Naive Prompt Optimization: Rethinking the Need for Complex Prompt Search

Yuan Chang; Xiaoqi Chen

Source-linked AI summary

Naive Prompt Optimization: Rethinking the Need for Complex Prompt Search

Yuan Chang, Xiaoqi Chen

arXiv:2608.27266v1Aug 27, 2026cs.AIcs.CL

TL;DR

Prompt optimization can improve agents without changing model weights, but recent methods increasingly rely on complex search. This paper introduces NPO, a single-lineage optimizer using teacher revisions and rollout feedback, and finds it broadly competitive with GEPA while using fewer rollouts and transferring gains across student models. The preliminary study also finds that stronger teachers improve NPO’s relative advantage, while GRPO helps on some games less amenable to prompt optimization.

Problem
Recent prompt optimization methods increasingly use sophisticated search, leaving open whether that complexity is necessary for efficiently improving agents across diverse tasks.
Method
NPO iteratively revises a prompt lineage with a teacher model using rollout traces and rewards, then compares prompt optimization methods and GRPO across benchmarks and interactive games.
Results
NPO achieves comparable or better performance than GEPA with fewer rollouts, benefits more from stronger teachers, remains broadly competitive across 22 TextArena games, and transfers gains across student models.
Takeaways & Limitations
Simple linear prompt optimization can rival substantially more complex search procedures, while stronger teacher reasoning and rich feedback can partially substitute for optimizer-side search complexity.
Takeaways & Limitations
The preliminary study uses a limited set of tasks and does not establish NPO’s relative effectiveness in more complex, long-horizon agent environments.

Abstract

from arXiv · show

Efficiently improving autonomous agents across diverse tasks is central to accelerating recursive self-improvement (RSI) in agentic AI, with prompt optimization emerging as a promising approach capable of delivering performance gains comparable to those achieved by fine-tuning model weights, while reducing computational costs in both optimization and serving. However, recent developments increasingly favor unnecessarily complex prompt optimizers. We introduce Naive Prompt Optimization (NPO), a lightweight single-lineage method that iteratively revises prompts using a teacher model with rollout feedback. NPO achieves comparable or better performance than GEPA with fewer rollouts, and its advantage increases with stronger teacher models, suggesting that stronger teacher reasoning can partially substitute for optimizer-side search complexity. In interactive games, NPO remains broadly competitive with GEPA, while GRPO performs better on some tasks less amenable to prompt optimization. We also show that NPO-optimized prompts elicit similar performance improvements when applied verbatim to other student models, especially across models within the same family. Overall, our preliminary results show that simple, linear prompt optimization can rival substantially more sophisticated and complex search procedures.

1 Introduction

The paper introduces NPO, a lightweight single-lineage prompt optimizer, to test whether complex prompt search is necessary. Across benchmarks and games, NPO is broadly competitive with GEPA, benefits more from stronger teachers, and transfers improvements across student models.

Motivation: Prompt optimization keeps model parameters fixed while improving instructions, enabling lightweight and portable adaptation through standard inference APIs.This contrasts with reinforcement learning methods that update model parameters from reward signals.
Contribution: NPO iteratively revises prompts with rollout traces and rewards through a teacher model, without maintaining multiple prompt lineages or explicit search algorithms.The paper compares this approach with prompt optimizers and GRPO across established benchmarks and interactive games.
Results: NPO’s advantage over GEPA widens as the teacher model becomes stronger.The result supports the paper’s hypothesis that stronger teacher reasoning can reduce the need for optimizer-side search complexity.
Results: NPO-optimized prompts can be applied directly to other student models and generally retain similar performance gains across model scales and families.The paper reports strongest transfer within model families, with effective but more variable transfer across families.
Results: NPO achieves comparable or better performance than GEPA while using fewer rollouts on IFBench and HotpotQA.IFBench evaluates instruction following under verifiable constraints, while HotpotQA is a multi-hop question-answering benchmark.
Results: Across 22 TextArena games, NPO and GEPA remain broadly comparable, while GRPO provides complementary gains on several tasks.These gains occur on games where prompt optimization is less effective.

2 Methodologies and Experiment Setup

The methodology centers on NPO’s linear prompt-revision loop and controlled comparisons against GEPA and GRPO. The experimental setup varies teacher and student models while reducing environmental and formatting noise.

