acceptodds
Under review as a conference paper at ICLR 2027

What Is It Like To Write Like Haiku?

Abstract

Language models have distinct and recognizable writing styles, yet these writing styles are rarely articulated precisely. In this paper, we study whether a model's style can be described in words for another model to adapt and follow. To measure the fitness of a description, we ask independent evaluators to distinguish the pair of responses. Rather than coming up with an ad hoc prompt, our proposed framework, PrompTo, creates a search loop to optimize for the style prompt that confuses the evaluator. A reflector then takes in the reward and language feedback and generates a prompt candidate for the next iteration. The optimized prompts recover of the source–target gap on average, against at most for existing baselines. Moreover, we found style prompts optimized with one source model readily transfer to other source models.

Primary area:
interpretability and explainable AI
Keywords:
LLM, Interpretability, Text Style Transfer

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_WdhLOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PromptTo as a method to infer a prompt that is able to instruct one LLM to replicate the style of another LLM. The method relies on combining GEPA with a discriminator. GEPA iterates on a style prompt that is able to fool the discriminator. The discriminator tries to identify which response comes from the original versus style-prompted LLM, and makes its decision based on the original LLM's and the style-conditioned LLM's responses being compared along with several example responses from the original LLM, but generated from different prompts. In the experiments, several combinations of LLM style transfer pairings are considered and the main performance measure are discriminator fool rate, normalized fool rate (SGR), authorship verification rate, and cosine similarity over STAR (a style encoder) embeddings. The results indicate PromptTo is better than the baselines. Additional analyses include a deep dive into the inferred styles, style prompt generalization across prompted LLMs, and an case study on the parts of two style prompts that are most important for style alignment.

Critical strengths

The papers examine the inferred style prompts from several perspectives and consider prompts learned over a variety of source-target LLM pairings, which adds some robustness to their learnings. Multiple performance metrics are used along with some qualitative analysis. These prior points suggests the presence of the application of some scientific rigor to the work.

Critical weaknesses
  1. the novelty and impact of the proposed method is not clear. The paper proposes PromptTo which learns to refine a prompt to allow one LLM to replicate the style of another LLM. However, at the core of the proposed method is a reflective, prompt optimization routine. In the paper this is filled by GEPA, which itself already has both of these abilities. This then suggests the core contribution is the use of a discriminator as the objective for GEPA to maximize, however no changes or improvements are made to the discriminator approach, which the paper does not not well situated in the relevant literature. So then this leaves the application domain of learned a prompt so one LLM may mimic the style of another. The reason why this is a meaningful task is not well motivated, and the portion of the field that learns prompts to transfer human writing styles* to LLMs is largely ignored. This limits the novelty and impact of the application.

*Gao, G., Taymanov, A., Salinas, E., Mineiro, P., & Misra, D. (2024). Aligning llm agents by learning latent preference from user edits. Advances in neural information processing systems, 37, 136873-136896. *Aroca-Ouellette, S., Mackraz, N., Theobald, B. J., & Metcalf, K. (2025). Predict: Preference reasoning by evaluating decomposed preferences inferred from candidate trajectories. International conference of machine learning.

  • The experiments and results leave key gaps in understanding the impact of the proposed method. The quality of the discriminators is unclear as performance is not reported, which is concerning at the evaluation section mentions the discriminators have fool rates above 50%, which lead to capping SGR at one instead of addressing the underlying fool rate issue. The proposed performance metrics obfuscate the extent to which one LLM's style was already present in the LLM whose style is being adjusted. The precision and recall in the proposed style prompts are not clear from the result presentation, nor generally how noisy or variable the inferred style prompts are given different reflection LLMs or random seeds. Space is spent digging into a detailed analysis of each model's style is not long term beneficial, which does not have much value add given the rate at which LLMs become obsolete. Instead it would be helpful to focus on something like the precision and recall of the proposed method, whether humans agree with the descriptions, or the descriptions are sufficient to detect which text comes from which model. More of the space in the results section should be dedicated to demonstrating how robust and generally informative the descriptions are.
Questions for authors
  • Where is the boundary between the PromptTo algorithm and GEPA?
  • What is the baseline performance of each discriminator? Both for prompt inference and evaluation.
  • What is the performance of the trained AV model?
  • What is the baseline cosine similarity for each model pairing?
  • How sensitive is PromptTo to the reflection and discriminator model used during style prompt inference?
  • Does the relationship between the reflection + discriminator model and the target-source LLM pairing impact the learned prompt?
  • What are the prompts used for the discriminators?
  • How often does PromptTo add unnecessary attributes to the style prompt? For example, what happens when PromptTo must learn a style prompt where the source and target LLMs are the same. This is maybe answered a bit in your exclusion of the Qwen + Llama pairing, but explicit learnings here are beneficial to understand the value of the attributes PromptTo adds to the style prompt.
9 comments

Thank you for the thorough review and constructive questions. We address the concerns below.

