Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
PRESTO: Preimage-Informed Instruction Optimization for Prompting Black-Box LLMs
Authors: Jaewon Chu, Seunghun Lee, Hyunwoo J. Kim
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on 33 instruction optimization tasks demonstrate the superior performance of PRESTO. Code is available at https://github.com/mlvlab/PRESTO. |
| Researcher Affiliation | Academia | Jaewon Chu1 Seunghun Lee2 Hyunwoo J. Kim2 1Korea University, 2KAIST EMAIL EMAIL |
| Pseudocode | No | The paper describes the components and framework of PRESTO but does not present any formal pseudocode or algorithm blocks. Figure 2 provides a high-level diagram, not pseudocode. |
| Open Source Code | Yes | Code is available at https://github.com/mlvlab/PRESTO. |
| Open Datasets | Yes | We evaluate our proposed method, PRESTO, on 30 instruction induction tasks [36], a benchmark widely used to assess instruction optimization performance, and 3 arithmetic reasoning tasks [37 39]. We compare PRESTO with six competitive instruction optimization baselines: APE [8], Instruct Zero [14], INSTINCT [16], Evo Prompt [10], ZOPO [15], and OPRO [9]. We use LLa MA3.18B-Instruct [1] as the white-box LLM fw to generate candidate instructions, and GPT-4.1 as the black-box model fb. Following previous works [14 16], we set the total query budget to 165, initialize with 40 soft prompts, and evaluate all methods over three different random seeds. To ensure a fair comparison, we follow the hyperparameter tuning procedure in [16]. Detailed hyperparameter configurations and experimental settings are provided in the supplement. |
| Dataset Splits | Yes | Where Dval = {(xi, yi)}M i=1 is a validation set, and Ωdenotes the search space of instructions, typically a discrete sequence domain (e.g., natural language prompts or token sequences). ... Once the optimal soft prompt z is obtained, the corresponding instruction v is generated by the white-box LLM fw, i.e., v = fw(z , E) and subsequently evaluated on a held-out test set Dtest. ... We evaluate our proposed method, PRESTO, on 30 instruction induction tasks [36], a benchmark widely used to assess instruction optimization performance, and 3 arithmetic reasoning tasks [37 39]. |
| Hardware Specification | Yes | The supplementary materials include detailed specifications, including hardware type. |
| Software Dependencies | No | The paper mentions using specific LLMs as models (LLaMA3.1-8B-Instruct and GPT-4.1) but does not provide specific version numbers for ancillary software dependencies such as programming languages, libraries (e.g., PyTorch, TensorFlow), or other solvers. |
| Experiment Setup | Yes | Following previous works [14 16], we set the total query budget to 165, initialize with 40 soft prompts, and evaluate all methods over three different random seeds. To ensure a fair comparison, we follow the hyperparameter tuning procedure in [16]. Detailed hyperparameter configurations and experimental settings are provided in the supplement. |