NPO: NPO executes the student with the current prompt, collects minibatch rollout traces and rewards, and feeds a sliding window of recent iterations to the teacher.The teacher uses this rich context to produce the next prompt version.
NPO: Unlike OPRO, NPO revises prompts using complete rollout traces and rollout-specific rewards rather than only previous prompts and scalar scores.This provides trajectory-level feedback to the iterative prompt optimizer.
GEPA: GEPA maintains a pool of prompt candidates and uses reflection, acceptance testing, and Pareto-based parent selection to evolve them.Its prompt heritage forms a branching tree, unlike NPO’s linear evolution.
GRPO: GRPO updates a task-specific LoRA adapter using group-centered and standard-deviation-rescaled rollout rewards while keeping the default prompt and backbone fixed.In two-player environments, the target policy plays against a fixed unmodified base-model opponent.
Experiment Setup: Shared pseudorandomness reuses matched environment seeds across methods, reducing noise from differences in instance difficulty.Corresponding rollouts begin from the same randomly generated configuration, although later actions and trajectories may differ.
Experiment Setup: Constrained decoding requires the response format <think> reasoning </think> [action] and restricts actions to those legitimate under the current environment state.These controls separate decision-making performance from formatting errors.
Experimental Variations: The experiments vary teacher strength with Qwen3-8B students and test unchanged prompt transfer from Qwen3-8B or Llama-3.1-8B to larger and cross-family students.The transfer evaluation includes Qwen3-14B, Qwen3-32B, Llama-3.1-70B-Instruct, and Llama-3.3-70B-Instruct.
Experimental Variations: For IFBench and HotpotQA, NPO uses reflection minibatch/window settings of 50/10 and 40/20, respectively, with each revised prompt evaluated on a separate 300-example validation set.These settings provide richer feedback while remaining within the teacher model’s context window.

3 Results

NPO matches or exceeds GEPA with fewer rollouts, benefits more from stronger teachers, transfers across student models, and remains competitive across interactive games. Leakage checks find growing training-answer overlap but negligible validation-answer overlap, making evaluation-answer leakage unlikely to explain the gains.

3.1 Is NPO more Rollout-Efficient than GEPA Given Stronger Teachers?: NPO can achieve the same improved performance as GEPA with fewer rollouts, with the advantage widening as teacher models become stronger.On IFBench and HotpotQA, stronger teachers improve NPO more consistently than GEPA.
3.2 Do Optimized Prompts Transfer Across Student Models?: Prompts optimized on smaller student models produce substantial gains on larger models within the same family, while cross-family transfer is meaningful but weaker and more variable.Figure 5 covers within-family and cross-family transfers across tasks and optimization methods.
3.3 Comparison against GRPO: Across 22 TextArena games, NPO and GEPA show no universal winner, and NPO achieves gains similar to GRPO on several games.The comparison uses the same rollout budget across environments.
3.4 Do Optimized Prompts Leak Evaluation Answers?: NPO and GEPA prompts accumulate overlap with training answers, but validation-answer overlap remains negligible and mostly reflects shared answers across distinct questions.These checks suggest the observed gains are unlikely to result from direct exposure to evaluation answers.

4 Discussion and Conclusion

NPO is presented as a strong simple baseline: it remains broadly comparable to GEPA across benchmarks and games, transfers across models, and can outperform complex search with stronger teachers. The study also identifies safety, privacy, task-scope, and long-horizon limitations, alongside directions for consolidating task-specific prompts.

Discussion and Conclusion: NPO remains broadly comparable to GEPA on IFBench, HotpotQA, and 22 TextArena games, while GRPO provides complementary gains on some games.The paper frames prompt optimization and parameter optimization as task-dependent alternatives rather than universally ordered methods.
Discussion and Conclusion: Prompts optimized on one student model can transfer without re-optimization, with same-family transfer generally preserving gains better than cross-family transfer.Transfer performance varies across models, tasks, and optimization methods.
Discussion and Conclusion: NPO’s simple iterative design is intended as a baseline for testing whether sophisticated prompt search is necessary.Its use of stronger teachers and rich feedback is reported to partially substitute for optimizer-side search complexity.
Responsible Use: NPO can expose rollout traces to untrusted teacher infrastructure, so masking and sanitization are required for sensitive tasks.The responsible-use statement also calls for preserving and explicitly verifying teacher safety guardrails.
Limitations: The preliminary study leaves open whether NPO’s advantage extends to complex long-horizon environments and does not establish prompt optimization versus RL effectiveness there.The authors also note limited task coverage, omitted frontier-scale closed students, RL instability, and task-dependent algorithm or reward preferences.
Future Directions: Future work could consolidate independently optimized prompts into compact multi-task prompts before additional NPO refinement.This is proposed as a way to preserve task-specific gains after consolidation.