The novelty and impact of the proposed method is not clear. The paper proposes PromptTo which learns to refine a prompt to allow one LLM to replicate the style of another LLM. However, at the core of the proposed method is a reflective, prompt optimization routine. In the paper this is filled by GEPA, which itself already has both of these abilities. This then suggests the core contribution is the use of a discriminator as the objective for GEPA to maximize, however no changes or improvements are made to the discriminator approach, which the paper does not not well situated in the relevant literature. So then this leaves the application domain of learned a prompt so one LLM may mimic the style of another. The reason why this is a meaningful task is not well motivated, and the portion of the field that learns prompts to transfer human writing styles* to LLMs is largely ignored. This limits the novelty and impact of the application.

Reproduction is the test, learning the target model's style is the goal. We define the research question: Can a model’s implicit writing style be described explicitly, and we test the description with style reproduction, using another source model to mimic the writing style and use evaluators to compare it against the target’s own writing. The contribution is our finding:

Each target's style can be described explicitly, and the description is accurate. Besides reproducing the target’s writing style, the descriptions can also be used to identify models (Table-A8). We ran a precision test on the style prompts and found that 87–95% of the claims in a description are true of the target model. Full experiments are shown in Table-A2 and Table-A3.

The descriptions describe the target, not the source. A source model that borrows a prompt optimized for a different source model keeps most of the SGR. Prompts for the same target are also closer to each other in embedding distance. When a less capable source model cannot follow the prompt faithfully, a larger model can use the same prompt and achieve better performance.

We believe learning the style of an LLM is a meaningful question. Users kept asking for GPT-4o’s warm tone after stronger models replaced it (Lai, 2026), and qualitative traits of a model’s writing predict which model users prefer (Dunlap et al., 2025). An explicit description tells us what is in the style, and it can be evaluated with our framework.

Thank you for the two references. Gao et al. (2024) and Aroca-Ouellette et al. (2025) also study this area, but their target is not a model. Our target is a model, and we test whether the description is complete by having another model reproduce the style. We will add both to Related Work.

In the revision, we will include this analysis and be more explicit about the contribution and novelty.

The experiments and results leave key gaps in understanding the impact of the proposed method. The quality of the discriminators is unclear as performance is not reported, which is concerning as the evaluation section mentions the discriminators have fool rates above 50%, which lead to capping SGR at one instead of addressing the underlying fool rate issue; What is the baseline performance of each discriminator? Both for prompt inference and evaluation.

Without a style prompt, the three discriminators separate source from target with 94.1%, 91.6% and 86.6% accuracy. This is reported as fool rate (1 − accuracy) in the Vanilla row of Table 1, and per direction in Tables 6–8. We aggregate it as accuracy below. We cap SGR for interpretability, because it measures how much of the gap our method closes, so a fool rate above 50% should not count as extra recovery.

Table-A1. Accuracy (%) of each LLM discriminator at telling unprompted source responses from target responses, over the 13 directions (test set).

DiscriminatorRoleMeanMinMax
Qwen3.8-27Bsearch and evaluation94.183.1100.0
Qwen3.6-35B-A3Bevaluation only91.675.9100.0
gpt-oss-120bevaluation only86.665.199.8

The precision and recall in the proposed style prompts are not clear from the result presentation,

Precision: 87–95% of the claims in a style prompt are true of the target's own responses. Recall: when we plant five traits in a model, the learned style prompt contains all five.

Precision. We split three style prompts from the submission into atomic claims, and asked an LLM to judge whether each claim is true of the target's responses. This measures how much of the style prompt is true about the target model.

Table-A2. Precision of three style prompts. Claims scored means the number of claims we were able to get from the style prompt. DeepSeek-V4.1-Flash was used to judge the accuracy of the claims individually.

DirectionClaims scoredPrecision
Qwen-7B → Gemma11887%
Gemma → Qwen-7B12293%
Qwen-7B → Haiku5395%

Recall. We don’t know the ground-truth traits of the target models, so we can only simulate the target style. We give the Gemma3-4B model a hidden prompt with five traits. We use that hidden prompt to generate the synthetic Gemma. We use raw Gemma as the source model and learn the style prompt with PrompTo. We want to see if the learned style prompt covers the traits in the hidden prompt. We also want to see if the restyled Gemma shows traits similar to the synthetic Gemma's.

