Pith. sign in

REVIEW 4 major objections 7 minor 7 cited by

Self-Supervised Prompt Optimization

T0 review · 4 major / 7 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read SPO claims prompt optimization can be driven solely by pairwise output comparisons, with no ground truth, no human feedback, and only three sampled questions.

desk verdict A cheap and clever reference-free prompt optimizer, but the 'SOTA' claim rests on a statistically nonexistent gap and the judge signal is never calibrated against correctness. read the letter →

arxiv 2502.06855 v3 pith:4P755FQ5 submitted 2025-02-07 cs.CL cs.AIcs.LG

Jinyu Xiang , Jiayi Zhang , Zhaoyang Yu , Xinbing Liang , Fengwei Teng , Jinhao Tu , Fashen Ren , Xiangru Tang
show 3 more authors
Sirui Hong Chenglin Wu Yuyu Luo
This is my paper · ORCID
classification cs.CLcs.AIcs.LG
keywords promptoptimizationself-supervisedlearningLLM-as-a-judgepairwisecomparisonreference-freeevaluationcost-efficientpromptingopen-endedtasks
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

Self-Supervised Prompt Optimization (SPO) tries to establish that prompt optimization can run without any external reference: no ground-truth answers, no human feedback, and no large validation set. Its evaluation and optimization signals come entirely from pairwise comparisons of LLM outputs on three sampled questions. If the claim holds, prompt search becomes practical for open-ended and low-resource tasks, at an optimization cost the paper puts at about $0.15 per dataset, 1.1% to 5.6% of the cost of existing methods. The paper reports that this reference-free loop matches or beats the tested baselines across five closed benchmarks and three open-ended MT-Bench tasks.

What carries the argument

The load-bearing object is the Output-versus-Output (OvO) evaluation framework, which draws the reward signal from a comparison between two model outputs instead of from a ground-truth answer. SPO uses OvO twice: the evaluation function asks an LLM to choose which of two responses better meets the task requirement, and the optimization function gives the current best prompt plus its outputs to an LLM that rewrites the prompt. The method's premise is that prompt quality shows up directly in output quality and that an LLM judge can recognize adherence to the requirement, so no external reference is needed to keep the loop moving.

What would settle it

Construct or find a task where an LLM judge systematically prefers fluent but wrong answers, run SPO on it, and check whether the selected prompt's accuracy on a held-out labeled set falls as judge preference rises across iterations.

Watch

Extended reading notes

Core claim

The central claim is that the relative quality of two prompts can be read off the relative quality of their outputs, so a prompt optimizer can hill-climb on preference comparisons alone. SPO instantiates that claim in an Optimize-Execute-Evaluate loop: an LLM optimizer rewrites the current best prompt, the execution model answers three questions with it, an LLM evaluator compares those answers against the previous best answer, and the winning output's prompt becomes the next candidate, with four randomized pairwise rounds to blunt position bias. The paper reports that prompts found this way exceed conventional prompting methods, reach top or near-top accuracy on GPQA, AGIEval-MATH, LIAR, WSC, and BBH-Navigate, and also improve open-ended writing, roleplay, and humanities outputs, including lifting a weaker execution model from 62.2 to 89.7 on BBH-Navigate.

Load-bearing premise

SPO assumes that an LLM judge's pairwise preference between two outputs, on just three questions, is a reliable reward for which prompt is genuinely better for the task.

Editorial extensions

If this is right

  • Prompt optimization no longer requires labeled answers, so tasks with no reference solution—open-ended writing, roleplay, product copy—become optimizable automatically.
  • At roughly $0.15 per dataset and three samples per iteration, prompt search becomes cheap enough to run on a single task without a training set.
  • Weaker execution models can gain substantially from SPO-found prompts; the paper reports Claude-3-Haiku rising from 62.2 to 89.7 on BBH-Navigate.
  • A single small model can play all three roles; GPT-4o-mini as optimizer, evaluator, and executor reaches 97.8 on BBH-Navigate, the best number in the ablation table.
  • The search peaks around ten iterations and degrades afterward, giving a practical early-stopping signal rather than requiring a separate validation metric.

