Influencing LLM Multi-Agent Dialogue via Policy-Parameterized Prompts

Hongbo Bo; Jingyu Hu; Weiru Liu

cs.AIMar 10, 2026

Influencing LLM Multi-Agent Dialogue via Policy-Parameterized Prompts

Authors: Hongbo Bo, Jingyu Hu, Weiru Liu

Organizations: University of Bristol, Bristol, United Kingdom

Abstract

Large Language Models (LLMs) have emerged as a new paradigm for multi-agent systems. However, existing research on the behaviour of LLM-based multi-agents relies on ad hoc prompts and lacks a principled policy perspective. Different from reinforcement learning, we investigate whether prompt-as-action can be parameterized so as to construct a lightweight policy which consists of a sequence of state-action pairs to influence conversational behaviours without training. Our framework regards prompts as actions executed by LLMs, and dynamically constructs prompts through five components based on the current state of the agent. To test the effectiveness of parameterized control, we evaluated the dialogue flow based on five indicators: responsiveness, rebuttal, evidence usage, non-repetition, and stance shift. We conduct experiments using different LLM-driven agents in two discussion scenarios related to the general public and show that prompt parameterization can influence the dialogue dynamics. This result shows that policy-parameterised prompts offer a simple and effective mechanism to influence the dialogue process, which will help the research of multi-agent systems in the direction of social simulation.

Figures & tables

Figure 1. The overall framework illustrates the process from an agent’s state representation to action generation, LLM-based action execution, and evaluation.

		Land					Education
Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.88_{\pm 0.33}$	$0.19_{\pm 0.40}$	$0.47_{\pm 0.42}$	$0.10_{\pm 0.30}$	$0.52_{\pm 0.08}$	$0.88_{\pm 0.33}$	$0.25_{\pm 0.43}$	$0.45_{\pm 0.41}$	$0.13_{\pm 0.34}$	$0.52_{\pm 0.09}$
	Light	$0.90_{\pm 0.30}$	$0.21_{\pm 0.41}$	$0.33_{\pm 0.35}$	$0.32_{\pm 0.47}$	$0.51_{\pm 0.09}$	$0.87_{\pm 0.34}$	$0.27_{\pm 0.44}$	$0.41_{\pm 0.36}$	$0.25_{\pm 0.44}$	$0.52_{\pm 0.11}$
	Struct	$0.74_{\pm 0.44}$	$0.27_{\pm 0.45}$	$0.58_{\pm 0.39}$	$0.31_{\pm 0.46}$	$0.49_{\pm 0.09}$	$0.89_{\pm 0.32}$	$0.09_{\pm 0.29}$	$0.54_{\pm 0.37}$	$0.17_{\pm 0.37}$	$0.52_{\pm 0.11}$
Q2	None	$0.87_{\pm 0.34}$	$0.37_{\pm 0.48}$	$0.47_{\pm 0.44}$	$0.11_{\pm 0.31}$	$0.49_{\pm 0.08}$	$0.85_{\pm 0.35}$	$0.15_{\pm 0.36}$	$0.41_{\pm 0.41}$	$0.37_{\pm 0.48}$	$0.48_{\pm 0.07}$
	Light	$0.89_{\pm 0.31}$	$0.31_{\pm 0.47}$	$0.32_{\pm 0.34}$	$0.27_{\pm 0.44}$	$0.45_{\pm 0.06}$	$0.89_{\pm 0.32}$	$0.06_{\pm 0.24}$	$0.35_{\pm 0.35}$	$0.41_{\pm 0.49}$	$0.45_{\pm 0.08}$

Table 1. Overall performance of policies under different rule templates on 5 queries over 10-round conversation. Values are the mean

\pm

std of 5 runs experiments. Because Rebuttal and Evid. are binary (0/1), the mean is just the proportion of 1s and the standard deviation can appear large relative to the mean.