A Detailed Optimization Trajectories in Cross-Student Transfer Experiments

Figure 8 organizes cross-student transfer optimization trajectories by dataset–method combination and teacher model. Its curves track how successive NPO prompts or GEPA candidates perform on larger and cross-family student models.

Optimization Trajectories: Each curve tracks transfer from successive NPO iterations or GEPA pool candidates to larger and cross-family student models.The trajectories show performance changes as optimized prompt versions or candidate prompts are evaluated on other students.
Optimization Trajectories: Columns represent dataset–optimization-method combinations, while rows represent teacher models in the cross-student transfer experiments.This layout enables comparisons across both optimization settings and teacher choices.

B Heatmap for the evaluated games in TextArena

Figure 9 compares unoptimized performance, GRPO, prompt optimization, and cross-student transfer across 22 TextArena games. Same-family Qwen transfer largely preserves gains, whereas cross-family transfer is less consistent.

Heatmap Layout: The heatmap reports average performance for each game–method combination across six sections.Win rates are displayed as decimals.
Direct Evaluation: Sections 1–3 compare the unoptimized Qwen3-8B baseline, GRPO, and held-out NPO/GEPA results using two teacher models.The teacher models are DeepSeek-V4-Flash-preview-0424 and GPT-5.5.
Cross-Student Transfer: Sections 4–6 apply prompts optimized on Qwen3-8B unchanged to Llama-3.3-70B-Instruct, Qwen3-14B, and Qwen3-32B.The transfer comparison separates same-family Qwen models from the cross-family Llama model.
Cross-Student Transfer: Same-family Qwen transfer largely preserves performance gains, while cross-family transfer to Llama-3.3-70B-Instruct is less consistent.This is the main comparison to make across the transfer sections.

C Extended Analysis of GRPO Group Size

Figure 10 examines how GRPO rollout group size affects training quality on TextArena games under a fixed setup. Results are shown for Minesweeper, with other environments reported to behave similarly.

Experimental Setup: The experiment varies GRPO rollout group size after supervised fine-tuning Qwen3-8B on GPT-4o-mini responses.The SFT responses typically contain about 300 tokens, and the GRPO setup is otherwise fixed.
Results: Figure 10 reports GRPO training performance under different rollout group sizes for the Minesweeper environment.The authors state that other environments lead to similar results.

D Constrained Decoding Implementation

Constrained decoding represents legitimate action strings as paths in a token-prefix trie and renormalizes probabilities only over valid continuations. The implementation uses local normalization so forced formatting tokens do not distort action preferences.

A token-prefix trie organizes legitimate action strings as token paths, with nodes for prefixes, edges for tokens, and leaves for complete actions.The trie supports constrained sampling from the finite set of environment-supplied actions.
At each prefix s, invalid next tokens are masked and logits are renormalized over the legitimate continuation set A(s).This produces locally constrained next-token probabilities.
A legitimate action’s probability is the product of locally normalized edge probabilities along its path, with duplicate token paths summed.Multiple tokenizations decoding to the same action string contribute to that action’s total probability.
Local normalization differs from global conditioning, which multiplies complete-path probabilities before applying one normalization across legitimate leaves.The evaluation uses the local formulation rather than global conditioning.
Local normalization assigns probability 1 to forced formatting transitions, preventing delimiters such as brackets and commas from changing relative action probabilities.Global conditioning can instead reflect prefix-dependent format likelihood rather than decision preference.

E Detailed Optimization Trajectories

The appendix compares complete NPO, GEPA, and GRPO optimization trajectories across all 22 TextArena games. Shared axes expose rollout efficiency, training-metric progress, and relative prompt-optimization behavior across environments.

