acceptodds
Under review as a conference paper at ICLR 2027

Reflective Prompt Tuning: From Diagnosing Recurring Failures to History-Aware Revision

Abstract

Large language models (LLMs) have become increasingly capable of following instructions and reasoning through complex problems, making prompting a flexible interface for adapting models without parameter updates. Yet prompt design remains labor-intensive and highly sensitive to formatting, phrasing, and instruction order, motivating automated prompt optimization methods that reduce this manual effort. Existing methods search over prompt candidates or revise prompts from feedback on individual examples or small batches, without maintaining a history of diagnosed failures across prompt versions. As a result, revisions are guided by local feedback rather than by which failures recur across examples and which persist or resolve across iterations. We propose Reflective Prompt Tuning (RPT), a diagnosis-driven prompt optimization framework that grounds each revision in a dataset-level view of the current prompt's recurring failures and in the optimization history. At each iteration, RPT evaluates the target model on the entire optimization set, critiques its incorrect responses, clusters the resulting critiques into recurring failure modes, and returns them with aggregate metrics as a structured report. An LLM optimizer then either revises the prompt, conditioned on this report and a memory of prior reports. RPT also optimizes for confidence calibration, using it in both its diagnostic reports and prompt selection. Across three reasoning tasks and four optimizer LLMs, we find that (i) RPT improves over initial prompts by up to 14.1 points and is the only method competitive on all three tasks, while also improving calibration and keeping prompts compact, (ii) full-set diagnosis, failure clustering, and memory each contribute, and memory lets the optimizer stop once performance plateaus, and (iii) revisions are aligned with the diagnosed failures, though weaker optimizers map distinct failures onto a few generic edits.

Primary area:
foundation or frontier models, including LLMs
Keywords:
prompt optimization, reflective prompt tuning, LLM function calling, diagnostic feedback, confidence calibration, prompt engineering

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_7dxoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Reflective Prompt Tuning (RPT), a diagnostic-driven framework for automated prompt optimization. RPT uses failure diagnosis on the full optimization set, failure clustering, and cross-iteration memory to drive prompt revision, and considers task score, calibration, and prompt length simultaneously when making the final selection. Experiments indicate that RPT offers advantages in both performance and prompt compactness across three tasks, with particularly notable results when using a GPT-5 optimizer. However, the current results do not allow for a fair disentanglement of RPT’s methodological contributions from the effects of higher computational budgets, different scaffolding, varying initial performance levels, or distinct prompt selection objectives.

Critical strengths

(1). The problem is practically significant; the issues of recurring failures and the utility of historical information in prompt optimization are indeed worthy of research. (2) . The methodological workflow of RPT is clearly articulated, and the interfaces between its modules are well-defined.

Critical weaknesses

(1). RPT’s diagnostic process involves extensive calls for target-model evaluation, critique, and clustering, whereas the baselines employ different optimization workflows and default budgets. (2). There are significant disparities in initial performance, making it impossible to attribute final performance gains solely to the optimization method. (3). The "w/o full-set" ablation removes both full-set coverage and clustering simultaneously, rendering the attribution of performance gains to individual components invalid. (4). RPT and GEPA-C utilize different protocols for calibration-based prompt selection. (5). RPT incurs high additional optimization costs; the actual token-based costs have not been sufficiently quantified.

Questions for authors

(1) Under a unified scaffold, an identical target-model call budget (or equivalent total budget), to what extent does RPT retain its advantage over ACE, GEPA, and MIPRO? (2) Is it possible to isolate and verify "full-set diagnosis" by varying only the failure sampling ratio, while keeping the clustering and reporting pipeline constant? (3) If GEPA-C were to employ the same calibration-aware selection objective as RPT, would RPT still maintain its advantage in terms of Brier score and ECE? (4) If the same historical prompts, metrics, and failure reports were incorporated into GEPA or other reflective baselines, would RPT still outperform them? (5) Are RPT’s results consistent across different target models, open-source models, and larger-scale tasks?

8 comments
AuthorsOpenReview ↗

Response to Reviewer 7dxo (1/4)

Thank you for taking the time to review our paper and for the constructive comments. Your comments led to new experiments and analyses, summarized below. Several of your weaknesses correspond to your questions, so we answer them together.

Summary of changes

  • Optimization budget and cost (Weaknesses 1 and 5, Question 1). We revised the cost analysis (new Appendix K) to report, for all four methods, target-model evaluations and other optimization calls, the target-model cost of reaching each selected prompt, and serving cost from provider-reported token usage. It also adds score-versus-budget curves and a held-out check on LiveBench-Math. Revised Section 4 now states that we do not claim an advantage at a matched call or dollar budget.
  • Full-set diagnosis (Weakness 3, Question 2). A new ablation, "50% failures, clustered", changes only the failure sampling ratio while keeping clustering and report generation fixed. The former "w/o full-set" row is renamed "sampled raw critiques" (Section 5.3, Table 2).
  • Calibration-aware selection (Weakness 4, Question 3). We applied RPT's selection objective to GEPA-C's saved candidate prompts on LiveBench-Math.
  • Other target models (Question 5). We added an open-weight target model, Qwen3.8-27B (new Appendix F).

Weaknesses 1 and 5, Question 1: Optimization budget, cost, and a unified scaffold

Thank you for raising this. We agree that the methods were not budget-matched: each ran in its own native configuration, with its own optimization budget, so final scores alone do not establish an advantage at equal cost. To address this thoroughly, we revised our Cost Analysis appendix (new Appendix K) to examine cost from several angles. The call counts, budget comparisons, target-model cost estimates, and serving-cost measurements all use the GPT-5-optimizer runs of all four methods, with GPT-4.1 as the target model:

  1. Call inventory for all four methods. For each task, we counted GPT-4.1 target evaluations (a candidate prompt applied to one training or development example) and all other optimization calls (critique, clustering, prompt revision).
  2. Token volume. We compared the input tokens that RPT and ACE provide over their full optimization runs.
  3. Score versus optimization budget. We plotted RPT's and ACE's development scores against cumulative target evaluations and against target-model dollar cost, and we compare the two methods at ACE's final budget.
  4. Cost to reach the selected prompt. For all four methods, we priced the GPT-4.1 target tokens used up to the checkpoint where each method's selected prompt first appeared.
  5. New held-out check on LiveBench-Math (732 queries). We evaluated six saved checkpoints (four from RPT, two from ACE) in one GPT-4.1 session on the LiveBench-Math test questions, with paired bootstrap intervals. This measures how test score changes with optimization budget.
  6. New serving-cost measurement (240 queries). For each task, we sampled 20 held-out questions and sent each one to GPT-4.1 once with each method's selected prompt, so all four methods answer the same questions (20 questions × 4 methods × 3 tasks). We computed the cost per response with and without prompt caching, with paired bootstrap intervals for the differences between methods.
  7. Prompt size and the length penalty. We compared selected-prompt sizes across all 12 task–optimizer settings, and we replayed RPT's selection without its length penalty.
  8. Validation of our token estimates. RPT's GPT-4.1 token counts during optimization are computed by tokenizing the requests in its saved trajectories. To check how accurate these estimates are, we applied the same counting to the RPT requests in the new LiveBench-Math run, where GPT-4.1 reported its actual usage. Each request is a single GPT-4.1 call answering one test question; the four RPT checkpoints give 481 such requests with reported usage. Our input counts were 0.64% below the reported total (a constant 11 tokens of message formatting per request), and output counts matched within 0.05%.

Main findings in the next response.

AuthorsOpenReview ↗

Response to Reviewer 7dxo (2/4)

Weaknesses 1 and 5, Question 1 (continued)

Main findings

RPT makes more calls, but its calls are much cheaper. RPT makes 5.6–8.6 times as many calls as ACE, but each ACE request carries its growing playbook, whereas RPT sends only its compact prompt. Over the full runs, RPT sent 29.2 / 17.8 / 76.9 million input tokens versus ACE's 168.7 / 25.5 / 224.8 million (HotPotQA / LiveBench-Math / Formula). As a result, reaching the selected prompt costs RPT far less in target-model spend than ACE:

GPT-4.1 target cost to reach selected prompt (USD, no caching)RPTACEGEPAMIPRO
HotPotQA13.20153.5914.96≥17.40
LiveBench-Math20.9639.782.83≥14.79
Formula42.2683.734.46≥11.03

RPT's cost is 0.09 / 0.53 / 0.50 times ACE's, and its development score at these checkpoints is within 2 points of ACE's on HotPotQA (+1.0) and Formula (−1.7). On HotPotQA, RPT is the cheapest of all four methods. GEPA has lower target cost on LiveBench-Math and Formula.

RPT finds its selected prompt early. RPT's selected prompts first appeared after only 34–38% of each run's total calls, so each run continued well past the point where it found the prompt it eventually selected. Currently, the optimizer LLM decides when to stop. An explicit stopping rule, such as stopping after a few iterations without improvement, could therefore reduce RPT's optimization cost substantially; testing such a rule remains future work.

Where RPT needs more budget. At ACE's final number of target evaluations, RPT matches ACE on HotPotQA (65.7% vs. 66.0% development score) but is still improving on Formula (73.3% vs. 88.3%) and LiveBench-Math (60.9% vs. 71.6%). On these tasks, RPT reaches its reported scores with more, but cheaper, target evaluations. This matches our new Appendix section "Performance After One Prompt Revision", where later iterations contribute most of the gain on these two tasks.

Held-out results on LiveBench-Math. In the new 732-query evaluation, the first checkpoints score alike (RPT 61.3% after 246 evaluations, ACE 62.2% after 260). RPT's iteration 6 reaches 72.8% after 1,476 evaluations, 4.4 points above ACE's selected prompt (68.4% after 512). RPT's own selected prompt scores 66.1%. All simultaneous 95% intervals for the RPT–ACE differences include zero, so neither method shows a reliable held-out advantage at these budgets.

RPT's prompts are much cheaper to serve. Across all 12 task–optimizer settings, ACE's selected prompt is at least 23 times larger than RPT's (in tokens). Values are USD per 1,000 responses, averaged over the queries after each method's first one; by then, the provider has cached the repeated prompt text and bills it at a lower rate:

Serving costRPTACEGEPAMIPRO
HotPotQA5.2775.585.539.72
LiveBench-Math10.2554.4911.1211.21
Formula4.6138.552.984.86
  • HotPotQA: RPT has both the highest test score (68.4%, vs. 66.6% ACE, 64.4% GEPA, 66.8% MIPRO) and the lowest serving cost. ACE costs 14.3 times as much (saving $70.31 per 1,000 responses; paired 95% interval [$53.63, $94.73]).
  • Formula: ACE scores 1.5 points higher (85.5% vs. 84.0%) but costs 8.4 times as much (saving $33.94; [$33.33, $34.63]). RPT scores 10.0 and 5.5 points above GEPA and MIPRO, at a serving cost between theirs.
  • LiveBench-Math: The table excludes two questions on which RPT's response hit GPT-4.1's 32,768-token output limit (ACE's also did on one). On the rest, RPT saves $44.24 over ACE ([$29.62, $62.70]), and its differences from GEPA and MIPRO have intervals including zero. Including the two cut-off responses, the saving over ACE is $22.19, with an interval that includes zero.
  • Without prompt caching, ACE costs 41, 20, and 8.7 times as much as RPT on HotPotQA, Formula, and LiveBench-Math.

The length penalty keeps prompts compact. With the penalty, RPT's selected Math and Formula prompts are 44% and 46% shorter than those selected without it. On HotPotQA, both rules select the same prompt.

Summary. RPT makes more calls than ACE, and on Formula and LiveBench-Math it needs more target evaluations to reach its reported scores. Because its prompts are compact, however, reaching its selected prompt costs less in target-model spend than ACE on all three tasks. With prompt caching, its serving cost is 8–14 times lower on HotPotQA and Formula, and 5.3 times lower on LiveBench-Math questions. Savings at serving time accumulate with every query. The saved records do not contain complete GPT-5 usage for every method, so we report target-model and serving costs rather than total bills and do not estimate a break-even point. We do not claim that RPT dominates at a matched call or dollar budget, and we revised Section 4 to state this explicitly.

