- 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 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. , , 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 = 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
[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, (Eq. 9).
- Scope: Assessed directionally by the Semantic Diff Analyzer (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 . 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.
| Metric | TextGrad | REVOLVE | TextReg |
|---|---|---|---|
| Directives per prompt | 19.6 | 7.2 | 7.0 |
| Narrow directives | 5.8 | 1.8 | 1.8 |
| Prompt tokens | 351 | 131 | 160 |
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.
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.
[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.
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:
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.
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.
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 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.