Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Prompt Smart, Pay Less: Cost-Aware APO for Real-World Applications

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A hybrid of two prompt optimizers matches OPRO's accuracy at roughly 18% lower API cost.

desk verdict Useful real-world cost data, but the headline APE-OPRO-vs-OPRO parity rests on a modified OPRO baseline and no significance tests. read the letter →

arxiv 2507.15884 v1 pith:WU37FGKT submitted 2025-07-18 cs.LG

Jayesh Choudhari , Piyush Kumar Singh , Douglas McIlwraith , Snehal Nair This is my paper
classification cs.LG
keywords LLMautomaticpromptoptimizationAPE-OPROOPROAPEProTeGimulticlassclassificationcost-awareevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that automatic prompt optimization does not have to be expensive to work on real-world, high-stakes classification. It introduces APE-OPRO, a hybrid that runs APE-style prompt generation before OPRO-style meta-prompt refinement, and reports that this combination matches OPRO's test weighted F1 on a 2,500-product, ten-destination commercial taxonomy while cutting API cost by roughly 18%. The authors' deeper point is that the cost of an optimization trajectory is set early: OPRO's cold start produces long, highly detailed prompts immediately, while APE's warm start keeps prompts short, and every later iteration inherits that difference. The paper also finds that ProTeGi gives the best accuracy at lower API cost but with longer runtimes, that depth and breadth beyond modest settings bring diminishing returns, and that APE's accuracy shifts with label-list formatting even when the prompt says order does not matter. These findings matter for practitioners who pay per token and need a principled, repeatable way to choose among prompt optimizers.

What carries the argument

The machinery is the warm-start schedule: APE's semantic-similarity generator creates short variants of the initial system prompt, macro-F1 scoring on the training subset selects the top three, and those become the initial entries of the OPRO metaprompt that asks the optimizer to write a new, higher-scoring prompt. The paper's token-cost formula, which sums the optimizer's and scorer's input and output tokens in every iteration, is what converts this trajectory difference into the roughly 18% dollar saving.

What would settle it

Rerun the comparison with a larger training sample per label (for example, sixteen examples per label instead of four) and with bootstrapped random splits; the central claim is falsified if APE-OPRO no longer matches OPRO's test weighted F1, or if the 18% cost gap shrinks to within run-to-run variance. A cheaper, direct check is to plot token counts for iterations 1–3 of both methods: if OPRO's first-iteration prompt is not substantially longer, the proposed warm-start mechanism is not what saves the money.

Watch

Extended reading notes

Core claim

The paper's central claim is that APE-OPRO matches OPRO's performance while significantly reducing API costs by about 18% on average across ten destinations, with the gap reaching nearly 2x for Amsterdam. The mechanism is initialization: APE-OPRO begins with semantically similar variants of a base instruction, scores them on a tiny training set with macro F1, and feeds the top three into OPRO's metaprompt, so the optimizer sees short, structured exemplars from iteration one. Plain OPRO starts from a bare metaprompt and immediately writes long prompts with elaborate label definitions, so its token count spikes early and stays high. The paper also argues ProTeGi is the strongest absolute performer at lower API cost but longer execution time, and reports that increasing depth or breadth beyond 5 iterations or 5 prompts yields diminishing or negative returns, and that APE—but not the other methods—is sensitive to label formatting.

Load-bearing premise

The whole cost-performance comparison assumes that macro F1 on a training set of at most four examples per label (about 50–75 products per destination) reliably predicts which prompt will maximize weighted F1 on the test set; if a few idiosyncratic training examples dominate selection, both the reported rankings and the 18% figure are split artifacts.