Detailed Optimization Trajectories: The trajectories cover all 22 TextArena environments with NPO, GEPA, and GRPO overlaid for direct comparison.Each game receives its own trajectory comparison.
Detailed Optimization Trajectories: Cumulative rollouts form the horizontal axis, while each game’s training metric forms the vertical axis.The layout supports comparison of optimization dynamics and performance improvement rates.
Optimized prompts at a glance: Despite fewer rollouts, NPO produces prompts similar in detail and depth to GEPA’s prompts.This comparison concerns the optimized prompts reported for the tasks.

HotpotQA GEPA

The HotpotQA materials show jointly optimized multi-hop prompts for successive retrieval, summarization, and answer-generation stages. The NPO prompt emphasizes structured intermediate fields and concise factual answering across the multi-hop pipeline.

HotpotQA GEPA: The first hop summarizes retrieved passages into a summary field while requiring a single strict JSON object.The first-hop prompt identifies the question and passages as inputs.
HotpotQA GEPA: The second hop generates a retrieval query from the original question and the first summary.This stage uses the intermediate summary to formulate the next retrieval step.
HotpotQA GEPA: The third hop summarizes newly retrieved passages using the question and prior context, while treating context as potentially incomplete or incorrect.The prompt prioritizes verification against newly retrieved evidence.
HotpotQA GEPA: The fourth hop produces a concise factual answer from the question and two intermediate summaries.The prompt warns against blindly trusting summaries that claim information is unavailable.
HotpotQA GEPA: HotpotQA is treated as a multi-hop task whose prompts for each hop are optimized jointly.The pipeline includes multiple prompt stages rather than a single answer-generation prompt.

Blackjack GEPA

The game prompts encode task-specific legal-action and output constraints for Blackjack and several comparison environments. Their recurring design prioritizes selecting valid moves and emitting exactly the required action format.

Blackjack GEPA: Blackjack prompts require one valid action, either [Hit] or [Stand], with no explanation or extra text.The Blackjack setup specifies using the current hand and dealer upcard.
Blackjack GEPA: The TextArena Blackjack prompt targets final reward across 20 hands under a 21-turn limit.Unplayed hands score zero, making turn usage part of the task objective.
Game prompt constraints: Briscola prompts optimize raw points captured from a 120-point deck game, with card values and trick rules specified explicitly.The NPO prompt states that scores above 60 win while encouraging the largest margin.
Game prompt constraints: Chess prompts require exactly one bracketed UCI move copied verbatim from the valid-move list.The NPO version also treats the observed side to move as authoritative.
Game prompt constraints: Across other games, prompts similarly constrain actions to legal moves copied from environment-provided lists or formats.Examples include Connect Four columns, Countdown indexed operations, Crusade moves, and Checkers coordinates.

KuhnPoker GEPA

The section presents task-specific optimized prompts for interactive games, including Kuhn Poker, Nim, Othello, Pig, Quantum Tic Tac Toe, and Lights Out. These prompts encode game rules, legal actions, and output constraints for the target agent.

KuhnPoker GEPA: Kuhn Poker prompts specify the player’s role, card ranking, betting actions, and exact one-line action output.The GEPA prompt restricts responses to [check], [bet], [call], or [fold].
LightsOut: Lights Out prompts define orthogonal toggling on a 5x5 board and instruct the agent to reach an all-off configuration.They also discourage repeating a press unless the board state makes it useful.
Nim: Nim prompts require counting pile objects and selecting a legal removal based on the piles’ configuration.The GEPA version emphasizes counting bullet symbols and the NPO version includes nim-sum strategy.
Othello: Othello prompts constrain play to listed valid moves and prioritize strategically important board positions such as corners.Both prompts specify a 4x4 board, while the NPO prompt explicitly prioritizes legal corners.
Pig: Pig prompts define exact one-line [roll] or [hold] outputs while using score and turn-total conditions to determine actions.The prompts differ in which game fields and winning-score rules they treat as authoritative.
Quantum Tic Tac Toe: Quantum Tic Tac Toe prompts enforce player-turn rules, legal move formatting, and classical winning-line conditions.The NPO prompt requires a final pair of distinct board coordinates appearing among the legal options.