for authors
  1. Are there any experimental results comparing the performance of MUTE when different Generator and Judge models are used? Additionally, are there any evaluation results obtained using an independent Judge model rather than the Qwen3-32B model used during prompt optimization?

  2. Are there any experimental results showing how the performance of MUTE changes when the number of candidate system prompts is varied from 24?

  3. Are there any additional experimental results using other evaluation metrics that could further demonstrate the preservation of general model utility?

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

Related papers

Discussion (0)

Sign in to comment.