Land		Education
Agent	LLM	Agent	LLM
Farmer	Qwen3	Rural Teacher	Qwen3
Conservationist	Llama3	Urban Parent	Llama3
Community Rep.	Mistral	Policy Maker	Mistral

Table 2. Agents and LLM assignments across two scenarios.

Panel of Figure 2 — Figure 2. Round-wise changes of three dialogue metrics under different rule templates. Each curve shows the mean value (with 95% confidence interval).

Weights	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
$w_{T}=1.0$	None	$0.85_{\pm 0.35}$	$0.28_{\pm 0.45}$	$0.43_{\pm 0.42}$	$0.10_{\pm 0.29}$	$0.48_{\pm 0.07}$
$w_{M}=1.0$	Light	$0.85_{\pm 0.36}$	$0.31_{\pm 0.46}$	$0.33_{\pm 0.36}$	$0.28_{\pm 0.45}$	$0.47_{\pm 0.07}$
$w_{D}=1.0$	Struct	$0.80_{\pm 0.40}$	$0.22_{\pm 0.41}$	$0.62_{\pm 0.36}$	$0.20_{\pm 0.40}$	$0.47_{\pm 0.08}$
$w_{T}=1.0$	None	$0.86_{\pm 0.34}$	$0.28_{\pm 0.45}$	$0.30_{\pm 0.35}$	$0.30_{\pm 0.46}$	$0.47_{\pm 0.08}$
$w_{M}=1.0$	Light	$0.84_{\pm 0.37}$	$0.33_{\pm 0.47}$	$0.31_{\pm 0.34}$	$0.28_{\pm 0.45}$	$0.44_{\pm 0.10}$
$w_{D}=1.5$	Struct	$0.82_{\pm 0.38}$	$0.17_{\pm 0.37}$	$0.49_{\pm 0.40}$	$0.17_{\pm 0.37}$	$0.48_{\pm 0.07}$

Table 3. Overall performance of policies with different weights and rule templates in the Land scenario. The detailed experimental results are presented in Appendix C .

Panel of Figure 3 — Figure 3. An Example of Adaptive Weight Changes. Farmer agent’s $W_{D}$ and $W_{M}$ changes in 10 rounds under the Q1 topic.

Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
None	$0.86_{\pm 0.35}$	$0.26_{\pm 0.44}$	$0.45_{\pm 0.39}$	$0.18_{\pm 0.38}$	$0.48_{\pm 0.07}$
Light	$0.85_{\pm 0.36}$	$0.39_{\pm 0.49}$	$0.37_{\pm 0.35}$	$0.29_{\pm 0.46}$	$0.45_{\pm 0.07}$
Struct	$0.82_{\pm 0.38}$	$0.13_{\pm 0.34}$	$0.52_{\pm 0.40}$	$0.16_{\pm 0.37}$	$0.48_{\pm 0.07}$

Table 4. Overall performance of policies with adaptive weight settings in the Land scenario, averaged over all agents. Detailed results are presented in Appendix G .

Panel of Figure 4 — Figure 4. Round-wise changes of three dialogue metrics with adaptive weights.

$\alpha$	Rule	Resp.	Reb.	Non-rep.	Evid.	Stance
0.05	None	$0.83\pm 0.38$	$0.24\pm 0.43$	$0.54\pm 0.38$	$0.15\pm 0.35$	$0.49\pm 0.08$
	Light	$0.83\pm 0.37$	$0.25\pm 0.43$	$0.42\pm 0.36$	$0.21\pm 0.41$	$0.46\pm 0.08$
	Struct	$0.84\pm 0.37$	$0.10\pm 0.30$	$0.62\pm 0.36$	$0.18\pm 0.39$	$0.49\pm 0.06$
0.1	None	$0.83\pm 0.37$	$0.20\pm 0.40$	$0.54\pm 0.37$	$0.17\pm 0.38$	$0.49\pm 0.08$
	Light	$0.84\pm 0.37$	$0.24\pm 0.42$	$0.41\pm 0.35$	$0.28\pm 0.45$	$0.46\pm 0.08$
	Struct	$0.83\pm 0.37$	$0.08\pm 0.28$	$0.58\pm 0.37$	$0.20\pm 0.40$	$0.49\pm 0.06$

Table 5. Effect of Different

\alpha

Values.