Reading between the lines

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

  • An extension the paper leaves implicit is judge calibration: before deploying SPO, one could compare the judge's pairwise choices against a handful of human-labeled outputs and stop optimization if the agreement is low, preserving most of the cost savings.
  • Because the same model can optimize, evaluate, and execute, SPO could plausibly run as a closed-loop self-improvement procedure during inference, though the paper only tests offline iterations.
  • The four randomized rounds reduce order effects but not stylistic preferences such as length or confidence; a testable extension would hold output length fixed and measure whether SPO's gains persist.
  • The method's three-sample budget suggests the binding constraint is preference alignment rather than statistical sample size, so pairing SPO with a small calibrated reward model would be a direct next experiment.
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 / 7 minor

Summary. The paper proposes Self-Supervised Prompt Optimization (SPO), an iterative prompt optimization framework that replaces ground-truth-based evaluation with pairwise comparisons of LLM outputs by an LLM judge, and that uses the same outputs to guide an LLM optimizer. On five closed benchmarks (GPQA, AGIEval-MATH, LIAR, WSC, BBH-Navigate) with GPT-4o-mini execution, the method reports an average accuracy of 66.9 against 66.6 for OPRO, at roughly 1.1% to 5.6% of the optimization cost of the compared methods, using three sampled questions and ten iterations; open-ended MT-Bench tasks show improved win rates after SPO optimization. The paper also includes ablations on model combinations and a case study.

Significance. If validated, the paper's contribution would be practically significant: it lowers the cost of prompt optimization by orders of magnitude and removes the need for labeled validation sets, which is a bottleneck for many real-world applications. The manuscript ships code, reports optimization costs explicitly, and includes stability measures (Appendix A.4) and prompt-level case studies, which are strengths. However, the empirical support for the headline claim is currently thin: the average gain over OPRO is 0.3 points with no significance tests, and the core reward signal—pairwise LLM judgment—is never calibrated against objective correctness. The significance of the paper therefore depends on additional validation that the LLM judge's preferences track task quality.