Editorial extensions

If this is right

  • A team already running OPRO can expect the same weighted F1 with about 18% lower API spend by switching to the APE-OPRO initialization, and the saving is largest exactly where OPRO's first-iteration prompts are most verbose.
  • The default 10-iteration, 10-prompt configuration is probably over-spending: single-run ablations show 5 iterations and 5 prompts usually suffice, with further depth or breadth increasing cost 2–3x without consistent gains.
  • If ProTeGi's wall-clock time is acceptable, it should be the first choice on cost-performance grounds, since it posts the highest test weighted F1 while staying cheaper than OPRO and APE-OPRO.
  • Label formatting should be fixed before optimization starts, because APE's final prompts omit label definitions and are measurably affected by hyphen, number, or letter prefixes despite instructions to ignore position.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves untested whether the warm-start benefit transfers to other OPRO-style tasks; a fair test is to run OPRO with a single short, well-scored prompt as its first candidate and see if the cost gap closes.
  • The 18% figure is likely split-dependent: because macro F1 is computed on at most four examples per label, a bootstrap across random training samples would reveal whether the cost-performance ranking is stable or driven by a handful of idiosyncratic products.
  • APE's formatting sensitivity hints that, without label definitions, positional cues are the only disambiguating signal the model can use; a probe comparing APE with and without label definitions under identical formatting would separate formatting effects from definition effects.
  • The flat, destination-specific taxonomy is proprietary, but the cost-geometry claim could be checked on any category-rich public dataset with single-label multiclass structure.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper reports an empirical comparison of four automatic prompt optimization methods (APE, OPRO, ProTeGi, and the proposed APE-OPRO hybrid, plus a CoT baseline) on a proprietary single-label multiclass product classification task at Viator. Cost is computed from token counts and list prices using Eq. (2). The headline claim is that APE-OPRO matches OPRO's test weighted F1 while cutting API cost by ~18%; ProTeGi is reported as the strongest performer per dollar, and depth/breadth ablations plus label-formatting sensitivity analyses are presented.

Significance. If the headline claims are supported, the paper would be a useful practitioner-oriented contribution: it attacks a realistic multiclass setting rather than binary benchmarks, proposes a simple hybrid that is easy to implement, and models API cost transparently. The authors also provide five-run variance estimates for the main comparison and report label-formatting sensitivity, which is an interesting behavioral observation. However, the central comparison is weakened by the modified OPRO baseline and the lack of statistical tests, and the proprietary data/code are not released, so the quantitative conclusions are hard to audit and generalize.

major comments (4)
  1. [Section 4.5.3, Section 5.2, Figures 5-6] The OPRO baseline is not the published OPRO algorithm: exemplars are removed from the metaprompt and the history is truncated to the top 3 prompts instead of the original 20. Both changes reduce the cost of the baseline and are likely to reduce its performance, so the observed 'match' and the ~18% cost gap may be artifacts of comparing APE-OPRO against a weakened OPRO variant. A comparison with the original OPRO configuration, or ablations of the two modifications, is required to support the headline claim.
  2. [Abstract and Section 5.2] The statements 'OPRO incurs 18% higher costs' and 'APE-OPRO reduces API costs by ~18%' are arithmetically inconsistent: if OPRO's cost is 18% above APE-OPRO's, then APE-OPRO's cost is about 15.25% below OPRO's. The paper should report the cost ratio with a clear reference direction and use the resulting percentage consistently.
  3. [Section 5.3, Figure 7] The performance-parity claim rests on five runs per destination for only three destinations, and no significance or equivalence test is reported. With standard deviations of roughly 0.02-0.04, the observed mean differences between APE-OPRO and OPRO may be within noise; the authors should report paired differences, confidence intervals, or an equivalence test (e.g., TOST) with a pre-specified margin to substantiate 'matches performance'.
  4. [Sections 4.2-4.3 and 5.4] Prompt selection uses macro F1 on training sets of 48-75 examples with at most four examples per label. For rare labels, a single misclassification changes macro F1 substantially, so the selected prompt and the resulting cost ranking could be dominated by a few training examples. The paper should report bootstrap or repeated-split stability of the selection and of the APE-OPRO vs OPRO gap, or at least discuss this limitation explicitly in the main text.
minor comments (4)
  1. [Template placeholders] The manuscript retains ACM template placeholders such as the CCS Concepts 'Do Not Use This Code' line, the 2018 copyright notice, and the generic conference name; these should be replaced or removed.
  2. [Figures 18-21] Several captions in Appendix A.8 and A.9 contain duplicated axis labels (e.g., 'Depth Depth Depth') or placeholder text such as 'Caption'; these should be cleaned up.
  3. [Sections 5.5 and 5.6] The depth and breadth ablations use a single run per configuration, so the conclusion that 'smaller depths may suffice' is not backed by variance estimates; this should be stated as a preliminary observation rather than a firm finding.
  4. [Appendix A.6] There are two sections titled 'APE-OPRO best Prompt for Lisbon' with different prompts; the authors should clarify which run or configuration each prompt corresponds to, or merge them under distinct headings.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all quantitative claims are empirical measurements against external baselines; no derivation reduces to its own inputs.

