pith. sign in

arxiv: 2605.19093 · v1 · pith:6NTBP2XDnew · submitted 2026-05-18 · 💻 cs.AI · cs.LG

Embedding by Elicitation: Dynamic Representations for Bayesian Optimization of System Prompts

Zhiyuan Jerry Lin , Benjamin Letham , Samuel Dooley , Maximilian Balandat , Eytan Bakshy This is my paper

Pith reviewed 2026-05-20 10:17 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords system prompt optimizationBayesian optimizationLLM elicitationembedding by elicitationaggregate feedbackGaussian process surrogateblack-box optimizationprompt engineering
0
0 comments X

The pith

An LLM can elicit and adapt a compact feature space to guide Bayesian optimization of system prompts using only aggregate scalar scores.

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

The paper introduces ReElicit to optimize system prompts when the only feedback available is a single scalar score per prompt rather than per-example labels or critiques. It frames the problem as sample-constrained black-box optimization over variable-length text and has an LLM generate a compact interpretable feature space from the task description and the history of past prompts and scores. A Gaussian process surrogate then models performance in that space and an acquisition function chooses promising feature vectors, which the LLM converts into new prompts. Re-eliciting the feature space after each batch of evaluations lets the representation change as more data arrives. A sympathetic reader would care because many real deployment settings provide only aggregate metrics such as overall accuracy or user satisfaction, making traditional prompt tuning difficult.

Core claim

ReElicit is a Bayesian optimization framework based on embedding by elicitation in which, given a task description and the history of evaluated prompts with scalar scores, an LLM elicits a compact interpretable feature space, maps prompts into it, and supports a Gaussian process surrogate and acquisition function that selects target feature vectors; the LLM then realizes those vectors as deployable system prompts, with the feature space being re-elicited as new evaluations arrive so that the representation adapts to the observed prompt-score history.

What carries the argument

Embedding by elicitation, in which an LLM generates a compact interpretable feature space from task description and evaluation history to enable Gaussian process surrogate modeling and acquisition over system prompts.

If this is right

  • The approach requires only one scalar score per prompt and no per-example labels, errors, or critiques.
  • Re-eliciting the feature space after new evaluations allows the representation to adapt dynamically to the observed history.
  • Across the ten tasks with a 30-evaluation budget ReElicit records the strongest aggregate performance among the tested aggregate-only baselines.
  • LLMs can function as adaptive semantic representation builders rather than only as prompt generators for optimization over natural-language artifacts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same elicitation mechanism could be tested on optimization of other natural-language artifacts such as model instructions or few-shot examples.
  • If the elicited features remain human-interpretable they might also help practitioners understand which prompt properties drive performance.
  • The method may extend to other black-box optimization domains where the search space is text and only aggregate feedback is available.

Load-bearing premise

An LLM given only a task description and the history of evaluated prompts with scalar scores can reliably produce a compact interpretable feature space whose dimensions capture the variations that matter for downstream Gaussian process modeling and acquisition.

What would settle it

Running the ten system prompt optimization tasks for a total of 30 evaluations each and finding that ReElicit does not achieve the strongest aggregate performance profile compared with the aggregate-only baselines would falsify the performance advantage.

Figures

Figures reproduced from arXiv: 2605.19093 by Benjamin Letham, Eytan Bakshy, Maximilian Balandat, Samuel Dooley, Zhiyuan Jerry Lin.

Figure 1
Figure 1. Figure 1: Feature stability and conservative adaptation. (a) Repeated extraction is highly stable: [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Predictiveness and actionability of the elicited feature space. (a) Dynamic features yield [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation convergence curves. Removing feature-gap refinement or replacing BO with [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Selected feature dimensionality over optimization rounds, reported as mean [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Best ℓ2 gap between the generated prompt’s extracted features and the BO-selected target over refinement steps. The trajectory shows that the refinement loop moves generated prompts closer to target feature vectors under the allotted refinement budget. Variant Pooled paired ∆ vs. ReElicit p No Refinement −0.009 ± 0.004 < 0.001 No BO −0.007 ± 0.003 < 0.001 Static Features −0.003 ± 0.004 0.168 Independent Ex… view at source ↗
read the original abstract

