- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes GibberCraft, a method for replacing readable system prompts with obfuscated hard-token sequences while preserving the original instruction-following behavior. It formulates prompt optimization using ADMM to alternate between continuous embedding optimization and discrete token projection, augmented with Gumbel-perturbed hard-token sampling to mitigate projection-induced degradation. Experiments on Llama 3.1 8B and Qwen3 8B report improved behavioral fidelity over Pape et al. (2025), 100% task success on two controlled instruction-following tasks, and up to 9.3× optimization speedup.
Critical strengths
- Clear formulation and promising empirical improvements. The paper provides a clear ADMM-based formulation of discrete prompt obfuscation, combining continuous optimization with hard-token projection and Gumbel-perturbed sampling. The experiments demonstrate promising improvements in instruction-following success and computational efficiency over the closest prior work.
Critical weaknesses
-
Limited evaluation of behavioral fidelity and confidentiality. The instruction-following evaluation primarily measures output format rather than semantic correctness or completeness, while the security evaluation focuses on lexical prompt extraction rather than recovery of the underlying instructions or business logic. These evaluations do not fully substantiate the paper's claims of behavioral preservation and prompt protection.
-
Limited task complexity and generalizability. The experiments focus on relatively simple tasks with short system prompts (10 to 32 tokens). It remains unclear whether the method generalizes to realistic applications involving longer prompts, multiple constraints, and complex workflows.
Questions for authors
See weaknesses, and:
-
The bold font in Table 1 does not consistently indicate the best performance, potentially misleading readers. The authors should clarify the formatting and qualify their claims, particularly where baseline outperforms GibberCraft.
-
The reported 100% instruction-following success rates are potentially misleading, as the LLM judge primarily assesses output format. Could the authors include task-specific correctness metrics, such as factual consistency for headline generation and entity-level F1 for name extraction?
-
Could the authors compare against other continuous-to-discrete prompt optimization methods beyond GCG?
-
Why does GibberCraft use only 128 training examples compared with 640 for Pape et al., how are these subsets selected, and how sensitive are the results to training-set size?
Thanks for your constructive feedback and suggestions. We clarify the concerns as follows:
Font Format and Advantage Clarification: We have removed the method-based bolding and clarified the comparison. GibberCraft achieves better similarity scores in most reported comparisons. Our contribution is efficient preservation of task behavior, with the clearest advantages appearing in tasks that require more specific instruction following.
On the simpler QA and summarization tasks, both methods produce responses that resemble those generated with the readable prompt, and several similarity scores are close. These metrics capture different aspects of similarity, including lexical overlap and embedding-based similarity. For example, on Qwen summarization, the method of Pape et al. achieves higher BLEU, while both methods obtain cosine similarity of 0.94, and GibberCraft achieves higher BERTScore. A marginal advantage on an individual metric does not establish overall superiority in this case.
In addition, QA and summarization overlap substantially with the models’ default capabilities, as indicated by the nontrivial response similarity obtained with an empty system prompt. A higher similarity score is therefore informative, but does not by itself establish better preservation of the requested instruction. This also motivates our evaluation on headline generation and name extraction tasks.
We therefore additionally evaluate headline generation and name extraction, where the empty-prompt model does not naturally produce the required output behavior. GibberCraft achieves substantially stronger results overall in these controlled settings. Most notably, it achieves 100% task-form compliance on name extraction for both models, compared with 3% and 9% for Pape. These results demonstrate our clearest advantage in preserving specific instruction-following behaviors.
Comparison with continuous-to-discrete prompt optimization: Following your suggestion, we added PEZ, a continuous-to-discrete prompt optimization method, as a baseline. Its results are reported in Tables 1 and 3. GibberCraft achieves stronger response-similarity results overall across the evaluated settings.
Entity-Level F1: We've included entity-level F1 for name extraction. Results are shown in Table 14 in the appendix.