AuthorsOpenReview ↗

Response to Reviewer 7dxo (3/4)

Weakness 2 (initial performance); Weakness 3 and Question 2 (full-set diagnosis)

Thank you for this suggestion. GEPA-C receives calibration diagnostics as auxiliary feedback but, unlike RPT, selects its final prompt by task score alone. To test whether RPT's calibration advantage comes from its final selection rule, we applied RPT's selection objective to GEPA-C's saved candidate prompts. We ran this test on LiveBench-Math, where GEPA-C's calibration differs most from RPT's (Figure 2), with the two strongest optimizers, GPT-5 and Gemini-3.1-Pro, and GPT-4.1 as the target model. Holding GEPA-C's generated candidates fixed, we selected among them with RPT's development-set selection objective (Eq. 2). Selection used only the development set; the test set was used only to evaluate the selected prompt. We compare the results with the RPT prompts reported in the paper for the same optimizers.

OptimizerGEPA-C candidate selected by RPT's objectiveGEPA-C test Brier ↓GEPA-C test ECE ↓RPT test Brier ↓RPT test ECE ↓
GPT-5Seed (candidate 0)0.5430.6180.1820.207
Gemini-3.1-ProCandidate 15 (unchanged)0.4750.5470.2440.309

Under GPT-5, RPT's objective selects GEPA-C's seed rather than its originally selected candidate, and GEPA-C's test Brier rises from 0.519 to 0.543. Under Gemini-3.1-Pro, it selects the same candidate that GEPA-C originally selected. In both settings, GEPA-C's Brier and ECE remain 1.8–3.0 times RPT's. Thus, applying RPT's selection objective to GEPA-C's candidates does not close the calibration gap observed in Figure 2, so in these settings RPT's calibration advantage is not explained by its final selection rule.

Question 4, incorporating RPT's history into GEPA or other reflective baselines.

Thank you for this question. RPT's history is not a separate input that can be passed to another optimizer. Each report pairs a prompt with metrics on the full optimization set and with failure modes clustered from critiques of all its failures (Eq. 5). The memory then orders these reports along a single revision trajectory, and the optimizer uses it both to rewrite the prompt and to decide when to stop. The baselines make different design choices at each of these points (Appendix T in revised version):

  • GEPA reflects on per-instance feedback from a minibatch and keeps a mutation only if its minibatch score improves. It also maintains a branching pool of candidates rather than one trajectory. Supplying RPT's reports would require evaluating every candidate on the full optimization set, replacing per-instance reflection with clustered diagnosis, and deciding which branch's history each mutation should see.
  • ACE reflects on one sample at a time and merges new entries into a playbook that it deliberately does not rewrite. Using dataset-level reports would require reflecting over the full set and letting the curator rewrite or remove entries, which runs against its incremental design.
  • MIPRO has no feedback stage that could receive the reports.

A baseline changed in all of these ways would no longer be GEPA or ACE. It would be RPT's diagnosis and memory inside a different search procedure, so RPT's contributions cannot be added to these methods without substantially changing them. We therefore tested whether these signals drive RPT's gains in two other ways. First, the ablations in Section 5.3 remove each signal from RPT in turn. Without memory (no earlier reports and no trajectory context δt\delta_t), task score falls on all three tasks, by 11 points on Formula, and the optimizer uses its full iteration budget on every task. Critiquing only sampled failures or only half of them lowers task score and worsens Brier score on every task. Second, we gave GEPA the parts of RPT that can be added without changing its search. GEPA-C receives calibration diagnostics as auxiliary feedback. RPT is still better calibrated on LiveBench-Math under every optimizer (Figure 2), and applying RPT's selection objective to GEPA-C's candidates does not close this gap (see our response to your previous question). Combining RPT's reports with GEPA's population search is an interesting direction for future work.

Reviewer_MsY5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Reflective Prompt Tuning (RPT), an LLM-based prompt optimizer. In each iteration, RPT runs the target model on the full optimization set, asks an LLM to critique every failure, and clusters the critiques into recurring failure modes. An optimizer LLM then reads this report together with the reports of earlier iterations, and either revises the prompt or stops. On HotPotQA, LiveBench-Math and Formula (GPT-4.1 as target, four optimizers), RPT is compared with ACE, GEPA and MIPRO, and with GPT-5 it has the best average score with prompts at least 23 times shorter than ACE's.

Critical strengths
  1. Under GPT-5, RPT reaches an average score of 74.9 compared with 71.2 for ACE, and its selected prompts contain 861–2,159 tokens compared with 53,819–106,320 for ACE (Table 6). Since the prompt is added to every query, this provides a meaningful deployment advantage. RPT also achieves the highest score on LiveBench-Math under all four optimizers.

  2. The paper reports where RPT underperforms, particularly on Formula, analyzes optimization costs, and finds that weaker optimizers map different failures to a few generic edits (Section 5.4).

Critical weaknesses
  1. UniPrompt (Task Facet Learning: A Structured Approach To Prompt Optimization) evaluates all training examples with the current prompt, collects feedback on errors, and, in one variant, clusters this feedback using a large language model (Section 3.1.1). It also provides the optimizer with previous edits and their accuracy changes (Section 3.1.2). These mechanisms closely resemble RPT's full-set critique clustering and use of optimization history. StraGo and PromptAgent also use failure feedback to guide prompt revisions, but none of these methods is cited or compared. RPT's distinguishing features are narrower: structured diagnostic reports, memory of previous failures and metrics, optimizer-controlled stopping, and calibration- and length-aware prompt selection.

  2. RPT achieves the highest average score only with GPT-5. ACE scores higher with the other three optimizers, outperforms RPT on Formula in all four settings, and RPT also trails GEPA on Formula in three settings. The 3.7-point average advantage over ACE under GPT-5 is driven primarily by LiveBench-Math (+10.6 points on a 122-question test set), while RPT trails ACE by 1.5 points on Formula. Nevertheless, the abstract and introduction describe RPT as "the only method competitive on all three tasks" without the GPT-5 qualification given in Table 1. Furthermore, all experiments use GPT-4.1 as the target model, and no repeated independent optimization runs are reported.

Questions for authors
  1. Table 4 shows that on HotPotQA, the first revised prompt (iteration 2) scores 68.2, close to the selected prompt's 68.4. Reporting results after a single revision across all tasks and optimizers would help quantify the additional benefit of continued optimization and clarify whether most improvements arise from the initial aggregated diagnosis or later revisions.

  2. In Table 2, the variant without full-set evaluation removes clustering as well as full-set coverage. Adding clustered diagnosis on a randomly sampled subset would isolate the contribution of full-set evaluation. The variant without memory also requires clarification regarding whether it retains the trajectory context (\delta_t) from Equation 5. If retained, this ablation compares complete historical reports with summarized historical statistics rather than removing all historical information.