major comments (4)
  1. [Section 3.2 and Algorithm 1 (lines 7–10)] The entire optimization loop is driven by phi_eval's pairwise choice, yet the paper never measures whether these choices agree with ground truth on the closed benchmarks that have it (GPQA, AGIEval-MATH, LIAR, WSC, BBH-Navigate). A systematic bias (e.g., verbosity, formatting, sycophancy) would propagate directly into the selected prompt, and the Limitations section's concession that "biases cannot be completely eliminated" does not quantify the effect. Please add a judge-agreement analysis on the closed benchmarks (or an ablation that replaces the judge with ground-truth labels and compares final prompts) to demonstrate that the reference-free signal is a valid reward for correctness.
  2. [Table 1 and Section 4.2] The "outperforms state-of-the-art" claim rests on a 66.9 versus 66.6 average over OPRO, a 0.3-point difference, with no significance test and no error bars for baselines. Moreover, on LIAR the best baseline (OPRO, 67.6) beats SPO (67.1), and on WSC the best baseline (PromptAgent, 82.7) beats SPO (82.0). Appendix A.4.1 reports standard deviations only for SPO and SPO*, and the reported STD values (0.29–1.15) are large enough that the 0.3-point average gap is within noise. Please provide paired bootstrap tests or confidence intervals for the Table-1 comparisons and report baseline standard deviations.
  3. [Section 4.2 (Ablation Study) and Appendix A.4] The number of samples (3) and iterations (10) were selected from ablations on BBH-Navigate, which is one of the five benchmarks in Table 1. This makes the BBH-Navigate result in the main table a post-selected number rather than a hold-out evaluation, and it biases the overall average toward the configuration tuned on that task. Please either pre-specify hyperparameters before seeing any benchmark, or perform the ablation on a separate hold-out task, and report the main results with the pre-specified configuration.
  4. [Section 4.1 (MT-Bench evaluation) and Section 3.1] The open-ended results use GPT-4o as the judge for win rates, while the optimization signal comes from GPT-4o-mini pairwise preferences. Since both are LLM preference judgments, the reported "significant improvement" may partly reflect alignment between the optimization target and the evaluation preference, not an improvement in task quality. Please add a human evaluation on a subset of MT-Bench (or an agreement analysis between GPT-4o judge, GPT-4o-mini judge, and human raters) to substantiate the open-ended claim.
minor comments (7)
  1. [Section 2.2, Eq. (3)] The formula for f_OVO includes G_i, although OvO is defined as not using ground truth; please remove G_i or clarify that it is a typographical leftover.
  2. [Appendix A.2 and A.5.2] The YAML configuration files include hand-written "requirements" and "faq" fields (e.g., BBH-navigate's requirements and three FAQ answers). The abstract's "without requiring external reference" should be qualified to "without ground-truth labels or human feedback," and the paper should state how much per-task human configuration is needed; an ablation with an empty or generic requirement string would help.
  3. [Algorithm 1] The same three questions are sampled once and reused across all N_max iterations; this fixed-sample overfitting is only acknowledged in Appendix A.4.3 and should be stated in the main text. Consider re-sampling questions each iteration.
  4. [Appendix A.3.2] The text states that baseline costs were "aligned" across methods, yet Table 1 reports costs differing by more than an order of magnitude; please clarify what was aligned (e.g., number of iterations, number of candidate prompts) and why the costs remain so different.
  5. [Table A3] Standard deviations are reported only for SPO/SPO*; please add standard deviations for the baseline methods to make the stability comparison fair.
  6. [Figure 2] Figure 2 is difficult to read in the current rendering (overlapping series, small labels); a per-dataset panel plot or table of cost-performance pairs would be clearer.
  7. [Throughout] Typos: "discoverd" (Section 4.3), "Claud-3.5-Sonnet" (Appendix A.5.2), and inconsistent "MT-bench"/"MT-Bench" capitalization.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: closed-task results are measured against external ground truth, and the open-ended judge-alignment concern is a validity limitation rather than a by-construction reduction.

full rationale

The paper's central closed-task claims are tested against external benchmark labels (GPQA, AGIEval-MATH, LIAR, WSC, BBH-Navigate), so the reported accuracies do not reduce to the optimization signal. Algorithm 1's acceptance rule is phi_eval's pairwise preference, but final closed-task performance is measured by ground-truth accuracy, an independent criterion. The cost-efficiency comparison is based on measured API costs against configured baselines, not on a fitted quantity renamed as a prediction. The paper cites several works by overlapping authors (e.g., AFlow, MetaGPT, Data Interpreter), but these appear in related-work context and are not load-bearing for the main claim. The open-ended MT-Bench results are judged by GPT-4o pairwise comparison, while the optimization signal is also an LLM pairwise preference (GPT-4o-mini); this is an evaluation-validity limitation, and the paper concedes judge bias in Limitations, but it is not a by-construction equivalence because the final judge differs from the optimization judge and because the closed-task results are externally verified. No equation reduces to its own input; no fitted parameter is renamed as a prediction; no uniqueness theorem is imported from the authors' prior work. Therefore no significant circularity is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's empirical claims rest on an LLM judge being a valid reward signal, on three sample questions being representative, and on ten iterations being enough; none of these is derived or calibrated. Hyperparameters are tuned on BBH-Navigate, which is also in the evaluation suite. No new physical or mathematical entities are introduced.