full rationale

This paper makes no mathematical derivation that could be circular. Its central claims are empirical: 'APE-OPRO matches OPRO's performance while significantly reducing API costs by ~18%' is supported by measured macro-F1 selection on a training split followed by held-out weighted F1 evaluation, and by token-count cost accounting in Eq. (2) using list prices. The method definitions (APE, OPRO, ProTeGi, APE-OPRO) come from external prior work; the few citations to [6] and [11] are used for background or for a reported execution-time property, not as load-bearing justification of this paper's results. The 'hybrid' APE-OPRO is defined operationally (APE-style initialization, then OPRO-style metaprompt iterations) rather than derived from a premise that already contains the conclusion. The only concerns a reader might raise are baseline fairness (Section 4.5.3 modifies OPRO by removing exemplars and using top-3 instead of top-20 prompts) and the inconsistent phrasing of the 18% figure (Abstract: '18% improvement over OPRO'; Section 5.2: 'OPRO incurs 18% higher costs'); these are correctness and interpretation concerns, not circularity, because the cost and F1 numbers are measured, not constructed to equal each other. Accordingly no circular step can be quoted and no reduction by construction can be exhibited, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims depend on experimental hyperparameters chosen by the authors, on the assumption that token costs are additive, on macro F1 as a selection signal over a tiny training set, and on the representativeness of three chosen destinations. None of these are fitted to data to produce a derivation; they are evaluation choices that bound the generality of the findings.

free parameters (4)
  • optimization depth (iterations) = 10
    Set uniformly across APE, OPRO, APE-OPRO and ProTeGi; the cost comparison and convergence results depend on this choice, and depth 10 vs 5 or 15 changes total cost up to 3x.
  • breadth (prompts per iteration) = 10
    Set to 10 for all methods; increasing breadth to 15 roughly doubled or tripled cost with marginal performance gains, so the 18% cost figure would shift under different breadths.
  • top-k selected prompts = 3
    After scoring, the top 3 prompts are fed back to the optimizer (Algorithm 1); this choice controls diversity and affects convergence and cost.
  • ProTeGi mini-batch size D_mini = D_train/2
    ProTeGi's selection uses half the training data instead of the original D_mini=64; the paper notes this may increase variance.
assumptions (4)
  • domain assumption Token pricing from OpenAI is linear and additive as in Equation 2.
    The cost comparison assumes that input and output token costs can be summed without discounts, caching, or batching; the paper explicitly excludes caching/batch savings, which could change the 18% figure.
  • domain assumption Macro F1 on the small training set is a valid prompt-selection criterion.
    Prompts are chosen by macro F1 over at most 4 examples per label (Section 4.3); if this signal is noisy, the selected prompts may not generalize, affecting all downstream comparisons.
  • domain assumption Human labels from two annotators (Cohen's kappa=0.81) are ground truth.
    All weighted F1 scores treat the manually assigned labels as correct; labeling noise is not modeled.
  • domain assumption The three destinations Rome, Amsterdam, and Lisbon are representative of the full set of ten.
    Most detailed experiments use only these three, yet conclusions about variance and convergence are generalized to all destinations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prompt Smart, Pay Less: Cost-Aware APO for Real-World Applications." pith.science (2026). https://pith.science/paper/WU37FGKT

@misc{pith2026250715884,
  author       = {Pith},
  title        = {Pith review of: Prompt Smart, Pay Less: Cost-Aware APO for Real-World Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WU37FGKT}},
  note         = {Machine review of arXiv:2507.15884}
}
abstract