  3. The critique model and optimizer use the same large language model, so changes in model capability affect both failure diagnosis and prompt revision. Fixing the critique model while varying the optimizer, and vice versa, would help determine which component explains the large performance differences on Formula, particularly whether stronger optimizers produce more accurate diagnoses or more effective revisions.

  4. Section 4 does not match optimization budgets, and Appendix I reports that RPT makes 5.6–8.6 times as many model calls as ACE under GPT-5, while comparable costs for GEPA and MIPRO are not provided. Reporting token consumption or monetary costs for all methods and comparing performance under matched budgets would clarify the cost-effectiveness of RPT. In addition, GEPA and MIPRO start 15–18 points below RPT on LiveBench-Math, partly due to their DSPy scaffolds (Appendix E.2).

5 comments
AuthorsOpenReview ↗

Response to Reviewer MsY5 (1/4)

Thank you for your very detailed and constructive review. Your specific suggestions shaped most of the new analyses in this revision. Below, we summarize the changes and then respond to each weakness and question.

Summary of changes

  • Related work and novelty (Weakness 1). We added UniPrompt, StraGo, and PromptAgent to Section 6, expanded Appendix T to compare their feedback, update, and memory mechanisms, and adjusted our novelty claims in the abstract and introduction.
  • Scope of claims (Weakness 2). The abstract, introduction, and Table 1 caption now state that RPT is competitive on all three tasks only with GPT-5 as the optimizer. We also added a second target model, Qwen3.8-27B (new Appendix F).
  • First revision versus later iterations (Question 1). New Appendix G reports test scores after one revision for all tasks and optimizers.
  • Ablations (Question 2). We added a "50% failures, clustered" condition, renamed the former "w/o full-set" row to "sampled raw critiques", and clarified that the no-memory variant also removes δt\delta_t (Section 5.3, Table 2).
  • Critic versus optimizer (Question 3). New Appendix U crosses GPT-5 and GPT-5-mini in the critique and optimizer roles on Formula.
  • Cost (Question 4). We revised Appendix K to report call counts, token volumes, and target-model and serving costs for all four methods, with budget curves and a held-out check on LiveBench-Math.

Weakness 1: Novelty relative to UniPrompt, StraGo, and PromptAgent Thank you for identifying these omissions. We have added UniPrompt, StraGo, and PromptAgent to Section 6 and expanded Appendix T to compare their feedback, update, and memory mechanisms. We have also revised the introduction and related work to acknowledge that UniPrompt’s feedback-based variant evaluates the full training set, clusters error feedback into edit batches, retains prior edits and accuracy changes, and uses validation-based early stopping; StraGo derives revision strategies from sampled successes and failures and caches scored prompts; and PromptAgent uses sampled-error feedback and prior prompt states in tree search.

RPT differs in the diagnostic information supplied for each revision. At every iteration, it evaluates the full optimization set. It returns a report linking the current prompt to recurring failure topics, their diagnosis counts and representative evidence, task score, Brier score, mean confidence, and aggregate metric changes relative to the previous and best iterations. The optimizer considers the current and earlier reports when revising or stopping. Final selection uses development-set task score, Brier score, and prompt length. We no longer present clustering, history, or early stopping as RPT-specific.

This is a methodological comparison. We have not evaluated UniPrompt, StraGo, or PromptAgent and make no empirical superiority claim over them, and limit primary performance comparisons to ACE, GEPA, and MIPRO.

Weakness 2: Qualified performance claims Thank you for identifying the scope of our claim. We agree with your reading of Table 1: RPT has the highest three-task mean only with GPT-5 as the optimizer. ACE has the highest mean with the other three optimizers and leads Formula in all four settings, and GEPA also exceeds RPT on Formula with those three. We revised the abstract and introduction to state that RPT's competitiveness on all three tasks holds with GPT-5 as the optimizer. The Table 1 caption makes the same qualification and acknowledges ACE's Formula lead.

Second target model. To test whether these findings depend on the target model, we added an experiment with Qwen3.8-27B as the target and GPT-5 as optimizer, on the same splits. The table gives initial → development-selected test scores (%); the best selected score for each task is in bold.

MethodHotPotQALiveBench-MathFormula
RPT58.0 → 71.678.2 → 83.677.5 → 84.0
ACE†58.6 → 69.270.3 → 73.476.0 → 87.0
GEPA50.6 → 70.674.3 → 77.877.0 → 86.0
MIPRO50.2 → 70.475.8 → 73.177.0 → 77.0

† ACE's runs stopped early as its playbook grew (on LiveBench-Math, requests reached Qwen's 65,536-token context limit), so we reported its best checkpoint before termination.

The pattern matches the GPT-4.1 results. RPT improves over its initial prompt on all three tasks and has the best HotPotQA and LiveBench-Math scores, while ACE again leads Formula. These runs also show a practical benefit of compact prompts: ACE's growing playbook ran into the context limit of this smaller target, which RPT's short prompts avoid. We added results with the new target model to the paper as well (new Appendix F, "Second Target Model: Qwen3.8-27B").

AuthorsOpenReview ↗

Response to Reviewer MsY5 (2/4)

Question 1: Benefit after the first revision Thank you for this suggestion. We evaluated each iteration-2 prompt (the prompt after one revision) on the held-out test split with GPT-4.1 as the target, for all three tasks and all four optimizers. These test scores were computed after optimization and were not used for revision or prompt selection. Each cell gives iteration 2 / selected prompt, with the selected iteration in parentheses.

TaskSeedGPT-5GPT-5-miniGemini-3.1-ProGemini-3.1-Flash-Lite
HotPotQA55.568.2 / 68.466.2 / 66.265.6 / 65.864.0 / 64.6
LiveBench-Math58.163.8 / 72.256.2 / 65.670.4 / 70.465.6 / 64.9
Formula72.369.0 / 84.026.0 / 74.074.0 / 74.072.5 / 73.5

HotPotQA. As you observed, most of the HotPotQA gain comes from the first revision, for all four optimizers: iteration 2 is 8.5–12.7 points above the seed, and the selected prompt adds less than 1 point. RPT thus reaches most of its gain early, which helps when the optimization budget is limited. Later iterations still improve calibration: with GPT-5, the Brier score falls from 0.287 after the first revision to 0.241 for the selected prompt.