free parameters (5)
  • number_of_iterations = 10
    Chosen from ablation on BBH-Navigate (Appendix A.4.3); the same dataset appears in the main results (Table 1), so the main benchmark partly tunes the stopping rule.
  • sample_size = 3
    Chosen from ablation on BBH-Navigate (Appendix A.4.2, Figure A1); an inverted-U relationship is reported and 3 is selected as the optimal balance.
  • optimization_model = Claude-3.5-Sonnet-20240620 (default), GPT-4o variant also tested
    Model choice affects results: SPO with Claude scores 66.9 average while SPO* with GPT-4o scores 66.3.
  • evaluation_model = GPT-4o-mini-0718
    LLM judge used for pairwise selection; no calibration against ground truth is reported, and biases are acknowledged in Limitations.
  • task_requirements_hand_design = Per-task requirement strings and FAQ examples
    Each closed and open-ended task is given manually written requirements and sample templates (Appendix A.2, A.5.2); this is human-authored guidance that shapes optimization and partially blurs the 'no external reference' framing.
assumptions (5)
  • domain assumption LLM pairwise judgments reliably identify better outputs for task requirements.
    Section 3.2 and Limitations: SPO assumes GPT-4o-mini pairwise comparisons identify better outputs; the paper acknowledges possible bias but offers no calibration.
  • domain assumption Prompt quality manifests in LLM outputs.
    Section 3.2 first observation; the paper cites chain-of-thought and related work as evidence, but treats this as a premise rather than proving it for all tasks.
  • domain assumption Three randomly sampled questions are representative enough for optimization.
    Algorithm 1 samples three questions per iteration; overfitting to sample questions is acknowledged in Appendix A.4.3, so representativeness is assumed rather than guaranteed.
  • domain assumption Judge biases do not affect the overall optimization trend.
    Section 3.2 states that 'biases do not affect the overall optimization trend because eval's feedback merely serves as a reference'; this is an unproved claim about the optimization dynamics.
  • domain assumption The LLM optimizer can improve prompts from the best current prompt and its outputs.
    Section 3.1 and Appendix A.1: the optimizer relies on the model's internal task understanding to produce better prompts; no failure cases or limits are characterized beyond the overfitting note.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Prompt Optimization." pith.science (2026). https://pith.science/paper/4P755FQ5

@misc{pith2026250206855,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Prompt Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4P755FQ5}},
  note         = {Machine review of arXiv:2502.06855}
}
read the original abstract

Well-designed prompts are crucial for enhancing Large language models' (LLMs) reasoning capabilities while aligning their outputs with task requirements across diverse domains. However, manually designed prompts require expertise and iterative experimentation. While existing prompt optimization methods aim to automate this process, they rely heavily on external references such as ground truth or by humans, limiting their applicability in real-world scenarios where such data is unavailable or costly to obtain. To address this, we propose Self-Supervised Prompt Optimization (SPO), a cost-efficient framework that discovers effective prompts for both closed and open-ended tasks without requiring external reference. Motivated by the observations that prompt quality manifests directly in LLM outputs and LLMs can effectively assess adherence to task requirements, we derive evaluation and optimization signals purely from output comparisons. Specifically, SPO selects superior prompts through pairwise output comparisons evaluated by an LLM evaluator, followed by an LLM optimizer that aligns outputs with task requirements. Extensive experiments demonstrate that SPO outperforms state-of-the-art prompt optimization methods, achieving comparable or superior results with significantly lower costs (e.g., 1.1% to 5.6% of existing methods) and fewer samples (e.g., three samples). The code is available at https://github.com/FoundationAgents/SPO.

Figures

Figures reproduced from arXiv: 2502.06855 by the authors.

Figure 1
Figure 1. Comparison of Prompt Optimization Meth￾ods. (a) illustrates the traditional prompt optimization process with external reference, where feedback from the ground truth of humans is used to iteratively im￾prove the best prompt. (b) presents our proposed self￾supervised prompt optimization, which utilizes pairwise comparisons of LLM’s own outputs to optimize prompts without relying on external reference. and ensuring al… view at source ↗
Figure 2
Figure 2. Comparison of Performance (y-axis) and Optimization Costs in Dollars (x-axis) across Six Prompt Optimization Methods. SPO demonstrates competitive performance, consistently ranking among the top two methods while maintaining significantly lower costs (ranging from 1.1% to 5.6% of the costs incurred by other methods) across all datasets. 2024). Through these innovations, PO methods have shown promise in reducing manu… view at source ↗
Figure 3
Figure 3. Components of the Evaluation Framework for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: A Running Example of SPO Framework: Pairwise evaluation on the outputs selects the better one from corresponding prompts . The best output and prompt pair are highlited with pentagrams, which will be updated after evaluation. Furthermore, using a case from MT-bench, we…
Figure 5
Figure 5. Figure 5: Win rates comparison between different LLMs [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