	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
S1	None	$0.67_{\pm 0.47}$	$0.28_{\pm 0.45}$	$0.25_{\pm 0.39}$	$0.14_{\pm 0.35}$	$0.48_{\pm 0.07}$
	Light	$0.59_{\pm 0.49}$	$0.26_{\pm 0.44}$	$0.26_{\pm 0.37}$	$0.16_{\pm 0.36}$	$0.44_{\pm 0.08}$
	Struct	$0.74_{\pm 0.44}$	$0.12_{\pm 0.33}$	$0.38_{\pm 0.41}$	$0.10_{\pm 0.29}$	$0.47_{\pm 0.07}$
S2	None	$0.71_{\pm 0.45}$	$0.26_{\pm 0.44}$	$0.39_{\pm 0.40}$	$0.06_{\pm 0.23}$	$0.48_{\pm 0.09}$
	Light	$0.72_{\pm 0.45}$	$0.30_{\pm 0.46}$	$0.33_{\pm 0.35}$	$0.42_{\pm 0.49}$	$0.46_{\pm 0.09}$
	Struct	$0.70_{\pm 0.46}$	$0.16_{\pm 0.37}$	$0.61_{\pm 0.36}$	$0.22_{\pm 0.42}$	$0.50_{\pm 0.09}$

Table 6. Overall performance of policies comparison between homogeneous (

S1

) and heterogeneous (

S2

S3

) backbone LLMs variants. Results reported as mean

\pm

std. Detailed performance is in Appendix E .

Figure 5. Ablation Study of Components T, M and D.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.90_{\pm 0.30}$	$0.33_{\pm 0.47}$	$0.30_{\pm 0.35}$	$0.37_{\pm 0.48}$	$0.52_{\pm 0.08}$
	Light	$0.83_{\pm 0.37}$	$0.31_{\pm 0.47}$	$0.23_{\pm 0.33}$	$0.49_{\pm 0.50}$	$0.47_{\pm 0.08}$
	Struct	$0.81_{\pm 0.40}$	$0.19_{\pm 0.40}$	$0.54_{\pm 0.40}$	$0.29_{\pm 0.46}$	$0.51_{\pm 0.08}$
Q2	None	$0.90_{\pm 0.30}$	$0.34_{\pm 0.47}$	$0.32_{\pm 0.36}$	$0.39_{\pm 0.49}$	$0.46_{\pm 0.07}$
	Light	$0.88_{\pm 0.33}$	$0.32_{\pm 0.47}$	$0.35_{\pm 0.33}$	$0.25_{\pm 0.44}$	$0.40_{\pm 0.06}$
	Struct	$0.87_{\pm 0.33}$	$0.14_{\pm 0.35}$	$0.50_{\pm 0.40}$	$0.17_{\pm 0.38}$	$0.46_{\pm 0.06}$

Appendix

Table 7.

w_{T}=1.0,w_{M}=1.0,w_{D}=1.5

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.86_{\pm 0.35}$	$0.27_{\pm 0.45}$	$0.57_{\pm 0.37}$	$0.27_{\pm 0.44}$	$0.56_{\pm 0.08}$
	Light	$0.88_{\pm 0.33}$	$0.44_{\pm 0.50}$	$0.38_{\pm 0.37}$	$0.38_{\pm 0.49}$	$0.49_{\pm 0.07}$
	Struct	$0.83_{\pm 0.37}$	$0.22_{\pm 0.42}$	$0.53_{\pm 0.40}$	$0.35_{\pm 0.48}$	$0.51_{\pm 0.08}$
Q2	None	$0.87_{\pm 0.33}$	$0.37_{\pm 0.48}$	$0.46_{\pm 0.41}$	$0.05_{\pm 0.23}$	$0.49_{\pm 0.06}$
	Light	$0.88_{\pm 0.33}$	$0.28_{\pm 0.45}$	$0.36_{\pm 0.37}$	$0.23_{\pm 0.42}$	$0.43_{\pm 0.06}$
	Struct	$0.86_{\pm 0.35}$	$0.25_{\pm 0.43}$	$0.47_{\pm 0.41}$	$0.12_{\pm 0.33}$	$0.49_{\pm 0.06}$

