acceptodds
Under review as a conference paper at ICLR 2027

TextReg: Mitigating Prompt Distributional Overfitting via Regularized Text-Space Optimization

Abstract

Large language models (LLMs) are highly sensitive to the prompts used to specify task objectives and behavioral constraints. Many recent prompt optimization methods iteratively rewrite prompts using LLM-generated feedback, but the resulting prompts often become longer, accumulate narrow sample-specific rules, and generalize poorly beyond the training distribution. We study this failure mode as *prompt distributional overfitting* and argue that it reflects a lack of representation control in discrete text-space optimization. We formalize this view through *representational inefficiency*, a dual-factor measure that decomposes prompt inefficiency into capacity cost and scope narrowness, attributing distributional prompt overfitting to their coupled growth during optimization. We propose **TextReg**, a regularization framework that realizes a soft-penalty objective through regularized textual gradients, combining Dual-Evidence Gradient Purification, Semantic Edit Regularization, and Regularization-Guided Prompt Update. Across multiple reasoning benchmarks, **TextReg** substantially improves out-of-distribution (OOD) generalization, with accuracy gains of up to +11.8% over TextGrad and +16.5% over REVOLVE. The code is available at https://anonymous.4open.science/r/textreg_code.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Prompt Optimization, Prompt Distributional Overfitting, Representational Inefficiency, Out-of-Distribution Generalization, Regularization

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

This paper proposes TextReg, a regularization framework for automatic prompt optimization. The authors argue that iterative rewriting based on textual feedback may lead to prompt distributional overfitting through increasing prompt length and accumulating sample-specific rules. To address this issue, TextReg introduces three stages: filtering textual feedback using local examples and historical rules, detecting excessive prompt growth and increased rule specificity, and guiding prompt updates with both task and regularization feedback. The method is evaluated by optimizing prompts on three source tasks (Logical Deduction and Tracking Shuffled Objects with three objects, and GSM8K) and testing them on the harder five- and seven-object variants and on SVAMP/MultiArith across four open-weight test engines. Compared with TextGrad and REVOLVE, the reported advantages are most pronounced on the harder BBH variants, where both baselines often fall below the unoptimized CoT prompt, while differences on the arithmetic transfer are small.

Critical strengths

S1. Well-motivated and coherent method design. This paper addresses a practical failure mode of automatic prompt optimization: accumulating increasingly specific instructions that transfer poorly. Distinguishing prompt length from rule specificity provides a useful design perspective, and the three-stage pipeline connects feedback filtering, edit-level diagnosis, and regularization-guided rewriting in a coherent way.

S2. Transparency and internal consistency. Anonymous code release, cost and call accounting, full templates, and candour about the ID loss (about two points on the BBH source tasks) and about prompts barely updating on GSM8K under the strict criterion.

S3. Clear organization and helpful figures. The two motivating factors (length growth, rule narrowing) map directly onto the two regularization channels, the experiments are organized around explicit questions (Q1–Q3), and the appendix tests the alternative explanations one by one. Fig. 1 makes the phenomenon concrete with a worked example, and Fig. 2 ties each stage to its equation using the same example.

Critical weaknesses

W1. The novelty of TextReg is not entirely clear. Section 2 acknowledges that overfitting in automatic prompt optimization is a known issue and discusses prior methods that use simplification instructions, external memory, or meta-reflection. TextReg combines feedback filtering, edit inspection, and textual regularization, but it is difficult to tell which parts are new and how they differ from existing approaches. The additional comparisons with DLPO and REMO are helpful, but better performance does not necessarily imply a new technical contribution. A more direct comparison of how these methods use feedback and guide prompt updates would help clarify what TextReg adds beyond prior work.

W2. No standard deviations (std). Every number is a mean over four seeds, so the std was sitting there and should have been reported; leaving it out is hard to understand. It matters because the margins that decide the paper are small. Once the baselines get a matched budget (Table 2), TextReg is ahead of the runner-up by 2.0, 1.4, 0.5 and 0.5 points, and 0.5 points on a 250-question test set is one question. The τC\tau_C sweep alone moves the number by 2.5 points. I could not tell from the paper whether any of the differences in Tables 2, 5, 6 or 7 would survive a second set of seeds.