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

  1. HarnessCompass: Guiding Automatic Harness Evolution toward Generalizable and Effective Agent Harnesses

    cs.LG 2026-08 conditional novelty 6.0 of 10

    A constrained, feedback-augmented, component-wise harness evolution loop improves coding agent Pass@1 on SWE-bench Verified and transfers to held-out tasks and another base model.

  2. Self-Supervised Skill Optimization

    cs.CL 2026-07 conditional novelty 6.0 of 10

    A comparative loop that extracts evidence for and against observed behaviors from judge preferences over unlabeled executions can optimize agent skill documents without any ground-truth labels or rewards.

  3. Synthetic Interaction Data for Scalable Personalization in Large Language Models

    cs.LG 2026-02 conditional novelty 6.0 of 10

    PersonaGym simulates noisy multi-turn user–assistant interactions to build PersonaAtlas, and PPOpt learns to rewrite user prompts from interaction history, improving judged personalization on synthetic benchmarks.

  4. ORPP: Self-Optimizing Role-playing Prompts to Enhance Language Model Capabilities

    cs.CL 2025-06 conditional novelty 6.0 of 10

    ORPP generates task-specific role-playing prompts through iterative reward-model-guided optimization on a small sample, then uses few-shot transfer to create prompts for new questions.

  5. Concise Reasoning, Big Gains: Pruning Long Reasoning Trace with Difficulty-Aware Prompting

    cs.AI 2025-05 conditional novelty 6.0 of 10

    Fine-tuning small models on difficulty-adapted, shortened reasoning traces (LiteCoT) yields equal or better benchmark accuracy than training on much longer traces, with far fewer tokens.

  6. Silence is Not Consensus: Disrupting Agreement Bias in Multi-Agent LLMs via Catfish Agent for Clinical Decision Making

    cs.CL 2025-05 conditional novelty 5.0 of 10

    A "catfish" agent that injects structured dissent into multi-agent LLM teams improves clinical question-answering accuracy by reducing premature consensus.

  7. Tournament of Prompts: Evolving LLM Instructions Through Structured Debates and Elo Ratings

    cs.AI 2025-05 conditional novelty 4.0 of 10

    DEEVO evolves better LLM prompts by debating outputs and selecting survivors with Elo ratings, without requiring labeled data or a hand-written fitness function.

Reference graph

Works this paper leans on

68 extracted references · 66 canonical work pages · cited by 7 Pith papers

  1. [1]

    Carefully read the entire question, identifying all relevant information and key concepts.,→

    work page

  2. [2]

    Choose the most appropriate problem-solving approach based on the question type.,→

    work page

  3. [3]

    State any relevant formulas, principles, or assumptions,→ b

    Solve the problem using these steps: a. State any relevant formulas, principles, or assumptions,→ b. Show all necessary calculations or conceptual analysis,→ c. Evaluate all answer options, explaining why incorrect options are wrong when relevant,→

    work page

  4. [4]

    Ensure that your response addresses all aspects of the question and demonstrates a clear problem-solving process

    Structure your response as follows: [Analysis] - Briefly state the main problem and key information (2-3 sentences max),→ [Solution] - Show your work step-by-step, including all relevant calculations and reasoning,→ - For conceptual questions, provide a clear, logical explanation,→ [Conclusion] - State the final answer in one clear sentence - Briefly expl...

    work page

  5. [5]

    Compare final position to starting point: - Calculate the distance from (0, 0)

    work page

  6. [6]

    Reasoning:

    Provide concise reasoning, labeled as "Reasoning:" - Explain key movements and their impact on position,→ - Justify your conclusion

    work page

  7. [7]

    Final Answer:

    State your final answer, labeled as "Final Answer:" Conclude your response with the answer in this XML format:,→ <answer>[Yes or No]</answer> Ensure your analysis adapts to all question types, handling both specific and ambiguous instructions. ,→ ,→ LIAR Prompt Analyze the given statement(s) carefully, following these steps for each question:,→

    work page

  8. [8]

    State the initial conditions: - Starting point: (0, 0) - Initial direction: positive x-axis (unless specified otherwise),→

    work page

Show all 68 references
  1. [9]

    Use a coordinate system: - x-axis: left (-) and right (+) - y-axis: backward (-) and forward (+)

  2. [10]

    Take X steps

    Analyze each step: - For ambiguous instructions (e.g., "Take X steps" without direction), assume forward movement,→ - Update coordinates after each step - Briefly explain any assumptions made

  3. [11]

    After analyzing all steps: - Summarize total movement in each direction - State the final position

  4. [15]

    Consider the statement, speaker 's background (if provided), and context.,→

  5. [16]

    Research and cite relevant facts and data related to the claim.,→

  6. [17]

    Evaluate the claim 's validity based on available evidence.,→

  7. [18]

    Consider potential biases or motivations of the speaker.,→ For each claim, structure your analysis as follows: CLAIM: [Restate the claim] CONTEXT: [Summarize speaker background and context] FACT-CHECK: - [Key fact 1 with source] - [Key fact 2 with source] - [Add more as needed...

  8. [19]

    Quickly identify key information

  9. [20]

    Devise fastest solution strategy

  10. [21]

    Execute minimal calculations

  11. [22]

    Manage time efficiently

    Double-check answer Rules: - Use shortcuts and estimation when possible - Skip detailed explanations - Omit units in final answer - Always present final answer as: \boxed{} Prioritize speed and accuracy. Manage time efficiently. Aim for shortest valid solution. Verify answer f...

  12. [23]

    Identify key elements and relationships

  13. [24]

    Consider multiple interpretations

  14. [25]

    Evaluate evidence for each possibility

  15. [26]

    Draw a logical conclusion

  16. [27]

    ""Please think step by step. Ensure the response concludes with the answer in the XML format:,→ <answer>[Yes or No]</answer>

    Summarize your reasoning. Conclude with: Final Answer: <answer>[Single letter option]</answer> A.4 Addtional Experiments A.4.1 Stability of SPO Check the results in Table A3. Table A3: Performance stability analysis of SPO. Stan- dard deviation values are reported for the main...

  17. [28]

    **Starting Point**: We begin at the starting point, facing forward.,→

  18. [29]

    Now we are 1 step to the right of the starting point.,→

    **Step 1**: Take 1 step right. Now we are 1 step to the right of the starting point.,→

  19. [30]

    From our current position (1 step right), moving 3 steps left means: ,→ ,→ - 1 step left brings us back to the starting point

    **Step 2**: Take 3 steps left. From our current position (1 step right), moving 3 steps left means: ,→ ,→ - 1 step left brings us back to the starting point. - 2 steps left takes us 1 step to the left of the starting point.,→ - 3 steps left takes us 2 steps to the left of the ...

  20. [31]

    "" Optimize Function’s Prompt PROMPT =

    **Step 3**: Take 2 steps right. From our current position (2 steps left), moving 2 steps right means: ,→ ,→ - 1 step right brings us back to 1 step left of the starting point.,→ - 2 steps right brings us back to the starting point.,→ After completing all the steps, we end up b...

  21. [32]

    **Initial Position**: We start at the origin, which we can define as (0, 0).,→

  22. [33]

    **Facing Direction**: We are facing forward, which we will assume is towards the positive y-axis.,→

  23. [34]

    - Moving right from (0, 0) means moving along the x-axis

    **Movement Instructions**: - **Step 1**: Take 1 step right. - Moving right from (0, 0) means moving along the x-axis. The new position is (1, 0).,→ - **Step 2**: Take 3 steps left. - Moving left means moving in the negative x-direction. From (1, 0), we move to (0, 0) after 1 s...