Appendix

Table 8.

w_{T}=1.0,w_{M}=1.5,w_{D}=1.0

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.88_{\pm 0.33}$	$0.33_{\pm 0.47}$	$0.32_{\pm 0.39}$	$0.08_{\pm 0.27}$	$0.61_{\pm 0.05}$
	Light	$0.83_{\pm 0.38}$	$0.41_{\pm 0.49}$	$0.27_{\pm 0.33}$	$0.34_{\pm 0.47}$	$0.56_{\pm 0.07}$
	Struct	$0.90_{\pm 0.30}$	$0.27_{\pm 0.45}$	$0.37_{\pm 0.38}$	$0.25_{\pm 0.44}$	$0.59_{\pm 0.07}$
Q2	None	$0.86_{\pm 0.35}$	$0.30_{\pm 0.46}$	$0.21_{\pm 0.34}$	$0.00_{\pm 0.00}$	$0.54_{\pm 0.06}$
	Light	$0.89_{\pm 0.31}$	$0.31_{\pm 0.46}$	$0.26_{\pm 0.33}$	$0.25_{\pm 0.44}$	$0.50_{\pm 0.06}$
	Struct	$0.79_{\pm 0.41}$	$0.48_{\pm 0.50}$	$0.29_{\pm 0.37}$	$0.08_{\pm 0.27}$	$0.51_{\pm 0.05}$

Appendix

Table 9.

w_{T}=1.5,w_{M}=1.0,w_{D}=1.0

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.88_{\pm 0.33}$	$0.31_{\pm 0.46}$	$0.37_{\pm 0.41}$	$0.23_{\pm 0.42}$	$0.56_{\pm 0.05}$
	Light	$0.87_{\pm 0.34}$	$0.35_{\pm 0.48}$	$0.40_{\pm 0.37}$	$0.33_{\pm 0.47}$	$0.51_{\pm 0.08}$
	Struct	$0.75_{\pm 0.43}$	$0.26_{\pm 0.44}$	$0.53_{\pm 0.40}$	$0.35_{\pm 0.48}$	$0.51_{\pm 0.08}$
Q2	None	$0.89_{\pm 0.31}$	$0.44_{\pm 0.50}$	$0.29_{\pm 0.39}$	$0.08_{\pm 0.27}$	$0.46_{\pm 0.05}$
	Light	$0.90_{\pm 0.30}$	$0.31_{\pm 0.46}$	$0.35_{\pm 0.34}$	$0.21_{\pm 0.41}$	$0.44_{\pm 0.04}$
	Struct	$0.87_{\pm 0.34}$	$0.21_{\pm 0.41}$	$0.52_{\pm 0.39}$	$0.31_{\pm 0.47}$	$0.51_{\pm 0.07}$

Appendix

Table 10.

w_{T}=1.0,w_{M}=0.5,w_{D}=1.0

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.86_{\pm 0.35}$	$0.19_{\pm 0.39}$	$0.42_{\pm 0.41}$	$0.20_{\pm 0.40}$	$0.54_{\pm 0.08}$
	Light	$0.88_{\pm 0.33}$	$0.31_{\pm 0.47}$	$0.33_{\pm 0.35}$	$0.55_{\pm 0.50}$	$0.51_{\pm 0.10}$
	Struct	$0.76_{\pm 0.43}$	$0.31_{\pm 0.46}$	$0.57_{\pm 0.38}$	$0.42_{\pm 0.49}$	$0.52_{\pm 0.07}$
Q2	None	$0.82_{\pm 0.39}$	$0.50_{\pm 0.50}$	$0.45_{\pm 0.42}$	$0.11_{\pm 0.32}$	$0.49_{\pm 0.08}$
	Light	$0.87_{\pm 0.34}$	$0.33_{\pm 0.47}$	$0.32_{\pm 0.34}$	$0.18_{\pm 0.39}$	$0.44_{\pm 0.05}$
	Struct	$0.79_{\pm 0.41}$	$0.12_{\pm 0.33}$	$0.57_{\pm 0.37}$	$0.05_{\pm 0.23}$	$0.48_{\pm 0.06}$