LiveBench-Math and Formula. On these tasks, later iterations contribute more. With GPT-5, the selected prompt improves over iteration 2 by 8.4 points on LiveBench-Math (iteration 7) and by 15.0 points on Formula (iteration 11). With GPT-5-mini, it improves by 9.4 points on LiveBench-Math (iteration 18). Gemini-3.1-Pro converges early: its first revision is selected on LiveBench-Math, and its selected Formula prompt (iteration 4) has the same text as its first revision. With Gemini-3.1-Flash-Lite on LiveBench-Math, the first revision scores 0.7 points above the selected prompt on the test, a small difference given that each is a single evaluation.

Early regressions. A first revision can also lower performance, and development-set selection keeps such a revision from being returned. GPT-5-mini's first Formula revision asked for a second value (percentage, decimal, or unit) in the single answer field (e.g., 0.15 (15.0%)); Formula's single-value scorer marked all 127 such test answers (of 200) incorrect, which largely explains the 26.0. No revision exceeded the seed's development objective, so selection returned the seed. GPT-5-mini's first LiveBench-Math revision also scored below the seed (56.2 vs. 58.1); later iterations raised it to 65.6 (iteration 18).

We added this table and analysis as a new appendix section, Appendix G: "Performance After One Prompt Revision".

Question 2: Full-set ablation and trajectory context

Thank you for pointing this out. To isolate the coverage of failure diagnosis, we added a new condition, 50% failures, clustered. At each iteration, the target still answers every training example, so task and calibration metrics are computed on the full set. RPT then critiques a random half of the incorrect responses, resampled at each iteration, and clusters these critiques with ClusterFusion as usual. This condition therefore changes only how many failures are diagnosed, not how many training examples are evaluated or how critiques are aggregated.

The previous "w/o full-set" row passed the raw 10-20 critiques without clustering, so it changed both coverage and aggregation. We therefore renamed it "Sampled raw critiques". The table gives held-out task scores with GPT-5 as optimizer and GPT-4.1 as target. The RPT and sampled-raw-critiques rows are from Table 2.

VariantHotPotQALiveBench-MathFormula
RPT68.4 / .24172.2 / .18284.0 / .129
Sampled raw critiques67.2 / .27963.3 / .24176.5 / .178
50% failures, clustered (new)67.0 / .29558.1 / .34782.0 / .146

Results. Compared with RPT, critiquing only half of the failures lowers task score on all three tasks and raises Brier score. The largest drop is on LiveBench-Math, where no revised prompt beats the seed on the development set (58.1 vs. 72.2). On HotPotQA and Formula, the drops are smaller (1.4 and 2.0 points). The two reduced-coverage variants are not consistently ordered: the 50% condition scores higher on Formula (82.0 vs. 76.5), while "sampled raw critiques" scores higher on LiveBench-Math (63.3 vs. 58.1). Both fall below RPT on every task, in task score and in Brier score, which supports diagnosing the full set of failures. Each row is one optimization trajectory, so we present these comparisons as descriptive.

Memory. The no-memory variant removes both the earlier diagnostic reports and trajectory context δt\delta_t from Eq. 5. The optimizer receives only the current prompt, iteration number, current metrics, and current clustered failure topics. This ablation therefore removes all explicit history.

We updated Section 5.3 and Table 2 accordingly.

AuthorsOpenReview ↗

Response to Reviewer MsY5 (3/4)

Question 3: Critic quality versus optimizer revision quality

Thank you for this comment. To separate diagnosis from revision, we crossed GPT-5 and GPT-5-mini in the critique and optimizer roles on Formula, the task where the two optimizers differ most in Table 1 (84.0 vs. 74.0). We ran the two new mixed-role combinations and compared them with the two same-model runs from the paper. All four use the same splits and seed prompt with GPT-4.1 as the target, so only the critique and optimizer models change.

Critique modelOptimizer modelFormula test accuracy (%) ↑Brier ↓
GPT-5GPT-584.0.129
GPT-5-miniGPT-584.5.137
GPT-5GPT-5-mini73.0.265
GPT-5-miniGPT-5-mini74.0.256

The optimizer, not the critic, accounts for the Formula gap. With GPT-5 as optimizer, RPT reaches at least 84.0% with either critic, and changing the critic changes test accuracy by only 0.5 points. This is expected given the two roles: the critic diagnoses one failed response at a time, whereas the optimizer must combine the recurring failure patterns across the full optimization set with the history of earlier prompts and metrics, and decide which instructions to add, rewrite, or remove, or whether to stop. The cross also has a practical implication: a smaller critique model paired with a strong optimizer matched the all-GPT-5 accuracy, which could lower the cost of diagnosis.

We added the four-cell comparison to a new Appendix U, with a pointer in Section 5.1.

Question 4: Matched budgets

Thank you for raising this. We agree that the methods were not budget-matched: each ran in its own native configuration, with its own optimization budget, so final scores alone do not establish an advantage at equal cost. To address this thoroughly, we revised our Cost Analysis appendix (new Appendix K) to examine cost from several angles. The call counts, budget comparisons, target-model cost estimates, and serving-cost measurements all use the GPT-5-optimizer runs of all four methods, with GPT-4.1 as the target model:

  1. Call inventory for all four methods. For each task, we counted GPT-4.1 target evaluations (a candidate prompt applied to one training or development example) and all other optimization calls (critique, clustering, prompt revision).
  2. Token volume. We compared the input tokens that RPT and ACE provide over their full optimization runs.
  3. Score versus optimization budget. We plotted RPT's and ACE's development scores against cumulative target evaluations and against target-model dollar cost, and we compare the two methods at ACE's final budget.
  4. Cost to reach the selected prompt. For all four methods, we priced the GPT-4.1 target tokens used up to the checkpoint where each method's selected prompt first appeared.
  5. New held-out check on LiveBench-Math (732 queries). We evaluated six saved checkpoints (four from RPT, two from ACE) in one GPT-4.1 session on the LiveBench-Math test questions, with paired bootstrap intervals. This measures how test score changes with optimization budget.
  6. New serving-cost measurement (240 queries). For each task, we sampled 20 held-out questions and sent each one to GPT-4.1 once with each method's selected prompt, so all four methods answer the same questions (20 questions × 4 methods × 3 tasks). We computed the cost per response with and without prompt caching, with paired bootstrap intervals for the differences between methods.
  7. Prompt size and the length penalty. We compared selected-prompt sizes across all 12 task–optimizer settings, and we replayed RPT's selection without its length penalty.
  8. Validation of our token estimates. RPT's GPT-4.1 token counts during optimization are computed by tokenizing the requests in its saved trajectories. To check how accurate these estimates are, we applied the same counting to the RPT requests in the new LiveBench-Math run, where GPT-4.1 reported its actual usage. Each request is a single GPT-4.1 call answering one test question; the four RPT checkpoints give 481 such requests with reported usage. Our input counts were 0.64% below the reported total (a constant 11 tokens of message formatting per request), and output counts matched within 0.05%.