Prompt design is a critical factor in the effectiveness of Large Language Models (LLMs), yet remains largely heuristic, manual, and difficult to scale. This paper presents the first comprehensive evaluation of Automatic Prompt Optimization (APO) methods for real-world, high-stakes multiclass classification in a commercial setting, addressing a critical gap in the existing literature where most of the APO frameworks have been validated only on benchmark classification tasks of limited complexity. We introduce APE-OPRO, a novel hybrid framework that combines the complementary strengths of APE and OPRO, achieving notably better cost-efficiency, around $18\%$ improvement over OPRO, without sacrificing performance. We benchmark APE-OPRO alongside both gradient-free (APE, OPRO) and gradient-based (ProTeGi) methods on a dataset of ~2,500 labeled products. Our results highlight key trade-offs: ProTeGi offers the strongest absolute performance at lower API cost but higher computational time as noted in~\cite{protegi}, while APE-OPRO strikes a compelling balance between performance, API efficiency, and scalability. We further conduct ablation studies on depth and breadth hyperparameters, and reveal notable sensitivity to label formatting, indicating implicit sensitivity in LLM behavior. These findings provide actionable insights for implementing APO in commercial applications and establish a foundation for future research in multi-label, vision, and multimodal prompt optimization scenarios.

Figures

Figures reproduced from arXiv: 2507.15884 by the authors.

Figure 1
Figure 1. Comparison of evaluated methods based on Mean [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. figure 1 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Workflow for single-label classification of Viator products. Product text from Viator’s product description pages is [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (19 more)
Figure 3
Figure 3. Figure 3: Prompt Optimization Framework 3 Cost Computation We compute the cost of APO for each destination as follows: • 𝑁 be the total number of iterations, • 𝑀 be the number of prompts generated per iteration, • 𝐶 op in (𝑖) and 𝐶 op out(𝑖) be the input and output token costs f…
Figure 4
Figure 4. Figure 4: Distribution of labels in Top 10 Destinations [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Distribution of token count for the best prompt [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Evaluation of method generalizability across diverse [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Change in Performance for depth or maximum [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Change in Performance for breadth or maximum [PITH_FULL_IMAGE:figures/full_fig_p008_10.png]
Figure 12
Figure 12. Figure 12: Test weighted F1 (averaged over 5 runs) on Top 10 destinations [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]
Figure 13
Figure 13. Figure 13: Train weighted F1 (averaged over 5 runs) on Top 10 destinations [PITH_FULL_IMAGE:figures/full_fig_p010_13.png]
Figure 14
Figure 14. Figure 14: Confusion Matrix for Lisbon As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]
Figure 15
Figure 15. Figure 15: Effect of Label list formatting in prompt templates on OPRO, APE-OPRO, and ProTeGi [PITH_FULL_IMAGE:figures/full_fig_p012_15.png]
Figure 16
Figure 16. Figure 16: Variance in the train weighted F1 score for all methods, across three different destinations for five independent trials. [PITH_FULL_IMAGE:figures/full_fig_p031_16.png]
Figure 17
Figure 17. Figure 17: Evaluation of method generalizability across diverse destination datasets, showing the avg. weighted F1 score and [PITH_FULL_IMAGE:figures/full_fig_p032_17.png]
Figure 18
Figure 18. Figure 18: Change in Performance for depth or maximum number of iterations for each method across destinations along with [PITH_FULL_IMAGE:figures/full_fig_p033_18.png]
Figure 19
Figure 19. Figure 19: presents the training weighted F1 scores across three key destinations for all four methods, along with the associated costs as depth increases. Note that the cost values are identical to those reported in [PITH_FULL_IMAGE:figures/full_fig_p033_19.png]
Figure 20
Figure 20. Figure 20: Change in Performance for breadth or maximum number of prompts in each iteration for each method across [PITH_FULL_IMAGE:figures/full_fig_p034_20.png]
Figure 21
Figure 21. Figure 21: displays the training weighted F1 score and corresponding cost as breadth increases. The observed diminishing or negative returns in F1 performance suggest that lower breadth values are likely adequate for effective prompt optimization. $0.2 $0.4 $0.7 $2.4 $4.0 $6.5 $…
Figure 22
Figure 22. Figure 22: Convergence analysis for APE-OPRO on 3 key destinations [PITH_FULL_IMAGE:figures/full_fig_p035_22.png]
Figure 23
Figure 23. Figure 23: Convergence analysis for OPRO on 3 key destinations [PITH_FULL_IMAGE:figures/full_fig_p036_23.png]
Figure 24
Figure 24. Figure 24: Convergence analysis for APE for 3 key destinations [PITH_FULL_IMAGE:figures/full_fig_p037_24.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. One Mask to Rule Them All: On Hidden Facts after Editing and How to Find Them