  24. [35]

    **Final Position**: After completing all the movements, we end up at (0, 0).,→

  25. [36]

    - Moved to (1, 0) after 1 step right

    **Summary of Journey**: - Started at (0, 0). - Moved to (1, 0) after 1 step right. - Moved to (0, 0) after 1 step left. - Moved to (-1, 0) after 2 steps left. - Moved to (-2, 0) after 3 steps left. - Moved to (-1, 0) after 1 step right. - Moved to (0, 0) after 2 steps right

  26. [37]

    "" By comparing the output results of the initial prompt at iteration 0 and the final prompt at itera- tion 10, the “iterative optimization of thought pro- cesses

    **Conclusion**: The final position (0, 0) is the same as the starting position (0, 0).,→ Final Answer: <answer>Yes</answer>""" By comparing the output results of the initial prompt at iteration 0 and the final prompt at itera- tion 10, the “iterative optimization of thought pr...

  27. [38]

    **Tech Bonanza: Up to 50% Off Top Brands on Black Friday!**,→

  28. [39]

    "",→ """ Performance After Optimization PROMPT =

    **Don 't Miss Out! Exclusive Black Friday Tech Deals for You**""",→ """ Performance After Optimization PROMPT = """Analyze the given advertising task and create compelling copy tailored to the specified platform(s) and industry. Consider current trends, competitor strategies, ...

  29. [40]

    ,→ ,→ ,→

    **Call to action**: The tweet asks followers to share their favorite way to take their coffee, encouraging them to participate and interact with the brand. ,→ ,→ ,→

  30. [41]

    **Incentive**: The promise of a free cup of coffee is a great incentive for people to engage with the tweet and share their thoughts. ,→ ,→

  31. [42]

    ,→ ,→ ,→

    **Hashtags**: The use of relevant hashtags like #coffee, #coffeeaddict, and #morningmotivation helps to reach a wider audience and attract new followers. ,→ ,→ ,→

  32. [43]

    Good morning

    **Personal touch**: The tweet starts with a friendly "Good morning" and addresses the audience as "coffee lovers", making it feel more personal and approachable. ,→ ,→ ,→ By responding to this tweet, the coffee brand can start a conversation with their followers, gather feedba...

  33. [44]

    Identify the target audience and platform (if specified).,→

  34. [45]

    State the content 's primary goal (e.g., awareness, engagement, promotion).,→

  35. [46]

    Choose an appropriate content format (e.g., text, image, video).,→

  36. [47]

    Develop a unique angle or concept related to current trends.,→

  37. [48]

    Craft a catchy headline or opening (max 10 words)

  38. [49]

    Write concise, platform-appropriate content (50-100 words).,→

  39. [50]

    Include 2-3 relevant hashtags

  40. [51]

    Add a clear call-to-action

  41. [52]

    "" ,→ ,→ ,→ ,→ ,→ ,→ Question =

    Suggest a visual element or enhancement. Provide one piece of social media content based on these guidelines, adapting to the given topic or brand. Focus on creating authentic, relatable content that resonates with the target audience and fits current social media trends. If n...

  42. [53]

    Deconstruct and reimagine the requested poetic form, pushing its structural limits while maintaining a semblance of its essence. ,→ ,→

  43. [54]

    ,→ ,→ ,→

    Approach the theme from an entirely unexpected angle, such as from the perspective of an inanimate object, a abstract concept, or a microscopic entity. ,→ ,→ ,→

  44. [55]

    ,→ ,→ ,→

    Employ surreal, avant-garde imagery that challenges the reader 's perception of reality while still conveying the core emotional truth of the theme. ,→ ,→ ,→

  45. [56]

    ,→ ,→ ,→

    Integrate unconventional elements such as mathematical symbols, scientific notation, or programming language to create a unique linguistic landscape. ,→ ,→ ,→

  46. [57]

    Experiment with typography, line breaks, and white space to create visual poetry that enhances the thematic exploration. ,→ ,→

  47. [58]

    Incorporate synesthesia-like sensory experiences, blending different senses in startling ways to evoke powerful emotional responses. ,→ ,→

  48. [59]

    "",→ Question =

    Conclude with an ending that simultaneously disorients and illuminates, leaving the reader with a profound sense of transformation. ,→ ,→ Before writing, imagine the most radical way to approach both the theme and the poetic form. Aim to create a piece that challenges the very...