Appendix

Table 11.

w_{T}=1.0,w_{M}=1.0,w_{D}=0.5

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.89_{\pm 0.31}$	$0.31_{\pm 0.46}$	$0.36_{\pm 0.41}$	$0.18_{\pm 0.39}$	$0.56_{\pm 0.07}$
	Light	$0.90_{\pm 0.30}$	$0.51_{\pm 0.50}$	$0.32_{\pm 0.36}$	$0.27_{\pm 0.44}$	$0.49_{\pm 0.08}$
	Struct	$0.83_{\pm 0.37}$	$0.33_{\pm 0.47}$	$0.55_{\pm 0.38}$	$0.34_{\pm 0.47}$	$0.57_{\pm 0.08}$
Q2	None	$0.89_{\pm 0.32}$	$0.38_{\pm 0.49}$	$0.29_{\pm 0.39}$	$0.03_{\pm 0.16}$	$0.48_{\pm 0.06}$
	Light	$0.90_{\pm 0.30}$	$0.30_{\pm 0.46}$	$0.22_{\pm 0.33}$	$0.46_{\pm 0.50}$	$0.43_{\pm 0.05}$
	Struct	$0.85_{\pm 0.36}$	$0.20_{\pm 0.40}$	$0.51_{\pm 0.40}$	$0.12_{\pm 0.33}$	$0.49_{\pm 0.06}$

Appendix

Table 12.

w_{T}=0.5,w_{M}=1.0,w_{D}=1.0

	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
None	None	$0.90_{\pm 0.30}$	$0.09_{\pm 0.29}$	$0.79_{\pm 0.16}$	$0.00_{\pm 0.00}$	$0.36_{\pm 0.09}$
	Light	$0.89_{\pm 0.31}$	$0.10_{\pm 0.30}$	$0.79_{\pm 0.14}$	$0.03_{\pm 0.16}$	$0.33_{\pm 0.08}$
	Struct	$0.85_{\pm 0.35}$	$0.03_{\pm 0.17}$	$0.86_{\pm 0.11}$	$0.01_{\pm 0.08}$	$0.40_{\pm 0.08}$
only T	None	$0.88_{\pm 0.33}$	$0.34_{\pm 0.47}$	$0.74_{\pm 0.20}$	$0.00_{\pm 0.05}$	$0.49_{\pm 0.09}$
	Light	$0.84_{\pm 0.37}$	$0.35_{\pm 0.48}$	$0.73_{\pm 0.23}$	$0.05_{\pm 0.21}$	$0.45_{\pm 0.09}$
	Struct	$0.87_{\pm 0.33}$	$0.34_{\pm 0.47}$	$0.78_{\pm 0.18}$	$0.03_{\pm 0.17}$	$0.50_{\pm 0.08}$