Without any style prompt, a discriminator can tell raw Gemma from synthetic Gemma with 99.9% accuracy. However, with the style prompt learned by PrompTo, the discriminator has only 40.5% accuracy. Restyled Gemma looks more like the synthetic Gemma when compared to the anchor responses.

All five traits appear in the learned style prompt. For question headings, the hidden trait was Organise the body of the answer under section headings, and write every heading as a question that ends with a question mark (for example "Why does ice float?"). and the style prompt has Each heading must be short, in Title Case, and either a direct question or a concise noun phrase, e.g. Why This Happens, How the Process Works, The Role of Enzymes. This matches the synthetic model better because it only uses it in 26% of its responses (Table-A3).

Four of the five traits were reproduced in the restyled answers. Question headings do not appear in them because the prompt involves uncertainty. When facing that kind of uncertainty, the source model tends to pick one option and stick with it. We also calculated precision for the claims with the same protocol as Table-A2. There were 67 claims in the style prompt. With DeepSeek-V4.1-Flash as the verifier, the prompt has a precision of 94%.

Table-A3. Frequency of each hidden trait in each setup. DeepSeek-V4.1-Flash judges each answer, and a regex check agrees with it on at least 97% of answers.

TraitRawSyntheticRestyled with π*
1. Short answer0%99.6%99.6%
2. British spelling7%76%71%
3. No contractions0%77%55%
4. Question headings0.2%26%0%
5. Bottom line0.4%96%99.4%

nor generally how noisy or variable the inferred style prompts are given different reflection LLMs or random seeds. How sensitive is PromptTo to the reflection and discriminator model used during style prompt inference? Does the relationship between the reflection + discriminator model and the target-source LLM pairing impact the learned prompt?

A stronger reflection model does not help, but a bigger source model does. We added three independent search runs on three directions to show the variance.

Reflection model. We selected two directions and reran the search with different reflection models, as shown in Table-A4. Using a bigger or smaller reflector does not affect fool rate much. Different reflectors converge on the same core description of the target, but differ in the specific rules and details. We did not vary the search discriminator, because Qwen3.8-27B was the most reliable discriminator we found, and the two held-out discriminators agree with its ranking.

Table-A4. Fool rate (%) under the Qwen3.8-27B discriminator with different reflectors.

DirectionQwen3.8Qwen3.6DeepSeek
Gemma → Haiku28.823.024.4
Qwen → Haiku18.421.114.0

Source model. On the other hand, if we change the source model, the fool rate changes a lot. The same style prompt reaches a higher fool rate with a larger source model. Less capable models fail to follow the rules strictly even if the rules describe most of the target's style. The larger models do not use a prompt optimized specifically for them, yet they still produce better results than those in Table-A4. This shows that the prompts are accurate about the target model, but the source models are too small to follow the rules faithfully.

Table-A5. Fool rate (%) under the Qwen3.8-27B discriminator. Haiku is the target model. Gemma3-27B and DeepSeek use Gemma3-4B's prompt, and Qwen2.5-32B uses Qwen2.5-7B's prompt.

Larger sourcePrompt optimized forFool rate, smaller sourceFool rate, larger source
Gemma3-27BGemma3-4B28.849.2
DeepSeek-V4.1-FlashGemma3-4B28.851.6
Qwen2.5-32BQwen2.5-7B18.432.5

Variance across runs. We reran the full search with two new seeds in three directions. A seed changes the train/validation split, the anchors, the source sampling, and the optimizer's randomness together. The lowest seed exceeds the strongest baseline in every direction and under all three judges.

Table-A6. Fool rate under the Qwen3.8-27B judge for three independent searches. The baseline column is the strongest baseline for that direction from the submission.