W3. The link between the formalization and the algorithm is unclear. I(p)I(p), s(r)s(r), Eq. (5) and Eq. (14) are introduced and then, in the paper’s own words, never computed, estimated or optimized. What actually runs is a three-way LLM classifier, a 20% length threshold, an LLM yes/no on specificity, and an extra paragraph in the optimizer prompt. The formalism also disagrees with the method: by Eq. (4) a long prompt made of general rules has I(p)I(p) = 0, yet the capacity channel fires on length alone. And RuleBank “recurrence” is counted over 36 training examples from one distribution, so it measures how common a rule is in-distribution, which is not scope over X.

Questions for authors

Please see the Critical Weaknesses section.

4 comments
AuthorsOpenReview ↗

[Part 1/2] Response to Reviewer cdgN

Thank you for your careful review and for recognizing the motivation and experimental design of our work. We provide our responses below and hope they will help you reconsider your evaluation and update your score.

Weakness 1: Novelty relative to prior prompt optimization methods.

Thank you for raising this concern. We would like to clarify the key differences in motivation and mechanism.

1. Distinction from prior methods. DLPO's regularization uses a fixed instruction to simplify or remove unnecessary content, without analyzing prompt edits. REMO focuses on accumulating optimization experience through a mistake notebook and adapting the optimizer through a meta-controller. Neither explicitly controls both prompt length growth and rule specificity during optimization.

2. Our technical contribution. TextReg introduces a regularized text-space optimization framework with three complementary mechanisms:

  • Dual-Evidence Gradient Purification filters case-specific feedback using local execution contexts and historical RuleBank evidence.
  • Semantic Edit Regularization detects length growth and scope narrowing in realized prompt edits and generates targeted regularization guidance.
  • Regularization-Guided Prompt Update integrates task and regularization feedback while preserving task fidelity.

Unlike fixed simplification or memory-based adaptation, TextReg regulates prompt evolution at both the feedback and edit levels, directly targeting the two sources of representational inefficiency.

3. Controlled comparison with static simplification. We add a fixed compress-and-generalize instruction to each TextGrad update under the same optimization setup as TextReg, without RuleBank or adaptive regularization.

OOD DatasetTextGradCompress + GeneralizeTextReg
LD 5obj49.856.156.9
LD 7obj46.149.551.5
Track 5obj59.459.668.4
Track 7obj57.753.564.1
Final prompt tokens351133160

Static simplification improves Logical Deduction but offers little or no benefit on Tracking. TextReg outperforms it across all four settings, supporting the value of adaptive regularization over fixed simplification.

Revision: We have strengthened the mechanism-level comparison in Related Work to clarify TextReg's technical contributions.

Weakness 2: Missing uncertainty estimates and threshold sensitivity.

Thank you for raising these concerns. We address statistical uncertainty and threshold sensitivity separately.

1. Statistical uncertainty. We have updated Tables 1, 2, 5, 6, and 7 to report mean ± std. All main-paper experiments use four optimization seeds, while supplementary experiments use 2–4 seeds due to computational and budget constraints. We have revised Appendix A.3 to clarify this distinction.

We agree that small margins should be interpreted cautiously. Nevertheless, TextReg achieves the highest or tied-highest mean accuracy compared with TextGrad and REVOLVE in 21 of 24 settings (Table 1), including cases with substantial baseline degradation. For example, on Phi-3.5 LD 5obj, TextReg achieves 57.9 versus 46.1 for TextGrad. Our conclusions therefore rely on the overall trend rather than isolated small gains.

2. Threshold sensitivity. Appendix E.3 evaluates τC\tau_C from 0.05 to 0.5. Across this ten-fold range, the 7-object OOD mean varies from 54.6% to 57.1%, while final prompt length remains between 152 and 181 tokens.

The 2.5-point variation warrants caution when interpreting small improvements, but the results do not indicate a sharp dependence on a single threshold value.

AuthorsOpenReview ↗

[Part 2/2] Response to Reviewer cdgN

Weakness 3: Inconsistency between the formalization and the actual algorithm.

Thank you for this detailed critique. We have revised the Problem Formulation and Method to clarify the purpose of the mathematical formalization and its connection to the implemented algorithm.