Appendix

Table 13. Ablation Study

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.81_{\pm 0.40}$	$0.35_{\pm 0.48}$	$0.18_{\pm 0.35}$	$0.13_{\pm 0.33}$	$0.49_{\pm 0.09}$
	Light	$0.68_{\pm 0.47}$	$0.28_{\pm 0.45}$	$0.23_{\pm 0.36}$	$0.28_{\pm 0.45}$	$0.44_{\pm 0.08}$
	Struct	$0.67_{\pm 0.47}$	$0.04_{\pm 0.20}$	$0.37_{\pm 0.42}$	$0.09_{\pm 0.29}$	$0.47_{\pm 0.07}$
Q2	None	$0.68_{\pm 0.47}$	$0.25_{\pm 0.43}$	$0.21_{\pm 0.37}$	$0.33_{\pm 0.47}$	$0.47_{\pm 0.08}$
	Light	$0.64_{\pm 0.48}$	$0.30_{\pm 0.46}$	$0.22_{\pm 0.34}$	$0.25_{\pm 0.43}$	$0.42_{\pm 0.06}$
	Struct	$0.68_{\pm 0.47}$	$0.00_{\pm 0.00}$	$0.36_{\pm 0.42}$	$0.01_{\pm 0.12}$	$0.46_{\pm 0.07}$

Appendix

Table 14. Homogeneous (S1): Overall performance of policies when all backbone LLMs are Qwen3. Report as mean

\pm

std over all agents and all rounds in

Land

scenario.

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.75_{\pm 0.44}$	$0.35_{\pm 0.48}$	$0.37_{\pm 0.40}$	$0.07_{\pm 0.25}$	$0.50_{\pm 0.10}$
	Light	$0.81_{\pm 0.40}$	$0.31_{\pm 0.46}$	$0.37_{\pm 0.35}$	$0.49_{\pm 0.50}$	$0.44_{\pm 0.10}$
	Struct	$0.79_{\pm 0.41}$	$0.12_{\pm 0.33}$	$0.61_{\pm 0.37}$	$0.23_{\pm 0.42}$	$0.55_{\pm 0.07}$
Q2	None	$0.70_{\pm 0.46}$	$0.31_{\pm 0.46}$	$0.36_{\pm 0.39}$	$0.07_{\pm 0.26}$	$0.46_{\pm 0.09}$
	Light	$0.69_{\pm 0.47}$	$0.28_{\pm 0.45}$	$0.26_{\pm 0.32}$	$0.33_{\pm 0.47}$	$0.46_{\pm 0.08}$
	Struct	$0.67_{\pm 0.47}$	$0.11_{\pm 0.31}$	$0.61_{\pm 0.34}$	$0.28_{\pm 0.45}$	$0.48_{\pm 0.09}$

Appendix

Table 15. Heterogeneous: Overall performance of policies with the variant of the backbones. Report as mean

\pm

std over all agents and all rounds in

Land

scenario.

Rule	Agent	Resp.	Rebuttal	Non-rep.	Evid.	Stance
None	Comm.	$0.90\pm 0.00$	$0.08\pm 0.04$	$0.84\pm 0.03$	$0.28\pm 0.14$	$0.46\pm 0.05$
	Conser.	$0.87\pm 0.03$	$0.24\pm 0.17$	$0.71\pm 0.06$	$0.02\pm 0.03$	$0.49\pm 0.06$
	Farmer 1(Qwen3)	$0.88\pm 0.02$	$0.50\pm 0.33$	$0.52\pm 0.15$	$0.13\pm 0.13$	$0.59\pm 0.04$
	Farmer 2(Llama)	$0.87\pm 0.03$	$0.63\pm 0.21$	$0.59\pm 0.16$	$0.10\pm 0.10$	$0.68\pm 0.06$
	Farmer 3(Mistral)	$0.88\pm 0.02$	$0.26\pm 0.17$	$0.80\pm 0.08$	$0.33\pm 0.20$	$0.51\pm 0.02$
Light	Comm.	$0.90\pm 0.01$	$0.09\pm 0.07$	$0.63\pm 0.03$	$0.21\pm 0.14$	$0.40\pm 0.03$

Appendix

Table 16. Performance of policies of each agent after adding two Farmer agents.