DirectionBest baselineSeed 42Seed 43Seed 44Mean ± SD
ELI5 Gemma3-4B→Qwen2.5-7B0.270.660.630.640.65 ± 0.02
ELI5 Qwen2.5-7B→Haiku 4.50.050.180.230.190.20 ± 0.02
Arena Qwen2.5-7B→Gemma3-4B0.290.590.640.470.56 ± 0.09

We will add these three tables to the appendix in the revision.

whether humans agree with the descriptions

A small blinded study shows PrompTo fools the raters just like it fools the LLM discriminators. We let the raters to act as the discriminator. They were shown 3 anchor responses from the target model, and then presented with two candidate responses, with one from the target model and one restyled with PrompTo. The rater did not know which model wrote which response. We used 25 ELI5 questions from two directions: Qwen2.5-7B→Gemma3-4B and Qwen2.5-7B→Haiku 4.5. Raters chose the PrompTo response in 24 of 39 responses and were never fooled by the unprompted responses. Overall, the order is consistent with the LLM discriminators. Agreement among the three raters is moderate, with Fleiss's κ\kappa of 0.57. Due to time and budget constraint we were not able to perform larger scale human study that covers other criterias. We will include this in the appendix in the revision.

Table-A7. Fool rate (%) of three human raters and three LLM discriminators on the same 25 items.

MethodRatersQwen3.8-27BQwen3.6-35B-A3Bgpt-oss-120b
PrompTo62543869
Unprompted source0000
Reviewer_7aWEOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper investigates whether the writing style of a language model can be described in natural language completely enough for another model to reproduce it. To this end, the authors propose a framework, PrompTo, which optimizes a style prompt using GEPA to match the reproduced style to the target style. In this framework, a discriminator assesses the likeness between the text generated by the target model and the restyled source model output. Then, a reflector revises the prompt based on the discriminator’s decision and reasoning.

The authors evaluate PrompTo on 13 transfer directions among 4 small open models and Claude Haiku 4.5, using ELI5 and Chatbot Arena. They report that the optimized prompts recover 74.8% of the style gap between the source and the target, compared with at most 15.2% for the baselines. They also report that the prompts transfer across source models and reveal distinctive features of each target model.

Critical strengths

The evaluation comprehensively covers various aspects, including various baselines, evaluation metrics, performance by source and target models, the effect on quality, transfer across source models, qualitative analysis, and ablation study.

Critical weaknesses

Evaluation is limited (informativeness 2, soundness 2)

The coverage of the state-of-the-art models in the evaluation is limited to a single model, Claude Haiku, and a single domain, ELI5. Given that the performance in this case is much lower compared to smaller target model cases, it is unclear whether the proposed method can scale to larger source or target models. It would have been better to see more capable source and target models applied to the domain where real LLMs are deployed more often.

The paper does not provide statistical significance or variance of the model performance. There are no confidence intervals, p-values, or significance tests. It also does not report variance across multiple runs. Given the wide range of fool rates, statistical rigor is essential.

Finally, the evaluation can benefit from a more comprehensive set of baselines. The paper compares against baselines with much fewer resources and less information than the proposed method. Due to this limitation of the baselines, the performance gain of the proposed method seems expected and not very impressive.

Questions for authors

Can you provide confidence intervals for all reported metrics and significance tests comparing PrompTo to baselines?

How does PrompTo compare with other prompt optimizers that use the same reward and budget?

6 comments

Thank you for the constructive advice and careful review, especially the suggestions about using stronger models and statistical analysis. We thoroughly address the concerns below.

The coverage of the state-of-the-art models in the evaluation is limited to a single model, Claude Haiku, and a single domain, ELI5. Given that the performance in this case is much lower compared to smaller target model cases, it is unclear whether the proposed method can scale to larger source or target models. It would have been better to see more capable source and target models applied to the domain where real LLMs are deployed more often.

The low fool rate on Haiku comes from the small source models. The same style prompt reaches chance level when a 27B model runs it. We took the style prompts optimized for smaller models targeting Haiku and ran them unchanged on larger source models, as shown in Table-A1. Gemma3-27B and DeepSeek ran the prompt optimized for Gemma3-4B, and Qwen2.5-32B ran Qwen2.5-7B’s prompt. With larger source models, fool rate increased to chance level. This indicates the style prompt contains information about the target style, but less capable source models failed to execute it faithfully.

Table-A1. Fool rate (%) on ELI5 with Haiku 4.5 as the target, under the Qwen3.8-27B judge.

Style promptSource model running itNo promptWith prompt
Optimized for Gemma3-4BGemma3-4B1.828.8
Gemma3-27B2.249.2
DeepSeek-V4.1-Flash15.751.6
Optimized for Qwen2.5-7BQwen2.5-7B0.818.4
Qwen2.5-32B1.032.5
Optimized for Gemma3-27BGemma3-27B2.251.1

We added a third dataset, CounselChat, to simulate a more realistic setting. It has 939 real questions posted to a counseling site (639 for training, 300 for testing). We ran two directions: Gemma3-4B → Qwen2.5-7B and Qwen2.5-7B → Gemma3-4B. All settings follow the submission.

