Agent-GWO: Collaborative Agents for Dynamic Prompt Optimization in Large Language Models

Xudong Wang; Chaoning Zhang; Chenghao Li; Shuxu Chen; Qigan Sun; Jiaquan Zhang; Fachrina Dewi Puspitasari; Tae-Ho Kim

cs.NEApr 14, 2026

Agent-GWO: Collaborative Agents for Dynamic Prompt Optimization in Large Language Models

Authors: Xudong Wang, Chaoning Zhang, Chenghao Li, Shuxu Chen, Qigan Sun, Jiaquan Zhang, Fachrina Dewi Puspitasari, Tae-Ho Kim,

Organizations: Kyung Hee University · University of Electronic Science and Technology of China · Nota Inc. · Tongji University

Abstract

Large Language Models (LLMs) have demonstrated strong capabilities in complex reasoning tasks, while recent prompting strategies such as Chain-of-Thought (CoT) have further elevated their performance in handling complex logical problems. Despite these advances, high-quality reasoning remains heavily reliant on manual static prompts and is sensitive to decoding configurations and task distributions, leading to performance fluctuations and limited transferability. Existing automatic prompt optimization methods typically adopt single-agent local search, failing to simultaneously optimize prompts and decoding hyperparameters within a unified framework to achieve stable global improvements. To address this limitation, we propose Agent-GWO, a dynamic prompt optimization framework for complex reasoning. Specifically, we unify prompt templates and decoding hyperparameters as inheritable agent configurations. By leveraging the leader-follower mechanism of the Grey Wolf Optimizer (GWO), we automatically select three leader agents ( $α$ , $β$ , and $δ$ ) to guide the collaborative updates of the remaining agents, enabling iterative convergence toward robust optimal reasoning configurations that can be seamlessly integrated for inference. Extensive experiments on multiple mathematical and hybrid reasoning benchmarks across diverse LLM backbones show that Agent-GWO consistently improves accuracy and stability over existing prompt optimization methods. The code will be released publicly.

Figures & tables

Figure 1: GWO abstracts each LLM agent as a “wolf” which is characterized by hyperparameters & reasoning prompt. These two guide the behavior of the LLMs during the optimization process.

Figure 2: The overall Agent-GWO framework operates by having each agent process a dataset, Prompt, and parameters to produce outputs, which are then scored. Only the top-3 agents ( $\alpha$ , $\beta$ , and $\delta$ ) are differentiated, while the rest (i.e., $\omega$ ) are updated using the prompt and parameters from the previous round. This iterative ranking and update process continues until termination, after which the final agent is set as the top-ranked $\alpha$ from the last iteration. Note: The left-bottom subfigure illustrates the learning process of each agent at each iteration and the right-bottom subfigure illustrates the optimization process.

GPT-4o-mini
Model	Math Reasoning Tasks						Hybrid Reasoning Tasks					Avg
Model	GSM8K	MATH	SVAMP	MultiArith	ASDiv	AQUA	MMLU	BBH	Date	CLUTRR	MATH_MIX	Avg
CoT	85.4%	74.8%	84.7%	89.5%	92.3%	65.3%	63.4%	66.6%	52.1%	66.2%	80.1%	74.6%
CoT-SC/n=5	89.9%	76.3%	85.8%	89.7%	93.3%	70.7%	67.1%	69.0%	54.7%	72.2%	81.6%	77.3%
Self-Refine	88.9%	75.1%	85.0%	89.9%	92.2%	67.8%	64.8%	67.5%	53.0%	68.4%	83.8%	76.0%
AFlow	93.5%	78.9%	88.7%	91.9%	93.4%	72.0%	69.5%	68.0%	66.0%	72.8%	85.3%	80.0%

Table 1: Performance comparison of Agent-GWO against baselines across math and hybrid reasoning tasks. Bold indicates the best result, and underlined indicates the second-best result.

Model / Method	Math Reasoning Tasks						Hybrid Reasoning Tasks					Avg. Increase
Model / Method	GSM8K	MATH	SVAMP	MultiArith	ASDiv	AQUA	MMLU	BBH	Date	CLUTRR	MATH_MIX	Avg. Increase
GPT-4o-mini
CoT	85.4%	74.8%	84.7%	89.5%	92.3%	65.3%	63.4%	66.6%	52.1%	66.2%	80.1%	-
CoT+GWO/n=5	95.1%	79.9%	92.4%	99.5%	93.8%	76.1%	73.7%	71.3%	77.1%	74.9%	85.0%	$\uparrow$ 8.9%
CoT+GWO/n=6	96.5%	81.5%	92.9%	99.8%	94.5%	76.8%	74.9%	72.2%	78.3%	75.8%	95.1%	$\uparrow$ 10.7%
GPT-4.1-mini

Table 2: Combined evaluation of adaptability of GWO to CoT on math and hybrid reasoning tasks.