Rule	Agent	Resp.	Rebuttal	Non-rep.	Evid.	Stance
None	Comm.(mistral)	$0.48\pm 0.20$	$0.02\pm 0.04$	$0.22\pm 0.10$	$0.14\pm 0.13$	$0.41\pm 0.03$
	Conser.(Llama)	$0.72\pm 0.28$	$0.23\pm 0.34$	$0.30\pm 0.19$	$0.26\pm 0.26$	$0.48\pm 0.04$
	Farmer(Qwen)	$0.80\pm 0.12$	$0.58\pm 0.35$	$0.24\pm 0.06$	$0.04\pm 0.08$	$0.54\pm 0.03$
Light	Comm.(mistral)	$0.40\pm 0.16$	$0.04\pm 0.07$	$0.22\pm 0.05$	$0.23\pm 0.21$	$0.39\pm 0.04$
	Conser.(Llama)	$0.65\pm 0.31$	$0.21\pm 0.29$	$0.31\pm 0.13$	$0.13\pm 0.14$	$0.44\pm 0.04$
	Farmer(Qwen)	$0.73\pm 0.29$	$0.53\pm 0.38$	$0.26\pm 0.08$	$0.11\pm 0.12$	$0.51\pm 0.05$

Appendix

Table 17. Overall performance of policies of each agent under different Backbone LLMs

Query	Rule	Resp.	Rebuttal	Non-rep.	Evid.	Stance
Q1	None	$0.88_{\pm 0.33}$	$0.21_{\pm 0.41}$	$0.50_{\pm 0.38}$	$0.23_{\pm 0.42}$	$0.53_{\pm 0.07}$
	Light	$0.89_{\pm 0.32}$	$0.51_{\pm 0.50}$	$0.36_{\pm 0.35}$	$0.47_{\pm 0.50}$	$0.50_{\pm 0.07}$
	Struct	$0.83_{\pm 0.38}$	$0.17_{\pm 0.38}$	$0.54_{\pm 0.40}$	$0.15_{\pm 0.36}$	$0.51_{\pm 0.07}$
Q2	None	$0.87_{\pm 0.33}$	$0.38_{\pm 0.49}$	$0.49_{\pm 0.39}$	$0.19_{\pm 0.40}$	$0.48_{\pm 0.07}$
	Light	$0.89_{\pm 0.31}$	$0.39_{\pm 0.49}$	$0.39_{\pm 0.33}$	$0.17_{\pm 0.38}$	$0.44_{\pm 0.05}$
	Struct	$0.81_{\pm 0.40}$	$0.11_{\pm 0.32}$	$0.49_{\pm 0.39}$	$0.17_{\pm 0.38}$	$0.47_{\pm 0.07}$

Appendix

Table 18. Overall performance with adaptive weight setting. Report as mean

\pm

std over all agents and all rounds in

Land

scenario.

Query	Rule	Resp.	Reb.	Non-rep.	Evid.	Stance
M+D	None	$0.89\pm 0.32$	$0.02\pm 0.15$	$0.42\pm 0.39$	$0.06\pm 0.23$	$0.46\pm 0.08$
	Light	$0.89\pm 0.32$	$0.01\pm 0.12$	$0.39\pm 0.36$	$0.14\pm 0.35$	$0.45\pm 0.09$
	Struct	$0.89\pm 0.31$	$0.02\pm 0.15$	$0.46\pm 0.38$	$0.05\pm 0.22$	$0.45\pm 0.08$
T+D	None	$0.85\pm 0.35$	$0.46\pm 0.50$	$0.74\pm 0.24$	$0.13\pm 0.34$	$0.50\pm 0.08$
	Light	$0.85\pm 0.35$	$0.34\pm 0.47$	$0.69\pm 0.25$	$0.55\pm 0.50$	$0.48\pm 0.09$
	Struct	$0.80\pm 0.40$	$0.30\pm 0.46$	$0.80\pm 0.17$	$0.31\pm 0.46$	$0.53\pm 0.08$

Appendix

Table 19. Ablation Study on EDU scenario

Multi-Agent LLM Systems

Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.