System prompts are a central control mechanism in modern AI systems, shaping behavior across conversations, tasks, and user populations. Yet they are difficult to tune when feedback is available only as aggregate metrics rather than per-example labels, failures, or critiques. We study this aggregate feedback setting as sample-constrained black-box optimization over discrete, variable-length text. We introduce ReElicit, a Bayesian optimization framework based on \emph{embedding by elicitation}. Given a task description, previously evaluated prompts, and scalar scores, an LLM elicits a compact, interpretable feature space and maps prompts into it. Leveraging a probabilistic Gaussian process surrogate, an acquisition function then selects target feature vectors, which the LLM realizes and refines into deployable system prompts. Re-eliciting the feature space as new evaluations arrive lets the representation adapt to the observed prompt-score history. We evaluate the setting using offline benchmark accuracy as a controlled aggregate proxy: the optimizer observes one scalar score per prompt and no per-example labels, errors, or critiques. Across ten system prompt optimization tasks with a 30 total evaluation budget, ReElicit achieves the strongest aggregate performance profile among representative aggregate-only prompt-optimization baselines. These results suggest that LLMs can serve as adaptive semantic representation builders, not only prompt generators, for Bayesian optimization over natural-language artifacts.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents ReElicit, a Bayesian optimization approach for system prompt tuning using only aggregate scalar feedback. An LLM is used to dynamically elicit a compact interpretable feature space based on the task and evaluation history, into which prompts are embedded. A Gaussian process models the prompt-score relationship in this space, and an acquisition function guides the selection of new feature vectors that are then converted back to prompts by the LLM. The feature space is re-elicited after each evaluation. On ten optimization tasks with a budget of 30 evaluations, ReElicit shows the best overall performance among compared aggregate-only methods.

Significance. Should the empirical results prove robust, this work highlights the potential of LLMs to construct task-adaptive representations for optimization over natural language, extending beyond their typical use in generation. This could impact automated design of AI system behaviors in settings with limited feedback.

major comments (2)
  1. The dynamic re-elicitation of the feature space after each new evaluation (as described in the method) risks introducing representational drift. Because the LLM may propose different dimensions or mappings in subsequent elicitations, prior prompts could be re-embedded to positions that alter their relative distances and correlations with scores. This could invalidate the GP posterior conditioned on earlier embeddings, especially under the tight 30-evaluation budget where the model is data-starved. The manuscript should provide analysis or ablations showing that the embeddings are sufficiently consistent to maintain GP coherence.
  2. The claim of strongest aggregate performance lacks supporting details on baseline implementations, the exact templates or prompts used for elicitation and realization, the number of independent runs, statistical significance tests, and variance measures. These omissions make it challenging to fully evaluate the reliability of the performance comparison.
minor comments (2)
  1. Consider adding a brief mention of the specific aggregate proxy used (offline benchmark accuracy) earlier in the abstract for clarity.
  2. Ensure consistent use of terms like 'feature space' and 'embedding' throughout to avoid potential confusion for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for their constructive feedback on our work. Below, we provide point-by-point responses to the major comments, and we outline the revisions we plan to make to address them.

read point-by-point responses

  1. Referee: The dynamic re-elicitation of the feature space after each new evaluation (as described in the method) risks introducing representational drift. Because the LLM may propose different dimensions or mappings in subsequent elicitations, prior prompts could be re-embedded to positions that alter their relative distances and correlations with scores. This could invalidate the GP posterior conditioned on earlier embeddings, especially under the tight 30-evaluation budget where the model is data-starved. The manuscript should provide analysis or ablations showing that the embeddings are sufficiently consistent to maintain GP coherence.

    Authors: We thank the referee for raising this valid point about potential representational drift. In our method, upon each new evaluation, the feature space is re-elicited and all previously evaluated prompts are re-embedded into the updated space before refitting the Gaussian process. This ensures that the surrogate model is always conditioned on a coherent representation of the full history under the current elicitation. While this introduces some variability in embeddings, it allows the representation to better capture the prompt-score relationships as data accumulates. To further address the referee's concern, we will add to the revised manuscript an ablation examining the consistency of re-embeddings, including metrics on how much prompt positions shift across elicitations and the resulting effect on acquisition function values and optimization trajectories. revision: yes

  2. Referee: The claim of strongest aggregate performance lacks supporting details on baseline implementations, the exact templates or prompts used for elicitation and realization, the number of independent runs, statistical significance tests, and variance measures. These omissions make it challenging to fully evaluate the reliability of the performance comparison.

