What Should the Reflector See? An Empirical Study of Evidence in Reflective Prompt Optimization

Reflective prompt optimization improves instructions using task feedback, but coupling candidate generation with full-validation-based selection can make search costly and sequential. We propose a two-stage framework that decouples reflective candidate-pool construction from validation allocation. During construction, sampled parents first answer training questions, and groups of their saved responses guide parallel reflection and revision. Each child is screened on its group's questions against the corresponding parent scores. Improving, nonduplicate prompts enter the pool and can become parents without full validation. After construction, an exploration policy selects candidates from the frozen pool under a fixed full-validation budget, and the prompt with the highest measured validation score is returned. This separation allows candidate-pool size and validation budget to be set independently. We evaluate random selection and a multilayer perceptron with similarity-based exploration (MLP + similarity) on five benchmarks using Qwen3.5-9B and Llama-3.1-8B-Instruct. With 100 retained candidates and 20 full validations, our framework improves over the initial solver by up to 11.22 percentage points. MLP + similarity improves over the initial solver in all ten comparisons, although neither selection policy consistently outperforms the other. These results support separating reflective search from budgeted validation while leaving room for stronger candidate-selection policies.

Prompt optimization is an attractive alternative to reinforcement learning for adapting language-model agents, since it is cheap and produces human-readable artifacts. On factual tasks, however, it carries a failure mode that scalar-reward reinforcement learning does not. An optimizer that reads the training examples can write their answers into the prompt, so an in-distribution gain may be memorization rather than transferable behaviour. We apply reflective prompt evolution (GEPA) to the two system prompts of a multi-agent deep-research agent and evaluate it under a topic-disjoint split in which no topic is shared across train, validation, and test. Held-out citation reward rises by 0.079, from 0.702 to 0.781, with a bootstrap 95% confidence interval of [0.051, 0.107], and the gain replicates across three seeds (mean 0.107). Nearly half of the winning prompt is a memorized answer key, yet a controlled ablation shows this content earns no reward even on the exact training questions it encodes. The large validation-minus-held-out gap is therefore small-sample selection under a noisy judge, not memorization. Varying the verifier isolates what makes the gain transfer. Decomposed per-dimension feedback is the causal ingredient, and collapsing it to a single scalar reward makes held-out transfer negative. The gain persists under an out-of-family judge and is largely reproduced by a hand-written version of the discovered recipe. Credible evaluation of prompt optimizers therefore requires held-out topics, multiple seeds, and verifier ablations. We release the candidate prompts, lineage diffs, figures, tables, and analysis.

LLM agents are becoming central to information retrieval: they issue retrieval queries, synthesize answers, and increasingly serve as judges for IR evaluation. Improving the prompts that control these agents is an optimization problem, but in applied IR settings it often looks less like blind search and more like debugging. Engineers need to know which behavior failed, which nearby behavior still worked, what distinguishes the two, and whether a prompt edit improves held-out quality without introducing regressions. We present Contrastive Reflection, an iterative prompt-optimization framework for agentic IR workflows. The framework starts from a task-centric quality definition: QA agents expose retrieval or reasoning traces, and grading agents expose dimension-level scores and rationales. These structured traces are used to identify error-anchored behavioral slices, add nearby successful examples from the same region, and ask a Teacher LLM to propose a targeted prompt edit. Candidate edits are accepted only when validation performance improves, optionally subject to regression checks. We instantiate the framework with a tree-based slice selector, but the contribution is the contrastive reflection loop rather than the tree itself. On a public HotpotQA retrieval-augmented QA setup, one tree-selected contrastive repair improves held-out exact-match accuracy from 51.4% to 60.4%. Failure-only and random-evidence variants improve less and break more previously correct examples. A light instruction-only comparison places the method near modern prompt optimizers: MIPROv2 reaches 59.4% and GEPA 57.0%. The result is an interpretable optimization loop for IR agents, aimed at making prompt repair more inspectable and validation-driven.