  49. [60]

    You select the specific mRNA sequence that corresponds to the protein you want to create, which is the key to triggering an immune response against a particular virus or disease

    **Recipe selection**: The mRNA vaccine is like choosing a recipe book. You select the specific mRNA sequence that corresponds to the protein you want to create, which is the key to triggering an immune response against a particular virus or disease. ,→ ,→ ,→ ,→ ,→

  50. [61]

    The mRNA is prepared in a lab, and it 's like measuring out the right amount of each ingredient (amino acids) to create the protein

    **Ingredient preparation**: The mRNA sequence is like a set of instructions that tells the body how to make the specific protein. The mRNA is prepared in a lab, and it 's like measuring out the right amount of each ingredient (amino acids) to create the protein. ,→ ,→ ,→ ,→ ,→

  51. [62]

    In this case, the mRNA is packaged in a tiny particle called a lipid nanoparticle (LNP), which helps it reach the cells in the body

    **Delivery**: The mRNA is delivered to the body, just like you would add the ingredients to the dish. In this case, the mRNA is packaged in a tiny particle called a lipid nanoparticle (LNP), which helps it reach the cells in the body. ,→ ,→ ,→ ,→

  52. [63]

    The cells read the mRNA instructions and start making the specific protein

    **Cellular cooking**: Once the mRNA reaches the cells, it 's like adding the ingredients to the pot. The cells read the mRNA instructions and start making the specific protein. This protein is then displayed on the surface of the cells, like a flag, to alert the immune system ...

  53. [64]

    "" ,→ ,→ ,→ ,→ Performance After Optimization PROMPT =

    **Immune response**: The immune system responds to the protein by producing antibodies and activating immune cells, like a chef responding to a delicious aroma by adding the finishing touches to the dish. The immune system then remembers how to respond to this specific protein...

  54. [65]

    Think of it like a blueprint for your cells to follow

    **mRNA**: The recipe card that contains the instructions for making a specific protein. Think of it like a blueprint for your cells to follow. ,→ ,→

  55. [66]

    **Cells**: The sous chefs in our body that receive the mRNA recipe card and use it to make the protein. ,→ ,→

  56. [67]

    "" The pre-optimization explanation, while using a cooking analogy, was lengthy and lacked focus, such as “Imagine you’re a master chef

    **Protein**: The final dish that our cells produce, which helps our immune system recognize and fight off diseases. ,→ ,→ **Narrative:** When we receive an mRNA vaccine, our cells take the recipe card (mRNA) and use it to make the specific protein. This protein then triggers a...

    2022

  57. [2022]

    These techniques, developed through human insights and extensive experimentation, pro- vide essential optimization seeds for automated prompt optimization research

    which improved reasoning across various tasks, techniques for enhancing single-shot reason- ing (Deng et al., 2023; Zheng et al., 2024; Wang et al., 2024e), and methods for output format speci- fication (Zhang et al., 2024a; He et al., 2024; Tam et al., 2024). These techniques...

    2024

  58. [2023]

    is a dataset designed to evaluate the perfor- mance of language models on graduate-level ques- tions across multiple disciplines, including biology, physics, and chemistry. For our experiments, we utilized the GPQA-Diamond subset as the test set, while constructing our trainin...

    2023

  59. [2024]

    liar, liar pants on fire

    Code generation with alphacodium: From prompt engineering to flow engineering. CoRR, abs/2401.08500. Jon Saad-Falcon, Adrian Gamarra Lafuente, Shlok Natarajan, Nahum Maru, Hristo Todorov, Etash Guha, Estefany Kelly Buchanan, Mayee Chen, Neel Guha, Christopher Ré, and Azalia Mi...

    2024 arXiv

  60. [6512]

    differentiation

    ijcai.org. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elic- its reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824– 24837. Yifan Wu,...

    2024 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.