Model Performance-Guided Evaluation Data Selection for Effective Prompt Optimization

Conference: ACL 2025
arXiv: 2505.10736
Code: None
Area: LLM Pre-training
Keywords: Prompt Optimization, Coreset Selection, Evaluation Data Selection, Real-Time Model Performance, Semantic Clustering

TL;DR

Proposed IPOMP, a two-stage evaluation data selection method. The first stage selects diverse samples through semantic clustering and boundary analysis, and the second stage iteratively replaces redundant samples using real-time model performance during the prompt optimization process. It improves prompt optimization performance by 1.6%-3.1% and stability by 50%+ on BIG-bench and LIAR, with less than 1% extra overhead.

Background & Motivation

Prompt Optimization is a key step in improving LLM task performance, but automated prompt optimization faces a neglected core problem: the selection of evaluation data.

Limitations of Prior Work:

Full evaluation is impractical: Evaluating each candidate prompt on the entire training set is too costly.

Random sampling is unreliable: Randomly selected subsets are likely unrepresentative, leading to unstable evaluations and sub-optimal prompts.

Existing coreset methods are inapplicable: - Semantic clustering performs poorly on highly similar task samples. - Model performance-based methods require pre-collecting evaluation data, which is expensive and has poor transferability. - No historical performance data is available for new or private datasets.

This is the first work to propose an evaluation data selection method specifically for prompt optimization scenarios.

Method

Overall Architecture

IPOMP is a two-stage method embedded into the iterative process of prompt optimization:

• Stage 1 (Diverse Sample Selection): Semantic clustering + boundary sample selection
• Stage 2 (Real-time Performance-Guided Refinement): Dynamically replacing redundant samples during each iteration.

Key Designs

1. Stage 1: Diverse Sample Selection:

   • All training samples are encoded by Sentence-BERT, followed by K-means clustering (\(k=5\)).
   • Proportional sampling of \(\alpha N\) samples from each cluster (semantic representativeness).
   • Find the furthest distance sample pairs in the semantic space to select \((1-\alpha)N\) samples (boundary diversity).
   • Use HNSW to accelerate boundary point detection, avoiding the full \(O(dN^2)\) distance calculation.
   • Design Motivation: Clustering ensures representativeness, while boundary samples compensate for coverage in extreme cases.

2. Stage 2: Real-time Performance-Guided Refinement:

   • Core Observation: Around 20% of the samples show highly correlated performance (>0.9) across different candidate prompts, indicating significant redundancy.
   • Record a performance matrix during each iteration: samples \(\times\) (number of output labels \(\times\) number of candidate prompts), using logits.
   • Hierarchical clustering groups highly correlated samples, randomly selecting a \(\beta\) ratio of redundant samples.
   • Replace these with the most dissimilar training samples in the semantic space.
   • Design Motivation: Leverage "free" model feedback during optimization to identify redundancy with zero extra inference overhead.

3. Key Differences from Existing Performance-Guided Methods:

   • Anchor-Point requires an extra warmup phase (~200 seconds + API cost).
   • Prediction-based requires training an evaluator on other LLMs.
   • IPOMP directly utilizes the byproducts of the optimization process, incurring zero extra inference cost.

Key Hyperparameters

• Evaluation set size \(N=20\), number of clusters \(K=5\)
• Cluster vs. boundary ratio \(\alpha=0.5\)
• Correlation threshold \(CT=0.9\), replacement rate \(\beta=0.5\)

Key Experimental Results

Main Results (Accuracy ± SD, averaged across three optimization methods)

Method	GPT-3.5 BIG-bench	GPT-4o-mini BIG-bench	GPT-3.5 LIAR	GPT-4o-mini LIAR
Random	0.719±0.035	0.704±0.041	0.742±0.041	0.748±0.048
Clustering	0.725±0.029	0.725±0.029	0.752±0.036	0.788±0.038
Boundary	0.727±0.040	0.723±0.038	0.770±0.035	0.797±0.047
Anchor-Point	0.745±0.028	0.756±0.027	0.801±0.027	0.807±0.024
Prediction-based	0.725±0.038	0.705±0.044	0.746±0.039	0.750±0.043
IPOMP	0.757±0.012	0.778±0.011	0.820±0.012	0.833±0.012

Ablation Study

Variant	GPT-3.5 BB Acc	GPT-4o-mini BB Acc
IPOMP (Full)	0.757±0.012	0.778±0.011
Stage 1 only	0.733±0.034	0.743±0.012
Random + Stage 2	0.737±0.021	0.738±0.012

Removing Stage 2: Acc drops by 2.4%, SD increases by 2.83x. Replacing Stage 1 with Random: Acc drops by 2%.

Time Overhead (BIG-bench, GPT-3.5, seconds)

Component	APO	APE	EVOPROMPT	Average
Stage 1	0.45	0.37	0.51	0.45
Stage 2	2.74	2.31	3.43	2.83
Anchor-Point Warmup	200.36	205.32	207.85	204.51

IPOMP extra overhead < 1%, while Anchor-Point warmup increases time by ~50%.

Key Findings

1. All coreset methods outperform Random, demonstrating that evaluation data selection is crucial for prompt optimization.
2. Compared to the second-best Anchor-Point, IPOMP achieves +1.6%~3.1% Acc gain and reducing SD by 50%+.
3. Stage 2 can serve as a plug-and-play plugin: bringing 2.3%/1.1%/1.5% gains to Random/Boundary/Clustering, respectively.
4. A sample size of 20 is the optimal trade-off point.
5. A single round of refinement can reduce highly correlated redundancy from 19% to 10%.

Highlights & Insights

1. Novel Problem Formulation: First to investigate evaluation data selection in prompt optimization.
2. Elegant Design of Real-Time Performance Utilization: Masterfully reuses existing model feedback from the optimization process.
3. Generality: Stage 2 can be applied plug-and-play to enhance any data selection method.
4. Improved Stability: Reducing SD by 50%+ signifies a substantial boost in reliability in production environments.

Limitations & Future Work

1. Limited Model Coverage: Only evaluated on GPT-3.5 and GPT-4o-mini.
2. Task Bias towards Classification: Performance on generation tasks remains unverified.
3. Logits Dependency: Limited applicability to APIs where logits cannot be obtained.
4. Prompt Optimization Techniques: Covers only APE, APO, and EVOPROMPT.

Related Work & Insights

• Three major categories of prompt optimization: non-directional (APE/EvoPrompt), directional (APO).
• Evolution of coreset selection from geometric methods to performance-based methods.
• Core insight: In prompt optimization, "what to evaluate" is as important as "how to optimize".

Rating

• Novelty: ⭐⭐⭐⭐ — First to define and address the evaluation data selection problem in prompt optimization.
• Experimental Thoroughness: ⭐⭐⭐⭐ — Evaluated across multiple datasets, models, and optimization techniques with detailed ablations.
• Writing Quality: ⭐⭐⭐⭐ — Clear motivation of the problem, systematic description of the method.
• Value: ⭐⭐⭐⭐ — High practical value for Stage 2 as a general plug-and-play plugin.

Related Papers

• [ACL 2025] Data Whisperer: Efficient Data Selection for Task-Specific LLM Fine-Tuning via Few-Shot In-Context Learning
• [ACL 2025] DavIR: Data Selection via Implicit Reward for Large Language Models
• [ICLR 2026] Predicting Training Re-evaluation Curves Enables Effective Data Curriculums
• [ICML 2025] LLM Data Selection and Utilization via Dynamic Bi-level Optimization
• [ICLR 2026] OptimSyn: Influence-Guided Rubrics Optimization for Synthetic Data Generation