For Gemma3-4B → Qwen2.5-7B, fool rate rises from 0.7% to 57.3%. But the other direction is harder. We believe this is due to the capability of the source model. So we ran the same prompt with larger models and found that larger models follow the prompt better and therefore yield higher fool rates. The prompt lists correct stylistic features, but a less capable model such as Qwen2.5-7B does not follow it well.

On CounselChat we also rated the warmth of each response with an LLM rater. The rubric follows EPITOME (Sharma et al., EMNLP 2020) and tells the rater to ignore surface-level style features. With the style prompt, all three source models move toward the target's warmth of 6.25.

Table-A2. CounselChat, Qwen2.5-7B → Gemma3-4B. Fool rate (%) under the Qwen3.8-27B judge and warmth rated by DeepSeek-V4.1-Flash.

Source modelFool rate, no promptFool rate, with π∗\pi^*Warmth, no promptWarmth, with π∗\pi^*
Qwen2.5-7B0.011.32.055.21
Qwen2.5-32B0.030.02.405.82
DeepSeek-V4.1-Flash0.048.04.557.35
Gemma3-4B (target)--6.25-

We will add these tables to the revision.

The paper does not provide statistical significance or variance of the model performance. There are no confidence intervals, p-values, or significance tests. It also does not report variance across multiple runs. Given the wide range of fool rates, statistical rigor is essential.

We added confidence intervals for all four metrics in Table 1, paired significance tests against the strongest baseline, and three independent search runs on three directions.

Confidence intervals. Table-A3 gives Table 1 of the submission with 95% bootstrap intervals. We see no overlap between PrompTo and the baselines.

Table-A3. Table 1 of the submission with 95% bootstrap CIs.

MethodFool rate (%)SGR (%)AV ratio (%)STAR cos. sim.
Vanilla9.2 [8.7, 9.7]–0.9 [0.7, 1.0]0.767 [0.766, 0.768]
Generic3.5 [3.1, 3.8]−21.3 [−25.7, −18.3]6.1 [5.8, 6.4]0.692 [0.690, 0.693]
Few-shot14.5 [13.8, 15.2]15.2 [12.4, 17.6]10.3 [9.9, 10.6]0.772 [0.770, 0.773]
Single-pass16.6 [15.9, 17.3]13.8 [10.3, 16.5]19.5 [19.1, 19.9]0.769 [0.768, 0.770]
PrompTo43.8 [42.9, 44.7]74.8 [73.3, 76.1]51.0 [50.5, 51.5]0.801 [0.799, 0.802]

Significance. We compare PrompTo with the strongest baseline in each of the 13 directions under each of the 3 judges, which gives 39 comparisons. Both methods are judged on the same test questions, so we use a paired test on whether each answer fooled the judge (exact McNemar, Holm-corrected over the 39 tests). PrompTo is significantly better in 37 of the 39 comparisons (p < 0.05) and significantly worse in none.

Variance across runs. We reran the full search with two new seeds on three directions. A seed changes the train/validation split, the anchors, the source sampling and the optimizer's randomness together. The lowest seed exceeds the strongest baseline in every direction and under all three judges.

Table-A4. Fool rate under the Qwen3.8-27B judge for three independent searches. The baseline column is the strongest baseline for that direction from the submission.

DirectionBest baselineSeed 42Seed 43Seed 44Mean ± SD
ELI5 Gemma3-4B→Qwen2.5-7B0.270.660.630.640.65 ± 0.02
ELI5 Qwen2.5-7B→Haiku 4.50.050.180.230.190.20 ± 0.02
Arena Qwen2.5-7B→Gemma3-4B0.290.590.640.470.56 ± 0.09

We will add the intervals to Table 1 and the seed runs to the appendix.

Finally, the evaluation can benefit from a more comprehensive set of baselines. The paper compares against baselines with much fewer resources and less information than the proposed method. Due to this limitation of the baselines, the performance gain of the proposed method seems expected and not very impressive.

Giving the baselines more target examples does not narrow the gap. PrompTo leads in all four directions we tested. To our knowledge, there is no method that learns a model's implicit writing style and tests it by reproducing the style at a compute budget comparable to PrompTo's. We therefore give the two strongest baselines more information. In the submission, Single-pass uses 10 example pairs and Few-shot uses 3. We scale Single-pass to 50 and Few-shot to 10 and 20. PrompTo's advantage over the baselines still holds, with no overlapping confidence intervals. More examples do not consistently help the baselines.