1. Formalization and implementation. Our mathematical formulation is intended to characterize the objectives and operations of discrete text-space optimization, rather than prescribe numerical computations. Representational inefficiency is now characterized through two structural factors (Eq. 4), capacity cost and scope narrowness, without assuming a specific numerical relationship or requiring either factor to vanish when the other does. The idealized objective (Eq. 5) expresses the principle of controlling these factors while optimizing task performance. Similarly, Eq. (14) conceptually characterizes task-faithful rewriting preferences, which the optimizer realizes through natural-language instructions rather than numerical minimization.

TextReg operationalizes this principle through two signals:

  • Capacity: Measured directly through relative prompt-length growth, ρC\rho_C (Eq. 9).
  • Scope: Assessed directionally by the Semantic Diff Analyzer MΔM_\Delta (Eqs. 10–11), with RuleBank providing supporting evidence.

All three stages are designed around this objective: Gradient Purification filters case-specific and stylistic feedback that may increase capacity or narrow scope; Semantic Edit Regularization detects these structural changes and generates corrective guidance; and Regularization-Guided Prompt Update integrates this guidance with task feedback while preserving task fidelity.

2. Role of RuleBank. As described in the paper, RuleBank recurrence serves only as a soft prior for potential generalizability, not a numerical estimate of scope over X\mathcal{X}. Since future OOD inputs are unavailable during optimization, recurrence across training contexts provides indirect evidence that a rule may generalize beyond observed examples. Importantly, no numerical mapping or hard threshold is imposed between recurrence and scope; the Semantic Diff Analyzer separately assesses the direction of scope changes.

3. Empirical evidence. We provide complementary evidence for both structural factors.

Capacity. TextGrad expands from 41 to 351 tokens while OOD accuracy drops from 42.0% to 39.8%. TextReg limits prompt growth to 160 tokens and improves OOD accuracy to 47.0%.

Scope. We use Gemini-2.5-Flash, independent of our GPT-4o optimization pipeline, to evaluate directive applicability on held-out OOD inputs.

MetricTextGradREVOLVETextReg
Directives per prompt19.67.27.0
Narrow directives5.81.81.8
Prompt tokens351131160

A directive is considered narrow if it is inapplicable to at least half of the OOD inputs. TextReg substantially reduces narrow directives relative to TextGrad. REVOLVE achieves a similar count but lower OOD accuracy.

Length-only controls (Table 2). A CAPO-style length penalty and DLPO's fixed simplification instruction both underperform TextReg, suggesting that reducing prompt length alone is insufficient to reproduce its OOD performance.

These results provide empirical support for controlling both capacity growth and scope narrowing through targeted textual regularization.

I have read the authors’ response and decided to maintain my original score.

AuthorsOpenReview ↗

Thank You & Looking Forward to Further Discussion

Thank you for reviewing our response. We respect your decision and remain happy to provide any further clarifications if needed.

Reviewer_DVq9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies prompt distributional overfitting in iterative prompt optimization (e.g., TextGrad, REVOLVE). Over successive steps, prompts often grow longer and accumulate sample-specific rules, which degrades out-of-distribution performance. The authors define representational inefficiency as I(p) = C(p) * (1 - s(p)), combining token length and rule-scope narrowness (Eq. 4). They propose TextReg, a three-stage prompt optimization pipeline comprising Dual-Evidence Gradient Purification, Semantic Edit Regularization, and Regularization-Guided Update. The authors evaluate TextReg on Big-Bench Hard and arithmetic transfer benchmarks using four 7–8B models, with appendix experiments reporting compute-matched baselines, trajectory logs, and all-weak backward models.

Critical strengths

The submission addresses a relevant practical problem in discrete prompt optimization and provides several solid empirical observations:

  1. Empirical documentation of prompt bloat: The paper tracks optimization trajectories and illustrates that iterative prompt updates can expand prompt length while hurting out-of-distribution accuracy, such as TextGrad expanding from 41 to 351 tokens in Table 3 (Section 5.1, Table 3).

  2. Compute and baseline reporting in the appendix: Appendix B includes baseline comparisons that account for backward call budgets and test explicit length penalties (DLPO, CAPO-pen in Table 2), which provides a more complete accounting of optimization cost (Table 2, Appendix B).

Critical weaknesses

Despite its practical motivation, several methodological and empirical aspects require further clarification:

  1. Unverified theoretical metric: The paper introduces representational inefficiency I(p) = C(p) * (1 - s(p)) and the regularized objective L + lambda * I(p) (Eq. 4–5), but neither quantity is ever computed or evaluated. The actual implementation relies on an LLM prompt heuristic without showing that the formal metric correlates with OOD generalization (Section 3.3, Section 4.3).