Parameter	Sampling Rule	Perf.(%)	Cost ($)
$t_{j}$	(0.2, 0.1)	$\downarrow$ 1.0–1.5%	$\downarrow$ 3–5%
	(0.4, 0.1)	$\downarrow$ 0.3–0.6%	$\downarrow$ 1–2%
	(0.6, 0.1)	baseline	baseline
	(0.8, 0.1)	$\downarrow$ 0.5–1.0%	$\uparrow$ 4–7%
$p_{j}$	(0.2, 0.1)	$\downarrow$ 0.8–1.3%	$\downarrow$ 2–4%
	(0.4, 0.1)	$\downarrow$ 0.2–0.5%	$\downarrow$ 1–2%

Table 3: Sensitivity analysis of agent-level decoding hyperparameters. Entries report relative performance and cost changes compared to the baseline configuration used in our framework.

Elite Size	Elite	GSM8K	MATH	MMLU
$m=1$	$\alpha$	92.4	76.8	72.0
$m=2$	$\alpha,\beta$	94.8	79.1	74.2
$m=3$	$\alpha,\beta,\delta$	95.9	80.2	75.3
$m=4$	$\alpha,\beta,\delta,\omega_{1}$	95.0	79.4	74.6
$m=5$	$\alpha,\beta,\delta,\omega_{1},\omega_{2}$	94.1	78.5	73.8

Table 4: Ablation on Elite Set Size. Results under different elite set sizes

m

\alpha

\beta

, and

\delta

denote the top three leaders in standard GWO, while

\omega_{1}

and

\omega_{2}

denote the 4th- and 5th-ranked agents by fitness, respectively.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Illustration of GWO algorithm.

Method	Model	GSM8K
CoMAT Leang et al. (2025)	GPT-4	93.7
CoT Wei et al. (2022)	GPT-4	94.5
FCoT Lyu et al. (2023)	GPT-4	95.0
MathPrompter Imani et al. (2023)	GPT-4	95.6
QuaSAR Radford et al. (2018)	GPT-4	96.5
MathDivide Srivastava and Gandhi (2024)	GPT-4	96.8

Appendix

Table 5: Accuracy on GSM8K comparing GPT-4 and GPT-4o-mini.

Method	Base	GSM8K	MATH
OMI2 Li et al. (2025b)	Qwen2.5	84.1	72.3
CODEI/O++ Li et al. (2025b)	Qwen2.5	85.7	72.1
PyEdu Li et al. (2025b)	Qwen2.5	85.8	71.4
CODEI/O Li et al. (2025b)	Qwen2.5	86.4	71.9
OC-SFT-1 Li et al. (2025b)	Qwen2.5	86.7	70.9
WI Li et al. (2025b)	Qwen2.5	87.0	71.4

Appendix

Table 6: Accuracy (%) on GSM8K and MATH using Qwen2.5-Coder-7B.

Automatic Prompt Optimization

Automatic prompt optimization uses model or task feedback to revise instructions for stronger, more reliable language-model performance.

See all 99 papers

Reflective Prompt Tuning through Language Model Function-Calling
May 20, 2026Farima Fatahi Bayat, Moin Aminnaseri, Pouya Pezeshkpour +1LLM Self-Refinement LLM Prompting
Megagon Labs
Beyond the Answer: Decoding the Behavior of LLMs as Scientific Reasoners
Mar 30, 2026Rohan Pandey, Eric Ye, Michael LiLLM Prompting LLM Interpretability
Carnegie Mellon University
RASPRef: Retrieval-Augmented Self-Supervised Prompt Refinement for Large Reasoning Models
Mar 27, 2026Rahul SoniLLM Input Optimization LLM Prompting
Independent Researcher

May 20, 2026J/K move · Enter open · S save

Reflective Prompt Tuning through Language Model Function-Calling

Farima Fatahi Bayat, Moin Aminnaseri, Pouya Pezeshkpour, Estevam Hruschka

Megagon Labs

LLM Self-Refinement LLM Prompting Automatic Prompt Optimization

Open paper AskPDF arXiv

Large language models (LLMs) have become increasingly capable of following instructions and complex reasoning, making prompting a flexible interface for adapting models without parameter updates. Yet prompt design remains labor-intensive and highly sensitive to formatting, phrasing, and instruction order, motivating automated prompt optimization methods that reduce manual effort while preserving inference-time flexibility. However, existing methods often search over prompt candidates or use fixed critique-refine pipelines driven by individual examples or small batches, limiting their ability to capture systematic error patterns and make targeted edits grounded in failure history. We propose Reflective Prompt Tuning (RPT), a framework that uses LLM function calling to simulate the iterative workflow of human prompt engineers. An LLM optimizer calls a diagnostic function that evaluates the target model over an entire optimization set, summarizes recurring failure modes, and returns a structured diagnostic report. The optimizer uses this report, together with an accumulated memory of prior reports, to revise the prompt for the next iteration. RPT further supports confidence-aware optimization by using calibration signals in diagnostic feedback and final prompt selection. Across three reasoning tasks, RPT improves over initial prompts by up to 12.9 points, remains competitive with state of the art, and improves confidence calibration. Our analyses show that RPT is especially effective on multi-hop and mathematical reasoning, producing targeted prompt revisions that align with diagnosed failure patterns and lead to gains in task performance and calibration.

Abstract

Automatic Prompt Optimization

Explore similar work

Reflective Prompt Tuning through Language Model Function-Calling

Beyond the Answer: Decoding the Behavior of LLMs as Scientific Reasoners

RASPRef: Retrieval-Augmented Self-Supervised Prompt Refinement for Large Reasoning Models