Main findings in the next response.

Reviewer_FC7FOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Reflective Prompt Tuning (RPT), a prompt optimization framework that diagnoses recurring failures across the full optimization set, clusters them into failure modes, and uses historical reports to guide prompt revisions. It jointly considers task performance, confidence calibration, and prompt length when selecting the final prompt. Experiments on HotPotQA, LiveBench-Math, and Formula show improvements over several prompt optimization baselines, with ablations supporting the contributions of full-set diagnosis, clustering, and memory.

Critical strengths
  • Dataset-level failure diagnosis with historical memory. RPT moves beyond feedback from individual examples by aggregating recurring failure modes and tracking them across iterations. The ablations support the value of these design choices, particularly on LiveBench-Math and Formula, where removing memory or full-set diagnosis substantially reduces performance.

  • Strong empirical results with compact prompts. RPT performs competitively across three distinct reasoning tasks and four optimizer models. With GPT-5, it achieves the best aggregate score among the compared methods. Its selected prompts are also substantially shorter than ACE's, addressing a practical deployment cost that accuracy-focused optimization often overlooks.

Critical weaknesses
  • Limited evaluation scope and generalization evidence. The experiments cover only three reasoning benchmarks, use GPT-4.1 as the target model, and rely on proprietary optimizer models. The results do not establish whether RPT generalizes to other target models, instruction-following tasks, or distribution shifts. This limits the strength of the broader claims about prompt optimization.

  • Baseline comparisons do not isolate the benefit of the method. The authors explicitly do not match optimization budgets across methods, even though RPT performs full-set diagnosis and incurs additional model calls. The reported results therefore do not establish whether its gains justify the additional optimization cost under comparable compute budgets. A more controlled comparison would strengthen the evidence for its practical advantage.

  • Important Baselines are missing. The paper only compares itself with 3 baselines with other important baselines missing such as OPRO, Learning from Contrastive Prompts (LCP).

Questions for authors

N/A

5 comments
AuthorsOpenReview ↗

Response to Reviewer FC7F (1/3)

Thank you for taking the time to review our paper and for your constructive feedback. Your review led to new experiments and revisions, which we summarize below before responding to each weakness.

  • Second target model (Weakness 1). We added an open-weight target model, Qwen3.8-27B, with GPT-5 as optimizer, on all three tasks. RPT again improves over its initial prompt on every task (new Appendix F, "Second Target Model: Qwen3.8-27B"). We also scope our claims to reasoning tasks with objectively scored answers.
  • Optimization cost (Weakness 2). We revised the cost analysis (Appendix K) to cover call and token counts for all four methods, development score as a function of optimization budget, the target-model cost of reaching each method's selected prompt, a new held-out check on LiveBench-Math, and a new serving-cost measurement. Section 4 now states explicitly that we do not claim RPT dominates at a matched call or dollar budget.
  • OPRO and LCP (Weakness 3). We added OPRO on all three tasks with the same optimizer, target, and splits. RPT scores higher on all three: by 15.6 points on LiveBench-Math and by smaller margins on HotPotQA and Formula. We also discuss OPRO and LCP in Related Work (Section 6).

Weakness 1: Evaluation scope and generalization

Thank you for this comment. We chose three reasoning tasks from different domains because RPT targets a specific setting: tasks whose failures have identifiable causes that recur across examples, so that diagnosing them over the full optimization set can guide revision. Each task tests a different kind of reasoning, with its own scoring rule and failure structure: multi-hop reasoning over textual evidence (HotPotQA, exact match), mathematical reasoning (LiveBench-Math, the benchmark's official scorers), and domain-specific numerical reasoning over financial data (Formula, numeric accuracy). All three have objectively scored answers, which makes verbalized confidence and its calibration well-defined. They also come from the baselines' own evaluations (HotPotQA and LiveBench-Math from GEPA, Formula from ACE), so the baselines are compared on tasks they have been evaluated on.

Across these tasks, RPT is useful in several distinct ways. It improves over its initial prompt on every task and optimizer, by up to 12.9, 14.1, and 11.7 points. Its diagnoses adapt to the domain rather than producing generic edits: the most persistent failures are surface-form errors on HotPotQA, arithmetic and semantic-alignment errors on LiveBench-Math, and formula-selection and cash-flow-timing errors on Formula, and the revisions that follow target them (Section 5.4, new Appendix G). RPT also lowers the Brier score in 11 of 12 settings, which matters when confidence is used for abstention or routing. Its prompts are at least 23 times shorter than ACE's, which makes them 8–14 times cheaper to serve on HotPotQA and Formula (Appendix K).

Per your suggestion, we expanded our experiment to another open-weight target model, Qwen3.8-27B, with GPT-5 as the optimizer for every method, on the same train, development, and test splits as the GPT-4.1(new Appendix F, "Second Target Model: Qwen3.8-27B"). The following table gives initial → development-selected test scores (%); the best selected score for each task is in bold.

MethodHotPotQALiveBench-MathFormula*
RPT58.0 → 71.678.2 → 83.677.5 → 84.0
ACE58.6 → 69.270.3 → 73.476.0 → 87.0
GEPA50.6 → 70.674.3 → 77.877.0 → 86.0
MIPRO50.2 → 70.475.8 → 73.177.0 → 77.0

The pattern matches the GPT-4.1 results. RPT improves over its initial prompt on all three tasks and has the best HotPotQA and LiveBench-Math scores, while ACE again leads Formula, with RPT 3.0 points behind. ACE's runs stopped early as its playbook grew (on LiveBench-Math, requests reached Qwen's 65,536-token context limit), so we reported its best checkpoint before termination. These runs also show a practical benefit of compact prompts: ACE's growing playbook ran into the context limit of this smaller target, which RPT's short prompts avoid. The results with the new target model are added to the paper as well.