    cs.LG 2026-04 conditional novelty 6.0 of 10

    ROME and MEMIT knowledge edits share a common weight subset isolable by a compact binary mask that reverses ~70–80% of edits and is necessary for editing success.

Reference graph

Works this paper leans on

34 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Jean-Yves Audibert and Sébastien Bubeck. 2010. Best arm identification in multi- armed bandits. In COLT-23th Conference on learning theory-2010 . 13–p

    work page 2010

  2. [2]

    Sarkar Snigdha Sarathi Das, Ryo Kamoi, Bo Pang, Yusen Zhang, Caiming Xiong, and Rui Zhang. 2024. GReaTer: Gradients over Reasoning Makes Smaller Lan- guage Models Strong Prompt Optimizers. arXiv preprint arXiv:2412.09722 (2024)

    work page Pith review arXiv 2024

  3. [3]

    Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. 2023. Promptbreeder: Self-referential self-improvement via prompt evolution. arXiv preprint arXiv:2309.16797 (2023)

    arXiv 2023

  4. [4]

    Han He, Qianchu Liu, Lei Xu, Chaitanya Shivade, Yi Zhang, Sundararajan Srini- vasan, and Katrin Kirchhoff. 2025. CriSPO: Multi-aspect critique-suggestion- guided automatic prompt optimization for text generation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 24014–24022

    work page 2025

  5. [5]

    Shanu Kumar, Akhila Yesantarao Venkata, Shubhanshu Khandelwal, Bishal Santra, Parag Agrawal, and Manish Gupta. 2024. SCULPT: Systematic Tuning of Long Prompts. arXiv preprint arXiv:2410.20788 (2024)

    work page Pith review arXiv 2024

  6. [6]

    Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, and Michael Zeng

  7. [7]

    Xinyu Tang, Xiaolei Wang, Wayne Xin Zhao, Siyuan Lu, Yaliang Li, and Ji-Rong Wen. 2025. Unleashing the potential of large language models as prompt opti- mizers: Analogical analysis with gradient-based model optimizers. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 25264–25272

    work page 2025

  8. [8]

    Zeliang Tong, Zhuojun Ding, and Wei Wei. 2025. EvoPrompt: Evolving Prompts for Enhanced Zero-Shot Named Entity Recognition with Large Language Models. In Proceedings of the 31st International Conference on Computational Linguistics . 5136–5153

    work page 2025

Show all 34 references
  1. [9]

    Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Haotian Luo, Jiayou Zhang, Nebojsa Jojic, Eric P Xing, and Zhiting Hu. 2023. Promptagent: Strategic planning with language models enables expert-level prompt optimization. arXiv preprint arXiv:2310.16427 (2023)

    2023 arXiv

  2. [10]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  3. [11]

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. 2023. Large language models as optimizers. arXiv preprint arXiv:2309.03409 (2023)

    2023 arXiv

  4. [12]

    Muchen Yang, Moxin Li, Yongle Li, Zijun Chen, Chongming Gao, Junqi Zhang, Yangyang Li, and Fuli Feng. 2024. Dual-Phase Accelerated Prompt Optimization. arXiv preprint arXiv:2406.13443 (2024)

    2024 arXiv

  5. [13]

    Qinyuan Ye, Maxamed Axmed, Reid Pryzant, and Fereshte Khani. 2023. Prompt engineering a prompt engineer. arXiv preprint arXiv:2311.05661 (2023)

    2023 arXiv

  6. [14]

    differentiation

    Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Zhi Huang, Carlos Guestrin, and James Zou. 2024. Textgrad: Automatic" differentiation" via text. arXiv preprint arXiv:2406.07496 (2024)

    2024 arXiv

  7. [15]