    Authors: We agree that providing these details is essential for assessing the robustness of our results. In the revised version of the manuscript, we will add a dedicated subsection in the experiments detailing the baseline implementations, including the specific prompts and templates for elicitation and realization steps. Additionally, we will report the number of independent runs conducted, include error bars or variance measures, and perform statistical significance tests such as paired t-tests to compare ReElicit against baselines. These enhancements will improve the transparency and credibility of the performance claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method with independent evaluation

full rationale

The paper introduces ReElicit as an LLM-based dynamic embedding method for Bayesian optimization over system prompts under aggregate feedback. No mathematical derivations, equations, or first-principles results are presented that reduce any claimed prediction or performance to a fitted quantity or self-defined input by construction. The central claims rest on empirical results across ten tasks with a fixed 30-evaluation budget, compared against baselines. The LLM elicitation step is an explicit modeling assumption rather than a derived quantity, and no self-citation chains, uniqueness theorems, or ansatzes are invoked to force the representation or outcomes. The method is therefore self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that LLMs can produce useful feature spaces without additional training or fine-tuning.

axioms (1)
  • domain assumption LLMs can elicit compact, interpretable feature spaces from task descriptions and prompt-score histories that are suitable for Gaussian process surrogates
    This premise is invoked to justify the embedding-by-elicitation step and is not derived within the paper.

pith-pipeline@v0.9.0 · 5782 in / 1234 out tokens · 28202 ms · 2026-05-20T10:17:24.832179+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 5 internal anchors

  1. [1]

    and Daulton, Samuel and Letham, Benjamin and Wilson, Andrew Gordon and Bakshy, Eytan , booktitle =

    Balandat, Maximilian and Karrer, Brian and Jiang, Daniel R. and Daulton, Samuel and Letham, Benjamin and Wilson, Andrew Gordon and Bakshy, Eytan , booktitle =. 2020 , url =

    work page 2020

  2. [2]

    Journal of Artificial Intelligence Research , volume=

    Bayesian optimization in a billion dimensions via random embeddings , author=. Journal of Artificial Intelligence Research , volume=

    work page

  3. [3]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    A systematic survey of automatic prompt optimization techniques , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2025

  4. [4]

    Advances in neural information processing systems , volume=

    Re-examining linear embeddings for high-dimensional Bayesian optimization , author=. Advances in neural information processing systems , volume=

    work page

  5. [5]

    International conference on machine learning , pages=

    Similarity of neural network representations revisited , author=. International conference on machine learning , pages=. 2019 , organization=

    work page 2019

  6. [6]

    Artificial intelligence and statistics , pages=

    Deep kernel learning , author=. Artificial intelligence and statistics , pages=. 2016 , organization=

    work page 2016

  7. [7]

    Advances in neural information processing systems , volume=

    Local latent space bayesian optimization over structured inputs , author=. Advances in neural information processing systems , volume=

    work page

  8. [8]

    arXiv preprint arXiv:2311.02213 , year=

    Joint composite latent space bayesian optimization , author=. arXiv preprint arXiv:2311.02213 , year=

    work page arXiv

  9. [9]

    The eleventh international conference on learning representations , year=

    Large language models are human-level prompt engineers , author=. The eleventh international conference on learning representations , year=

    work page

  10. [10]

    The Twelfth International Conference on Learning Representations , year=

    Large language models as optimizers , author=. The Twelfth International Conference on Learning Representations , year=

    work page

  11. [11]

    TextGrad: Automatic "Differentiation" via Text

    Textgrad: Automatic" differentiation" via text , author=. arXiv preprint arXiv:2406.07496 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    gradient descent

    Automatic prompt optimization with “gradient descent” and beam search , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    work page 2023

  13. [13]

    Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution

    Promptbreeder: Self-referential self-improvement via prompt evolution , author=. arXiv preprint arXiv:2309.16797 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Optimizing instructions and demonstrations for multi-stage language model programs , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2024

  15. [15]

    LLM Prompt Duel Optimizer: Efficient Label-Free Prompt Optimization

    Llm prompt duel optimizer: Efficient label-free prompt optimization , author=. arXiv preprint arXiv:2510.13907 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    arXiv preprint arXiv:2507.03910 , year=

    Return of the latent space COWBOYS: Re-thinking the use of VAEs for Bayesian optimisation of structured spaces , author=. arXiv preprint arXiv:2507.03910 , year=

    work page arXiv

  17. [17]

    arXiv preprint arXiv:2306.03082 , year=

    Instructzero: Efficient instruction optimization for black-box large language models , author=. arXiv preprint arXiv:2306.03082 , year=

    work page arXiv

  18. [18]

    Applied Sciences , volume=

    Bayesian optimization for instruction generation , author=. Applied Sciences , volume=. 2024 , publisher=

    work page 2024

  19. [19]

    arXiv preprint arXiv:2510.04384 , year=

    LLM Based Bayesian Optimization for Prompt Search , author=. arXiv preprint arXiv:2510.04384 , year=

    work page arXiv

  20. [20]

    arXiv preprint arXiv:2512.15076 , year=

    An Exploratory Study of Bayesian Prompt Optimization for Test-Driven Code Generation with Large Language Models , author=. arXiv preprint arXiv:2512.15076 , year=

    work page arXiv

  21. [21]

    GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

    Gepa: Reflective prompt evolution can outperform reinforcement learning , author=. arXiv preprint arXiv:2507.19457 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    The Thirteenth International Conference on Learning Representations , year=

    Searching for optimal solutions with LLMs via bayesian optimization , author=. The Thirteenth International Conference on Learning Representations , year=

    work page

  23. [23]

    arXiv preprint arXiv:2503.21023 , year=

    Data mixture optimization: A multi-fidelity multi-scale bayesian framework , author=. arXiv preprint arXiv:2503.21023 , year=

    work page arXiv

  24. [24]

    AutoML 2025 ABCD Track , year=

    Ax: A platform for adaptive experimentation , author=. AutoML 2025 ABCD Track , year=

    work page 2025

  25. [25]

    Bayesian Analysis , year=

    Constrained Bayesian optimization with noisy experiments , author=. Bayesian Analysis , year=

    work page

  26. [26]

    Proceedings of the IEEE , volume=

    Taking the human out of the loop: A review of Bayesian optimization , author=. Proceedings of the IEEE , volume=. 2015 , publisher=

    work page 2015

  27. [27]

    Advances in Neural Information Processing Systems , volume=

    Combinatorial bayesian optimization using the graph cartesian product , author=. Advances in Neural Information Processing Systems , volume=

    work page

  28. [28]

    Advances in neural information processing systems , volume=

    Boss: Bayesian optimization over string spaces , author=. Advances in neural information processing systems , volume=

    work page

  29. [29]

    Advances in Neural Information Processing Systems , volume=

    GAUCHE: a library for Gaussian processes in chemistry , author=. Advances in Neural Information Processing Systems , volume=

    work page

  30. [30]

    Advances in neural information processing systems , volume=

    Combining latent space and structured kernels for Bayesian optimization over combinatorial spaces , author=. Advances in neural information processing systems , volume=

    work page

  31. [31]

    ACS central science , volume=

    Automatic chemical design using a data-driven continuous representation of molecules , author=. ACS central science , volume=. 2018 , publisher=

    work page 2018

  32. [32]

    arXiv preprint arXiv:2412.07820 , year=

    Hyperband-based Bayesian optimization for black-box prompt selection , author=. arXiv preprint arXiv:2412.07820 , year=

    work page arXiv

  33. [33]

    2021 , eprint=

    Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint=

    work page 2021

  34. [34]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Measuring Massive Multitask Language Understanding , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

    work page

  35. [35]

    Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them

    Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them , author=. arXiv preprint arXiv:2210.09261 , year=

    work page internal anchor Pith review Pith/arXiv arXiv