  2. Gains shrink under compute-matched budgets: When backward calls are matched (180 calls in Table 2), TextReg's advantage over TextGrad and DLPO drops to 0.5–1.4 points on hard variants (e.g., Tracking 7-obj: 64.1 vs 63.6), while arithmetic transfer gains remain below 1 point (Table 1, Table 2).

  3. Absence of statistical variance: All main tables report 4-seed means with no standard deviations, confidence intervals, or significance tests, making small margins (<= 1 point) difficult to distinguish from run-to-run noise (Table 1, Table 2).

  4. Missing simple prompt-compression baseline: The ablations do not test a simpler control, such as TextGrad followed by a standard LLM prompt to "compress and generalize" without the RuleBank pipeline.

  5. Heuristic classification bias and in-distribution drop: The semantic diff analyzer explicitly defaults to CASE_PATCH (Appendix G.3), which may induce over-pruning and explain the 3.8% drop in in-distribution accuracy observed on Tracking 3-obj (Table 4).

Questions for authors

Please refer to the detailed concerns and clarification points listed in the Critical Weaknesses section above.

6 comments
AuthorsOpenReview ↗

[Part 1/2] Response to Reviewer DVq9

Thank you for your detailed review and constructive feedback. We appreciate your recognition of the practical relevance of our work and its empirical contributions. We provide our responses below and hope they will help you re-evaluate your evaluation and update your score.

Weakness 1: The representational-inefficiency objective is not directly evaluated.

Thank you for this important observation. We have revised the Problem Formulation and Method to clarify the relationship between the conceptual objective and its implementation.

1. Formalization and implementation. Representational inefficiency is now characterized through two structural factors (Eq. 4), capacity cost and scope narrowness, without prescribing a specific numerical relationship. The idealized objective (Eq. 5) expresses the principle of controlling these factors while optimizing task performance. TextReg operationalizes them through:

  • Capacity: Measured directly through relative prompt-length growth, ρC\rho_C (Eq. 9).
  • Scope: Assessed directionally by the Semantic Diff Analyzer MΔM_\Delta (Eqs. 10–11), with RuleBank providing contextual evidence.

All three stages are designed around this objective: Gradient Purification filters case-specific and stylistic feedback that may increase capacity or narrow scope; Semantic Edit Regularization detects these structural changes and generates corrective guidance; and Regularization-Guided Prompt Update integrates this guidance with task feedback while preserving task fidelity.

2. Empirical evidence. We provide complementary evidence for both factors.

Capacity. TextGrad expands from 41 to 351 tokens while OOD accuracy drops from 42.0% to 39.8%. TextReg limits growth to 160 tokens and improves OOD accuracy to 47.0%.

Scope. We use Gemini-2.5-Flash, independent of our GPT-4o optimization pipeline, to evaluate directive applicability on held-out OOD inputs.

MetricTextGradREVOLVETextReg
Directives per prompt19.67.27.0
Narrow directives5.81.81.8
Prompt tokens351131160

A directive is considered narrow if it is inapplicable to at least half of the OOD inputs. TextReg substantially reduces narrow directives relative to TextGrad. REVOLVE achieves a similar count but lower OOD accuracy.

Length-only controls (Table 2). A CAPO-style length penalty and DLPO's fixed simplification instruction both underperform TextReg, indicating that reducing prompt length alone is insufficient to reproduce its OOD performance.

These results support the two-factor design and its implementation through targeted textual regularization, rather than direct numerical optimization of the conceptual objective.

Weakness 2: Reduced gains under compute-matched budgets.

Thank you for raising this concern. We would like to clarify two points.

1. Performance under matched budgets. Although compute matching narrows some margins, TextReg still outperforms the matched-budget TextGrad and REVOLVE baselines across all four BBH OOD settings (Table 2). Compared with TextGrad, TextReg achieves an average improvement of 2.3 points, with only one setting showing a margin below 1 point. The average improvement over REVOLVE is even larger, at 7.2 points.

2. Advantage given to the baselines. This comparison is particularly favorable to TextGrad and REVOLVE. With 180 backward calls, they use best-of-N search to generate multiple candidate prompts and select the best based on task performance. In contrast, TextReg uses 174 calls to produce a single prompt update per step, with additional computation devoted to controlling representational inefficiency rather than exploring more candidates.