For the optimizer role, we use frontier proprietary models from two families and at two scales, which lets us measure how optimizer strength affects each method (Section 5.1). This follows common practice in LLM-based prompt optimization: OPRO (Yang et al., 2024), Pryzant et al. (2023), and TextGrad (Yuksekgonul et al., 2025) also use proprietary models as optimizers. We agree that instruction-following tasks and distribution shift remain untested, and we scope our claims to reasoning tasks with objectively scored answers.

AuthorsOpenReview ↗

Response to Reviewer FC7F (2/3)

Weakness 2: Optimization budgets and cost

Thank you for raising this. We agree that the methods were not budget-matched: each ran in its own native configuration, with its own optimization budget, so final scores alone do not establish an advantage at equal cost. To address this thoroughly, we revised our Cost Analysis appendix (new Appendix K) to examine cost from several angles. The call counts, budget comparisons, target-model cost estimates, and serving-cost measurements all use the GPT-5-optimizer runs of all four methods, with GPT-4.1 as the target model:

  1. Call inventory for all four methods. For each task, we counted GPT-4.1 target evaluations (a candidate prompt applied to one training or development example) and all other optimization calls (critique, clustering, prompt revision).
  2. Token volume. We compared the input tokens that RPT and ACE provide over their full optimization runs.
  3. Score versus optimization budget. We plotted RPT's and ACE's development scores against cumulative target evaluations and against target-model dollar cost, and we compare the two methods at ACE's final budget.
  4. Cost to reach the selected prompt. For all four methods, we priced the GPT-4.1 target tokens used up to the checkpoint where each method's selected prompt first appeared.
  5. New held-out check on LiveBench-Math (732 queries). We evaluated six saved checkpoints (four from RPT, two from ACE) in one GPT-4.1 session on the LiveBench-Math test questions, with paired bootstrap intervals. This measures how test score changes with optimization budget.
  6. New serving-cost measurement (240 queries). For each task, we sampled 20 held-out questions and sent each one to GPT-4.1 once with each method's selected prompt, so all four methods answer the same questions (20 questions × 4 methods × 3 tasks). We computed the cost per response with and without prompt caching, with paired bootstrap intervals for the differences between methods.
  7. Prompt size and the length penalty. We compared selected-prompt sizes across all 12 task–optimizer settings, and we replayed RPT's selection without its length penalty.
  8. Validation of our token estimates. RPT's GPT-4.1 token counts during optimization are computed by tokenizing the requests in its saved trajectories. To check how accurate these estimates are, we applied the same counting to the RPT requests in the new LiveBench-Math run, where GPT-4.1 reported its actual usage. Each request is a single GPT-4.1 call answering one test question; the four RPT checkpoints give 481 such requests with reported usage. Our input counts were 0.64% below the reported total (a constant 11 tokens of message formatting per request), and output counts matched within 0.05%.

Main findings

RPT makes more calls, but its calls are much cheaper. RPT makes 5.6–8.6 times as many calls as ACE, but each ACE request carries its growing playbook, whereas RPT sends only its compact prompt. Over the full runs, RPT sent 29.2 / 17.8 / 76.9 million input tokens versus ACE's 168.7 / 25.5 / 224.8 million (HotPotQA / LiveBench-Math / Formula). As a result, reaching the selected prompt costs RPT far less in target-model spend than ACE:

GPT-4.1 target cost to reach selected prompt (USD, no caching)RPTACEGEPAMIPRO
HotPotQA13.20153.5914.96≥17.40
LiveBench-Math20.9639.782.83≥14.79
Formula42.2683.734.46≥11.03

RPT's cost is 0.09 / 0.53 / 0.50 times ACE's, and its development score at these checkpoints is within 2 points of ACE's on HotPotQA (+1.0) and Formula (−1.7). On HotPotQA, RPT is the cheapest of all four methods. GEPA has lower target cost on LiveBench-Math and Formula.

RPT finds its selected prompt early. RPT's selected prompts first appeared after only 34–38% of each run's total calls, so each run continued well past the point where it found the prompt it eventually selected. Currently, the optimizer LLM decides when to stop. An explicit stopping rule, such as stopping after a few iterations without improvement, could therefore reduce RPT's optimization cost substantially; testing such a rule remains future work.

Where RPT needs more budget. At ACE's final number of target evaluations, RPT matches ACE on HotPotQA (65.7% vs. 66.0% development score) but is still improving on Formula (73.3% vs. 88.3%) and LiveBench-Math (60.9% vs. 71.6%). On these tasks, RPT reaches its reported scores with more, but cheaper, target evaluations. This matches our new Appendix section "Performance After One Prompt Revision", where later iterations contribute most of the gain on these two tasks.

AuthorsOpenReview ↗

Response to Reviewer FC7F (3/3)

Weakness 2: Optimization budgets and cost (continued) Held-out results on LiveBench-Math. In the new 732-query evaluation, the first checkpoints score alike (RPT 61.3% after 246 evaluations, ACE 62.2% after 260). RPT's iteration 6 reaches 72.8% after 1,476 evaluations, 4.4 points above ACE's selected prompt (68.4% after 512). RPT's own selected prompt scores 66.1%. All simultaneous 95% intervals for the RPT–ACE differences include zero, so neither method shows a reliable held-out advantage at these budgets.

RPT's prompts are much cheaper to serve. Across all 12 task–optimizer settings, ACE's selected prompt is at least 23 times larger than RPT's (in tokens). Values are USD per 1,000 responses, averaged over the queries after each method's first one; by then, the provider has cached the repeated prompt text and bills it at a lower rate:

Serving costRPTACEGEPAMIPRO
HotPotQA5.2775.585.539.72
LiveBench-Math10.2554.4911.1211.21
Formula4.6138.552.984.86
  • HotPotQA: RPT has both the highest test score (68.4%, vs. 66.6% ACE, 64.4% GEPA, 66.8% MIPRO) and the lowest serving cost. ACE costs 14.3 times as much (saving $70.31 per 1,000 responses; paired 95% interval [$53.63, $94.73]).
  • Formula: ACE scores 1.5 points higher (85.5% vs. 84.0%) but costs 8.4 times as much (saving $33.94; [$33.33, $34.63]). RPT scores 10.0 and 5.5 points above GEPA and MIPRO, at a serving cost between theirs.
  • LiveBench-Math: The table excludes two questions on which RPT's response hit GPT-4.1's 32,768-token output limit (ACE's also did on one). On the rest, RPT saves $44.24 over ACE ([$29.62, $62.70]), and its differences from GEPA and MIPRO have intervals including zero. Including the two cut-off responses, the saving over ACE is $22.19, with an interval that includes zero.
  • Without prompt caching, ACE costs 41, 20, and 8.7 times as much as RPT on HotPotQA, Formula, and LiveBench-Math.