See all 1,039 papers

Jun 12, 2026

Communication Policy Evolution for Proactive LLM Agents

LLM agents have rapidly evolved into autonomous systems, yet a persistent information gap remains between users and agents: communication is costly, while users' identical preferences further limit information exchange. To investigate how agents should communicate across modalities, this paper formalizes Communication Policy, establishes textual and UI-based policies, and then evaluates communication policies across diverse environments, personas, and model combinations. Building information asymmetry for proactive agents, we set up two complementary settings, User-Agent and Planner-Executor. Experimental results reveal complementary strengths between interaction channels: text-based interaction often facilitates task performance, while structured UI improves agents' response quality and persona compliance. Motivated by that, a hybrid method combines these advantages. We further propose Communication Policy Evolution (CPE), a self-evolution framework for refining communication policies through rollout and prompt-level evolving. Without model modification, CPE achieves the best task success across multiple settings using prompt refinement alone. Our findings identify communication behavior as a critical yet underexplored design dimension for LLM agents.

Xinbei Ma, Jiyang Qiu, Yao Yao +10

1Shanghai Jiao Tong University · 2OPPO Research Institute

LLM Agent Evaluation LLM Agent Self-Improvement LLM Agents Prompt Optimization

Ask PDF

Jun 16, 2026

Environment-Grounded Automated Prompt Optimization for LLM Game Agents

LLM agents in interactive environments are highly sensitive to their prompts, yet prompt engineering remains a manual, task-specific process. We introduce an automated prompt optimization framework for LLM agents that decomposes the observation-to-action pipeline into a goal-conditioned descriptor agent and an action selection agent, and iteratively refines each module's prompt through an LLM-driven evolutionary loop guided by environment returns. We propose a behavior analyzer to attribute episode outcomes to specific prompt components, and a mutator to propose targeted revisions to the prompt, before validating them through environment rollouts. We evaluate on all five BabyAI tasks in the BALROG benchmark, comparing our pipeline against BALROG's RobustCoTAgent under both plain and guided prompt initializations. Optimization improves performance consistently across tasks and conditions, without requiring updates to the model weights. On PutNext, a multi-step coordination task where the RobustCoTAgent achieves 0% success, our framework reaches up to 72.5% success rate using the same underlying LLM with optimized prompts. These results suggest that a multi-agent framework, combined with automatic prompt optimization, enhances LLMs without the need for fine-tuning or extensive human supervision.

Rean Clive Fernandes, Lukas Fehring, Theresa Eimer +2

Lamarr institute for ML and AI TU Dortmund University · Leibniz University Hannover · L3S Research Center

Multi-Agent LLM Systems Game-Playing Agents Automatic Prompt Optimization LLM Agents

Ask PDF

Sep 1, 2026

Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs

Prompt optimization can improve multi-agent LLM systems, but the prompts being optimized often serve two entangled roles: generating task-relevant content and specifying execution-critical protocols, such as message routing, output formatting, and termination signals, on which the underlying code relies. As a result, a prompt edit intended to improve content generation can inadvertently corrupt the protocol and cause the entire agent pipeline to fail. Our key observation is that these two roles have different representations: execution protocols are typically structured, while task-relevant content is usually expressed in unstructured language. Based on this, we propose control-data flow separation, where execution-critical control is represented as typed, validated program objects, while task-relevant language remains the optimizable data flow for agent communication. This design allows optimizers to improve multi-agent behavior without exposing the routing or formatting interface to prompt drift. Across synthetic reasoning, collaborative review generation, and insurance rating workflows, our framework empirically achieves 100% eventual protocol validity while consistently improving task performance.

Wentao Zhang, Syed Shariyar Murtaza, Junaid Ahmad Bhatti +4

University of Waterloo · Manulife

LLM Agent Orchestration LLM Agent Workflow Optimization Prompt Optimization Multi-Agent Communication Multi-Agent LLMs

Ask PDF

Abstract

Multi-Agent LLM Systems

Explore similar work

Communication Policy Evolution for Proactive LLM Agents

Environment-Grounded Automated Prompt Optimization for LLM Game Agents

Control-Data Flow Separation: Stable Prompt Optimization in Multi-Agent LLMs