Despite this search advantage for the baselines, TextReg maintains higher OOD accuracy across all four settings, supporting the effectiveness of its targeted regularization beyond additional optimization compute.

AuthorsOpenReview ↗

[Part 2/2] Response to Reviewer DVq9

Weakness 3: Missing standard deviations and statistical uncertainty.

Thank you for highlighting the importance of statistical uncertainty. All experiments in the main paper use four seeds, and we have updated Tables 1 and 2 to report mean ± std. Due to computational and budget constraints, supplementary experiments in the appendix use 2–4 seeds. We have revised Appendix A.3 to clarify these differences.

1. Sources of variability. Some settings exhibit relatively large standard deviations due to independent LLM-driven optimization trajectories. Each prompt undergoes 12 iterative updates, where variations in generated feedback and revisions can lead to different final prompts and performance.

2. Interpretation of results. While small individual differences should be interpreted cautiously, TextReg achieves the highest or tied-highest mean accuracy compared with TextGrad and REVOLVE in 21 of 24 settings (Table 1). On arithmetic benchmarks, baseline accuracy is already high, leaving limited room for further improvement. Our conclusions therefore rely on the consistent overall trend across tasks and models rather than isolated small gains.

Weakness 4: Missing a simple compress-and-generalize baseline.

Thank you for suggesting this control. We added a fixed compress-and-generalize instruction to every TextGrad optimizer update, using the same backbone, initial prompt, optimization steps, batch size, and seeds as TextReg.

At each iteration, the instruction asks the optimizer to remove redundant or overly specific content and rewrite example-specific instructions as general rules. Unlike TextReg, this baseline does not use RuleBank, Dual-Evidence Gradient Purification, or Semantic Edit Regularization to provide adaptive regularization guidance.

New Experiment: Compress-and-Generalize Baseline (OOD Acc %)

OOD DatasetTextGradCompress + GeneralizeTextReg
LD 5obj49.856.156.9
LD 7obj46.149.551.5
Track 5obj59.459.668.4
Track 7obj57.753.564.1
Final prompt tokens351133160

Insight. Static compression and generalization improves Logical Deduction but provides little benefit on Tracking 5obj and degrades Tracking 7obj. Unlike this fixed instruction, TextReg adaptively identifies and addresses structural inefficiencies during optimization while preserving task performance. Its stronger OOD results suggest that targeted regularization is more effective than generic simplification across these tasks.

Weakness 5: Default-to-CASE_PATCH may cause over-pruning and ID degradation.

Thank you for raising this concern. We would like to clarify two points.

1. Motivation for conservative classification. Defaulting to CASE_PATCH is intended to improve the reliability of GENERALIZED_RULE classification. A rule is classified as generalizable only when supported by a high-frequency RuleBank entry or relevance across multiple execution contexts. Without such evidence, treating a case-specific rule as generalizable risks introducing seemingly broad but actually overfitted instructions. We therefore favor conservative classification to prevent unsupported specialization.

Importantly, a CASE_PATCH classification by the Semantic Diff Analyzer does not directly remove a rule from the prompt. Regularization guidance is integrated with task feedback, with necessary task corrections prioritized and candidate updates subject to validation.

2. ID–OOD trade-off. The observed ID decrease cannot be directly attributed to classification bias. Unregularized optimization can accumulate lengthy, case-specific instructions that improve performance on the source distribution but generalize poorly to OOD inputs. For example, TextGrad achieves higher ID accuracy than TextReg on LD 3obj (92.2% vs. 91.5%) and Tracking 3obj (89.8% vs. 86.0%), yet performs worse on OOD variants. By limiting such overfitting, TextReg may sacrifice some ID accuracy for stronger OOD generalization. We consider this trade-off worthwhile given its overall OOD improvements, while acknowledging that conservative classification may occasionally reject useful rules.

Revision: We have revised Appendix D to clarify the ID–OOD trade-off and explain why controlling prompt overfitting may lead to modest reductions in ID accuracy.

AuthorsOpenReview ↗

Discussion Deadline Approaching; We Welcome Further Feedback

Dear Reviewer DVq9,

We sincerely appreciate your thoughtful review and valuable feedback, which have helped us improve our manuscript. With the discussion deadline approaching, we hope that our responses and additional experiments have helped address your concerns.