The length penalty keeps prompts compact. With the penalty, RPT's selected Math and Formula prompts are 44% and 46% shorter than those selected without it. On HotPotQA, both rules select the same prompt.

Summary. RPT makes more calls than ACE, and on Formula and LiveBench-Math it needs more target evaluations to reach its reported scores. Because its prompts are compact, however, reaching its selected prompt costs less in target-model spend than ACE on all three tasks. With prompt caching, its serving cost is 8–14 times lower on HotPotQA and Formula, and 5.3 times lower on LiveBench-Math questions. Savings at serving time accumulate with every query. The saved records do not contain complete GPT-5 usage for every method, so we report target-model and serving costs rather than total bills and do not estimate a break-even point. We do not claim that RPT dominates at a matched call or dollar budget, and we revised Section 4 to state this explicitly.

Weakness 3: Missing baselines (OPRO and LCP)

Thank you for pointing out these baselines. We added OPRO on all three tasks, with GPT-5 as the optimizer and GPT-4.1 as the target, on the data splits. We adapted the authors' released implementation to our tasks and kept OPRO's core loop: the optimizer sees a history of scored instructions and proposes eight new candidates per step. Starting from the same seed instruction as GEPA and MIPRO, we ran 20 steps and, as in OPRO, selected the instruction with the highest training score. The table reports test scores (%).

TaskOPRO initialOPRO selectedRPT selected
HotPotQA48.068.268.4
LiveBench-Math53.156.672.2
Formula74.083.084.0

OPRO improves on its own seed on every task, and RPT scores higher on all three. On HotPotQA and Formula, the differences are small (0.2 and 1.0 points, i.e., one and two test answers). On LiveBench-Math, RPT leads by 15.6 points. OPRO's optimizer sees only instruction–score pairs, whereas RPT's also sees which failures recur and how they change across prompt versions; LiveBench-Math is also the task where weakening RPT's diagnosis costs the most in our ablations (Section 5.3). OPRO does not elicit confidence, so it cannot be compared on calibration. We added the setup and results to a new Appendix M, "Additional Baseline: OPRO".

We also cite Learning from Contrastive Prompts (LCP) in the related work (Section 6). However, we could not include LCP as a baseline: the paper links no implementation, and key steps such as candidate generation and stopping are underspecified, so we make no empirical claim against LCP.

We hope these additions address your concerns, and we would be glad to answer any further questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reading of our paper and their constructive feedback. Their thoughtful suggestions led to new experiments and analyses that strengthened the paper and clarified its claims. Revised text in the paper is shown in blue. The main changes are:

  • Optimization cost and budget. We revised the cost analysis (Appendix K) to cover all four methods: recorded calls and token volumes, development score versus optimization budget, the target-model cost of reaching each selected prompt, a held-out remeasurement of LiveBench-Math checkpoints (732 queries), and serving cost measured from token usage (240 queries). RPT makes 5.6–8.6 times as many calls as ACE, but reaching its selected prompt costs 0.09–0.53 times ACE's target-model spend, and with prompt caching its prompts are 8–14 times cheaper to serve on HotPotQA and Formula. We do not claim an advantage at a matched call or dollar budget (more details in our responses to reviewers FC7F, MsY5, and 7dxo).
  • Second target model. We repeated the experiments with an open-weight target model, Qwen3.8-27B, with GPT-5 as optimizer (Appendix F). As with GPT-4.1, RPT improves over its initial prompt on all three tasks and has the best HotPotQA and LiveBench-Math scores, while ACE leads Formula. ACE's growing playbook also ran into this model's context limit, which RPT's compact prompts avoid (more details in our responses to reviewers FC7F, MsY5, and 7dxo).
  • Ablations of full-set diagnosis and memory. We added a "50% failures, clustered" condition that changes only how many failures are critiqued, keeping full-set scoring, clustering, and report generation fixed; it lowers task score and raises Brier score on all three tasks. We renamed the former "w/o full-set" row "sampled raw critiques" and clarified that the no-memory variant carries no explicit history (Section 5.3, Table 2; more details in our responses to reviewers MsY5 and 7dxo).
  • Additional baseline. We added OPRO on all three tasks (Appendix M). RPT scores higher on all three: by 15.6 points on LiveBench-Math and by 0.2 and 1.0 points on HotPotQA and Formula (more details in our response to reviewer FC7F).
  • Calibration-aware selection. Applying RPT's selection objective to GEPA-C's saved candidate prompts on LiveBench-Math does not close the calibration gap in Figure 2, so in these settings RPT's calibration advantage is not explained by its selection rule (more details in our response to reviewer 7dxo).
  • Where RPT's gains arise. We report test scores after RPT's first revision for every task and optimizer (Appendix G): most of the HotPotQA gain comes from the first revision, while later iterations contribute most on LiveBench-Math and Formula. On Formula, swapping GPT-5 and GPT-5-mini between the critique and optimizer roles shows that their performance gap comes from the optimizer, not the critic: with GPT-5 as optimizer, RPT reaches at least 84.0% with either critic, whereas with GPT-5-mini it reaches at most 74.0% with either (Appendix U; more details in our response to reviewer MsY5).
  • Related work and novelty. We added UniPrompt, StraGo, and PromptAgent to the introduction and related work, compared their feedback, update, and memory mechanisms with RPT's (Appendix T, Table 18), and refined how we position RPT's contributions relative to these methods (more details in our response to reviewer MsY5).

We sincerely thank all reviewers and the area chair for their time and effort, which have substantially improved our work. We remain available throughout the discussion period and would be glad to clarify any aspect of our responses.

Dear reviewers,

please review the authors’ rebuttal and provide your responses as soon as possible. Your timely feedback is important for moving the discussion forward.

Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.