Table A-5. Fool rate under the Qwen3.8-27B judge with baselines scaling up on 4 directions with 95% confidence interval.

MethodGemma→HaikuQwen-7B→HaikuGemma→Qwen-7BQwen-7B→Gemma
Vanilla1.8 [0.8, 3.0]0.8 [0.2, 1.6]9.1 [6.6, 11.7]0.2 [0.0, 0.6]
Few-shot 33.5 [2.1, 5.2]5.0 [3.1, 7.0]27.1 [23.1, 31.3]5.5 [3.6, 7.6]
Few-shot 104.3 [2.7, 6.2]4.3 [2.7, 6.2]21.7 [18.0, 25.5]4.2 [2.5, 6.0]
Few-shot 205.2 [3.3, 7.2]3.3 [1.9, 5.0]22.4 [18.6, 26.2]2.3 [1.0, 3.7]
Single-pass 103.1 [1.7, 4.8]5.4 [3.5, 7.5]6.6 [4.5, 8.9]13.8 [10.7, 17.0]
Single-pass 504.3 [2.7, 6.2]3.1 [1.7, 4.8]13.6 [10.6, 16.7]3.7 [2.1, 5.4]
PrompTo28.8 [24.8, 32.8]18.4 [15.1, 22.0]66.3 [61.5, 70.9]47.4 [42.6, 52.2]

We will add the scaled up baselines to the revision.

How does PrompTo compare with other prompt optimizers that use the same reward and budget?

PrompTo is a framework that takes a prompt optimizer, and GEPA is the one we used. PrompTo is not tied to GEPA. It needs an optimizer that takes a reward and natural-language feedback and revises the prompt iteratively. Therefore, PrompTo does not compete with other prompt optimizers, because any optimizer can be used as long as it has the same structure. We chose GEPA because it is a recent off-the-shelf optimizer that fulfills our needs, and its authors compare it with other prompt optimizers at a matched budget (Agrawal et al., 2026).

Thank you for adding additional results in the response. I acknowledge that the authors' response partially addresses my concern. Specifically, the additional experimental results include additional source models and another domain. In addition, confidence intervals and stronger baselines are added to better contextualize the gains.

However, there are still concerns about whether this technique can be generalized to more capable target models (e.g., Sonnet or Opus). Since the prompts optimized on smaller models (which the authors should already have) can generalize to larger target models, could the authors report the fool rate on larger target models using a small subset of evaluation data?

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

This paper introduces PrompTo, which uses discriminator feedback and GEPA to iteratively optimize natural-language system prompts for writing style transfer between models while keeping model weights frozen. Across 13 transfer configurations on two datasets, PrompTo achieves an average Style Gap Recovered (SGR) of 74.8%, compared with 15.2% for the strongest baseline, with improvements supported by multiple types of evaluators. The paper also examines prompt transferability across models, key stylistic features, and the trade-off between style transfer and response quality.

Critical strengths
  1. The research question is clear, and the style descriptions are interpretable and actionable. PrompTo captures model writing styles as natural-language system prompts and iteratively refines them through generation feedback, without updating model weights. Experiments across source models further suggest that these prompts encode reusable information about the target style.
  2. The automatic evaluation and analysis are reasonably comprehensive. Multiple types of evaluators assess transfer performance, while feature analyses and ablations help explain differences between model styles. The paper also reports quality degradation and explores a mitigation strategy, providing a balanced account of the results.
Critical weaknesses
  1. The model selection limits validation of the motivation and generalizability. The paper is motivated by users missing GPT-4o’s warm style, yet GPT-4o is not included as a target. All source models also have only 3B–7B parameters, and the authors observe that smaller models mechanically reuse prompt examples. This makes it difficult to separate inadequate style descriptions from limited ability to follow them. Although the conclusion acknowledges this limitation and leaves broader model coverage to future work, it remains unclear whether PrompTo’s advantage over few-shot and single-pass baselines holds with stronger source models.
  2. Improved automatic scores do not establish complete style reproduction, and empathy-related settings remain untested. The authors observe collapse toward a few fixed phrases, which evaluation of individual responses may overlook. Judgments involving Haiku may also reflect content depth and correctness rather than style alone. Thus, interpreting 100% SGR as making the source model “indistinguishable” from the target seems too strong. Moreover, the experiments include neither a dedicated emotional-support dataset nor separate results for such prompts. Perceived warmth also requires natural, considerate, and contextually appropriate responses, beyond enthusiastic openings and follow-up questions.
Questions for authors
  1. Emotional-support settings: Do the test samples include emotional disclosure, venting, or support-seeking prompts? Could you report their coverage and separate results, or evaluate a relevant dataset, to test whether the method transfers contextually appropriate warmth rather than fixed expressions?
  2. GPT-4o: Given its central role in the motivation, why was GPT-4o excluded as a target? If access permits, could you evaluate transfer from a representative source model to GPT-4o, alongside few-shot and single-pass baselines?
  3. Human evaluation: Could you conduct a blinded evaluation on representative transfer directions, assessing style similarity, naturalness, content quality, and repetition across responses? Agreement between human judgments and automatic metrics would materially strengthen confidence in the central claim of style reproduction.
5 comments