If our responses have addressed your main concerns, we would be grateful if you could reconsider your rating. We also remain happy to address any additional questions or concerns.

Thank you again for your time and consideration.

Warm regards,

Authors

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

TextReg reduces prompt overfitting by filtering example-specific feedback and guiding updates toward compact, generalizable instructions.

Critical strengths
  • Addresses a practical problem in automatic prompt optimization.
  • Includes compute-matched controls, stronger baselines, and useful ablations.
  • Evaluates transfer across datasets and models.
Critical weaknesses
  • The “representational inefficiency” explanation is not independently validated.
  • Most evidence comes from a few structured reasoning tasks.
  • Novelty is incremental relative to DLPO and REMO, though both are evaluated.
  • Small gains over stronger baselines need clearer uncertainty estimates.

[1] DLPO: Towards a Robust, Efficient, and Generalizable Prompt Optimization Framework from a Deep-Learning Perspective, 2025 [2] Reflection-Enhanced Meta-Optimization Integrating TextGrad-style Prompt Optimization with Memory-Driven Self-Evolution

Questions for authors

What evidence shows that the gains come from reducing rule specificity, beyond simply producing shorter or better-written prompts?

2 comments
AuthorsOpenReview ↗

[Part 1/2] Response to Reviewer GDvC

Thank you for your thoughtful review and for recognizing the practical relevance of our work, the compute-matched controls, and the transfer evaluation. We address your concerns below and hope they will help you reconsider your evaluation and update your score.

Weakness 1 & Question: Evidence for the role of rule specificity in prompt overfitting.

Thank you for raising this important question. We argue that prompt overfitting involves both excessive length growth and the accumulation of narrowly applicable rules. We provide three complementary analyses.

1. Length reduction alone is insufficient. REVOLVE produces shorter prompts than TextReg (131 vs. 160 tokens, Table 3), yet achieves lower OOD accuracy (43.9 vs. 47.0). Similarly, DLPO's static simplification and a CAPO-style length penalty both underperform TextReg (Table 2). These results suggest that reducing prompt length alone is insufficient to explain the OOD gains.

2. Rule applicability. We use Gemini-2.5-Flash, independent of our GPT-4o optimization pipeline, to evaluate the applicability of final-prompt directives to held-out OOD inputs.

MetricTextGradREVOLVETextReg
Directives per prompt19.67.27.0
Narrow directives5.81.81.8
Prompt tokens351131160

A directive is considered narrow if it is inapplicable to at least half of the OOD inputs. TextReg substantially reduces narrow directives relative to TextGrad. REVOLVE achieves a similar count but lower OOD accuracy, suggesting that neither prompt length nor narrow-directive count alone fully explains the observed performance.

3. Controlled rule injection. To examine the effect of rule specificity while controlling for prompt length, we append either four case-specific rules derived by GPT-4o from training failures or four general procedural rules of matched length to the same TextReg prompts. We evaluate the modified prompts on Qwen2-7B-Instruct and Meta-Llama-3-8B-Instruct.

OOD DatasetOriginal (TextReg)+ Narrow rules+ General rules
LD 5obj54.352.354.9
LD 7obj49.945.950.3
Track 5obj49.946.948.5
Track 7obj43.739.343.0

At matched length, adding narrow rules reduces OOD accuracy by up to 4.4 points, while adding general rules has a smaller effect. This provides direct evidence that rule specificity also plays an important role in OOD generalization.

Insight. These results show that both prompt length and rule specificity are important factors in OOD generalization. TextReg addresses both through feedback purification and semantic edit regularization, while preserving task-relevant corrections during prompt updates.

Weakness 2: Limited evaluation beyond structured reasoning tasks.

We appreciate this concern and would like to clarify two points.

1. Motivation for structured reasoning tasks. Structured reasoning tasks provide objectively verifiable answers, enabling reliable and reproducible evaluation of OOD generalization. Moreover, benchmarks such as BBH offer controlled difficulty shifts (3→5/7 objects), allowing us to examine generalization under well-defined distribution changes. We also cover mathematical reasoning through cross-dataset transfer from GSM8K to SVAMP and MultiArith.