    Chenrui Zhang, Lin Liu, Jinpeng Wang, Chuyuan Wang, Xiao Sun, Hongyu Wang, and Mingchen Cai. 2023. Prefer: Prompt ensemble learning via feedback-reflect- refine. arXiv preprint arXiv:2308.12033 (2023)

    2023 arXiv

  8. [16]

    Lechen Zhang, Tolga Ergen, Lajanugen Logeswaran, Moontae Lee, and David Jurgens. 2024. SPRIG: Improving Large Language Model Performance by System Prompt Optimization. arXiv preprint arXiv:2410.14826 (2024)

    2024 arXiv

  9. [17]

    Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2022. Large language models are human-level prompt engineers. In The Eleventh International Conference on Learning Representations . Conference acronym ’XX, June 03–05, 2018...

    2022

  10. [19]

    label":

    label 3 Treat all labels as equally likely and independent of their position in the list. You must carefully read the product description and select the single most fitting label. A.3.3 Initial system prompt formatting with alphabets. Prompt Smart, Pay Less: Cost-Aware APO for...

    2018

  11. [20]

    - Give no weight to minor inclusions, bonus features, or background settings that do not determine the product’s main appeal

    **Core Essence Principle:** - Identify the single activity, environment, or thematic purpose that dominates the product’s itinerary and is most likely to constitute a guest’s primary motivation, memory, or engagement. - Give no weight to minor inclusions, bonus features, or ba...

  12. [21]

    Which activity do you observe them most immersed in over 70

    **Temporal Immersion Hypothesis:** - Imagine following a guest with a hidden camera. Which activity do you observe them most immersed in over 70

  13. [22]

    Trip Report

    **“Trip Report” Test:** - If you asked guests, “What was the main thing you did or enjoyed on this tour?” The answer they give—*in the singular*—reveals the correct label

  14. [23]

    Combo Experience Tour

    **Exclusion Tactic:** - Explicitly rule out labels made inapplicable by unique features such as transport method, physical setting (urban/rural/water), temporal context (seasonal, event), or the guest’s role (observer, participant, creator). - Reserve “Combo Experience Tour” a...

  15. [24]

    indispensability factor

    **Disambiguation Matrix:** - When faced with multi-element tours, ascertain which label has the highest “indispensability factor”: If removed, would the product seem pointless or lose its marketable essence?

  16. [25]

    extras

    **Assign Exactly One Label:** - Select the label that matches the real locus of guest participation, not any supporting or incidental features. — **Ultra-Precise, Disambiguated Label Definitions** - **Amsterdam City Exploration by Land** - *Defining Feature:* Guests traverse a...

    2018

  17. [26]

    Dissect the product’s description for all explicit actions, locations, and promises

  18. [27]

    memory snapshot

    Visualize the key “memory snapshot” a participant would treasure

  19. [28]

    Dominance Test

    Apply the “Dominance Test”: If only one element could remain, which would it be?

  20. [29]

    Golden Rules

    Use provided label definitions, especially their “Golden Rules” and “Unique Exclusions” to resolve ambiguous cases

  21. [30]

    star scene

    Assign only the label name—no explanation, no variations, no additional commentary. — **EXCLUSIVE LABEL DEFINITIONS (with Golden Rules & Unique Exclusions):** - **Aerial Scenic Viewing** - *Golden Rule*: Experience revolves around physically flying over landscapes/cityscapes (...

    2018

  22. [31]

    label":

    A sunset cruise for city views and snacks: "label": "City Sightseeing Cruises"

  23. [32]

    label":

    Multi-day featuring Sintra and a separate day in Fatima: "label": "Combo Experience Tour"

  24. [33]

    label":

    Guided tour centering on historic UNESCO castles: "label": "Iconic Sites Tour"

  25. [34]

    label":

    Sintra-focused palaces/gardens tour with a short coastal visit: "label": "Sintra Day Tours" Before outputting, carefully go through each step and strictly select the ONE most fitting label. Prompt Smart, Pay Less: Cost-Aware APO for Real-World Applications Conference acronym ’...

    2018

  26. [2023]

    gradient descent

    Automatic prompt optimization with" gradient descent" and beam search. arXiv preprint arXiv:2305.03495 (2023)

    2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.