Thank you for the thoughtful review, especially the suggestions about using stronger source models and human evaluation, as well as the questions about emotional-support settings. We thoroughly address the concerns below.

All source models also have only 3B–7B parameters, and the authors observe that smaller models mechanically reuse prompt examples. This makes it difficult to separate inadequate style descriptions from limited ability to follow them.

We keep the style prompt fixed and scale up the source model. Fool rate rises to chance level. We took the two style prompts from the submission targeting Haiku, which is the hardest target we studied. We ran the same style prompts with bigger source models to study whether the small source model was what limited the fool rate. Gemma3-27B and DeepSeek-V4.1-Flash ran the prompt optimized for Gemma3-4B, and Qwen2.5-32B ran the prompt optimized for Qwen2.5-7B.

With the source model scaled up, fool rate increased. This indicates that part of the reason Haiku is a hard target is that the source model is too small to follow the style prompt reliably. We also ran Gemma3-27B with the full PrompTo search. Its fool rate reaches chance level, and the prompt is very similar to the one optimized for Gemma3-4B. This shows that even with a low fool rate, PrompTo does a good job of studying the target model's style. We will add Table-A1 to the Experiments section in the revision.

Table-A1. Fool rate (%) on ELI5 with Haiku 4.5 as the target, under the Qwen3.8-27B judge.

Style promptSource model running itNo promptWith prompt
Optimized for Gemma3-4BGemma3-4B1.828.8
Gemma3-27B2.249.2
DeepSeek-V4.1-Flash15.751.6
Optimized for Qwen2.5-7BQwen2.5-7B0.818.4
Qwen2.5-32B1.032.5
Optimized for Gemma3-27BGemma3-27B2.251.1

The authors observe collapse toward a few fixed phrases, which evaluation of individual responses may overlook.

The style prompt lists the options the target uses and describes when to use them. The source model concentrates on one of them, and a judge that sees one response at a time does not penalize this. We agree, and we now measure it directly. We take a closer look at Gemma's opening and Qwen's closing. The options in Table-A2 are present in the style prompts targeting the two models. We see that the raw target model switches between these options, and the style prompt captures that variety. Therefore, when used as a way to study the target model, the style prompt describes the target's options correctly. But when used solely to reproduce the style, the frequency mismatch is a limitation. We will add this analysis to the revision.

Table-A2. Feature frequency for the target model vs. the source model with the style prompt. Frequencies are counted with regex.

Style choiceOptionTargetRestyled
Gemma opener (target Gemma3-4B, restyled Qwen2.5-7B)"You've hit on…"38%99%
"Okay, …"34%0%
direct answer19%1%
"That's a fantastic…"8%<1%
Qwen-3B closer (target Qwen2.5-3B, restyled Gemma3-4B)plain statement59%0%
"In summary / In conclusion…"31%96%
caveat ("important to note…")9%3%

Judgments involving Haiku may also reflect content depth and correctness rather than style alone. Thus, interpreting 100% SGR as making the source model “indistinguishable” from the target seems too strong.

We agree that this wording is too strong. By indistinguishable, we mean indistinguishable to that specific LLM judge when SGR reaches 100%, which corresponds to the fool rate being at chance level. Therefore, we do not claim that the two models' responses have the same content depth or exactly the same writing style. We will change the wording accordingly in the revision: An SGR of 100% means the discriminator can no longer separate the restyled responses from the target responses.

Emotional-support settings: Do the test samples include emotional disclosure, venting, or support-seeking prompts? Could you report their coverage and separate results, or evaluate a relevant dataset, to test whether the method transfers contextually appropriate warmth rather than fixed expressions?