2. Broader task coverage. To examine generality beyond structured reasoning, we additionally evaluate TextReg on GPQA-main (knowledge-intensive QA) and CommonGen (open-ended generation) in Appendix F (Table 7). TextReg achieves the best performance on both tasks (33.9% and 88.0%), outperforming TextGrad and REVOLVE.

These results provide preliminary evidence beyond the primary reasoning benchmarks, while broader evaluation remains valuable.

AuthorsOpenReview ↗

[Part 2/2] Response to Reviewer GDvC

Weakness 3: Novelty relative to DLPO and REMO.

Thank you for raising this concern. We would like to clarify the key differences in motivation and mechanism.

1. Distinction from prior methods. DLPO's regularization uses a fixed instruction to simplify or remove unnecessary content, without analyzing prompt edits. REMO focuses on accumulating optimization experience through a mistake notebook and adapting the optimizer through a meta-controller. Neither explicitly controls both prompt length growth and rule specificity.

2. Our contribution. TextReg introduces a regularized text-space optimization framework with three complementary mechanisms:

  • Dual-Evidence Gradient Purification filters case-specific feedback using local and historical evidence.
  • Semantic Edit Regularization detects length growth and scope narrowing in prompt edits and generates targeted guidance.
  • Regularization-Guided Prompt Update integrates task and regularization feedback while preserving task fidelity.

Unlike fixed simplification or memory-based adaptation, TextReg directly regulates the structural evolution of prompts.

3. Empirical evidence. TextReg achieves the highest OOD accuracy across all four BBH settings (Table 2), supporting the effectiveness of its targeted design for mitigating prompt overfitting.

Revision: We have strengthened the mechanism-level comparison in Related Work to clarify TextReg's technical contributions.

Weakness 4: Missing uncertainty estimates.

Thank you for emphasizing the need to report uncertainty. We have updated Tables 1 and 2 to report mean ± std.

1. Sources of variability. Some settings exhibit relatively large standard deviations, mainly due to the variability of independent LLM-driven optimization trajectories. Each prompt undergoes 12 iterative updates, and differences in generated feedback and revisions can lead to different final prompts and performance.

2. Interpretation of results. To account for this variability, we evaluate independently optimized prompts across multiple runs, test engines, and OOD datasets. While small individual differences should be interpreted cautiously, TextReg achieves the highest or tied-highest mean accuracy compared with TextGrad and REVOLVE in 21 of 24 settings (Table 1), demonstrating a favorable overall trend.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

Dear Reviewers,

Thank you for your thorough and constructive reviews. We sincerely appreciate your valuable feedback, which has helped us significantly improve our paper. Below, we summarize the key concerns raised and our corresponding responses:

  1. Formalization and Implementation Reviewers DVq9, cdgN
    We have revised the Problem Formulation and Method to clarify the two factors of representational inefficiency—prompt capacity and rule specificity—and how they are operationalized through length monitoring and semantic edit analysis rather than direct numerical optimization. We also clarified the role of RuleBank and the ID–OOD trade-off that can arise when mitigating prompt overfitting.

  2. Technical Novelty Reviewers GDvC, cdgN
    We have clarified TextReg's distinctions from DLPO and REMO, highlighting its adaptive regularization at both the feedback and edit levels, rather than relying on fixed simplification or memory-based optimization.

  3. New Evidence for Prompt Overfitting Reviewers GDvC, DVq9, cdgN
    We have added held-out rule-applicability analysis and controlled rule-injection experiments, showing that narrowly applicable rules can harm OOD generalization even at matched prompt lengths. These findings further support jointly controlling prompt capacity and rule specificity.

  4. Additional Baselines and Generalization Reviewers GDvC, DVq9, cdgN
    We have introduced a compress-and-generalize baseline, showing that fixed simplification alone cannot match TextReg's OOD performance. We also clarified our benchmark selection and highlighted additional evaluations on GPQA-main and CommonGen, supporting broader applicability beyond structured reasoning.

  5. Statistical Reliability and Evaluation Fairness Reviewers GDvC, DVq9, cdgN
    We have added standard deviations, clarified sources of optimization variability, and further discussed the existing threshold sensitivity and compute-matched evaluations. TextReg achieves the highest or tied-highest mean accuracy compared with TextGrad and REVOLVE in 21 of 24 settings, supporting its overall OOD advantage.

Thank you again for your valuable feedback. We welcome any further questions or discussion.

Sincerely,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.