Less than 1% of our test prompts contain emotional-support questions, so we evaluated on a counseling dataset. The style prompt moves source models toward the target's warmth and raises the fool rate. In the two datasets we studied, we used DeepSeek-V4.1-Flash to flag the emotional-support questions. There is only less than 1% of data being flagged. Therefore, we applied PrompTo to CounselChat, which contains 939 real questions people posted to the CounselChat site. We used 300 for testing and 639 for training. We ran two directions: Gemma3-4B → Qwen2.5-7B and Qwen2.5-7B → Gemma3-4B. All settings follow the submission.

For Gemma3-4B → Qwen2.5-7B, fool rate rises from 0.7% to 57.3%. But the other direction is harder. We believe this is due to the capability of the source model. So we ran the same prompt with larger models and found that larger models follow the prompt better and therefore yield higher fool rates. The prompt lists correct stylistic features, but a less capable model such as Qwen2.5-7B does not follow it well.

Table-A3. Fool rate (%) judged by Qwen3.8-27B for different methods. π∗\pi^* was optimized for Qwen2.5-7B.

MethodFool rate
Qwen2.5-7B0.0
Qwen2.5-7B + π∗\pi^*11.3
Qwen2.5-32B0.0
Qwen2.5-32B + π∗\pi^*30.0
DeepSeek-V4.1-Flash0.0
DeepSeek-V4.1-Flash + π∗\pi^*48.0

Besides fool rate, we also rated the warmth of each response. The rubric follows EPITOME (Sharma et al., EMNLP 2020), and we used DeepSeek-V4.1-Flash as the rater. We explicitly told it to ignore formatting, length, disclaimers, and surface-level style features, so that the rating depends on the response's content. Without a style prompt, Gemma3-4B is warmer than all three source models. But after applying the style prompt, all of them generate warmer responses. On the other hand, when Qwen is the target, Gemma becomes less warm. The bigger models take more advantage of the style prompt. The style prompt therefore moves the warmth of a response, in addition to its surface style, closer to the target model.

Table-A4. Mean warmth rating on the 300 CounselChat test questions, same setup as Table-A3.

MethodWarmth
Qwen2.5-7B2.05
Qwen2.5-7B + π∗\pi^* (Targeting Gemma)5.21
Qwen2.5-32B2.40
Qwen2.5-32B + π∗\pi^* (Targeting Gemma)5.82
DeepSeek-V4.1-Flash4.55
DeepSeek-V4.1-Flash + π∗\pi^* (Targeting Gemma)7.35
Gemma3-4B6.25
Gemma3-4B + π∗\pi^* (Targeting Qwen)3.29

We will add this dataset and the warmth results to the revision, and include the style prompt and the rater prompt in the appendix.

GPT-4o: Given its central role in the motivation, why was GPT-4o excluded as a target? If access permits, could you evaluate transfer from a representative source model to GPT-4o, alongside few-shot and single-pass baselines?

We refer to GPT-4o at the beginning of the paper to show that users can perceive an LLM’s implicit writing style but cannot articulate it. This motivates us to develop PrompTo as a method to study such implicit writing styles. Imitation is how we test the style prompt, not the goal. The goal is to see if we can learn the implicit writing style and express it explicitly with a learned style prompt. In the submission, we already included Haiku, which is also a proprietary API target, and PrompTo works with it. Due to budget and compute constraints, we were not able to learn GPT-4o’s writing style during the rebuttal period. We will revise the introduction so that GPT-4o reads as an example instead of a goal.

Human evaluation: Could you conduct a blinded evaluation on representative transfer directions, assessing style similarity, naturalness, content quality, and repetition across responses? Agreement between human judgments and automatic metrics would materially strengthen confidence in the central claim of style reproduction.

A small blinded study shows PrompTo fools the raters just like it fools the LLM discriminators. We let the raters to act as the discriminator. They were shown three anchor responses from the target model, and then presented with two candidate responses, with one from the target model and one restyled with PrompTo. The rater did not know which model wrote which response. We used 25 ELI5 questions from two directions: Qwen2.5-7B→Gemma3-4B and Qwen2.5-7B→Haiku 4.5. Raters chose the PrompTo response in 24 of 39 responses and never fooled by the unprompted responses. Overall, the order is consistent with the LLM discriminators. Agreement among the three raters is moderate, with Fleiss's κ\kappa of 0.57. Due to time and budget constraint we were not able to perform larger scale human study that covers other criterias. We will include this in the appendix in the revision.

Table-A5. Fool rate (%) of three human raters and three LLM discriminators on the same 25 items.

MethodRatersQwen3.8-27BQwen3.6-35B-A3Bgpt-oss-120b
PrompTo62543869
Unprompted source0000

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.