At a Glance
Adjusting the parts and weights of prompts—without retraining models—can reliably change how multiple language-model agents argue, cite evidence, and shift stance during multi-round discussions.
ON THIS PAGE
What They Found
Treating the prompt as a lightweight policy (split into task/persona, memory, knowledge, rules, and weights) lets you nudge agent behaviour in predictable ways. Different prompt templates and weight schedules produced measurable differences in responsiveness, rebuttal frequency, evidence use, and how group stance evolved over 10-round dialogues. An adaptive weight scheduler can shift how much agents rely on task instructions, memory, or retrieved evidence over time, creating distinct conversation patterns without any extra training. For example, the approach aligns with ideas from the Tree of Thoughts Pattern to structure reasoning steps within prompts.
Need expert guidance?We can help implement this
Data Highlights
1Each experiment used 3 agents engaging in 10 rounds of dialogue, with the top-3 retrieved passages appended to each agent's prompt per round.
2For each scenario (land use and education) the team ran 5 topic queries × 5 independent runs, enabling comparison across 25 dialogues per scenario.
3Three different model backends were used (Qwen3-8B, Llama3-8B, Mistral-7B) and baseline comparisons held weights w_T = w_M = w_D = 1.0 while varying rule templates and scheduling.
What This Means
Engineers building multi-agent systems and technical leaders evaluating agent governance can use prompt-parameter controls to steer collective behaviour without retraining models. Researchers running social simulations or testing agent-to-agent evaluation setups can adopt these prompt knobs to run controlled experiments on evidence use, rebuttal, and stance dynamics. See the Model Context Protocol (MCP) Pattern for related context-management approaches.
Key Figures
Fig 1: Figure 1 . The overall framework illustrates the process from an agent’s state representation to action generation, LLM-based action execution, and evaluation.
Fig 2: (a) Responsiveness–Land
Fig 3: (a) W D W_{D}
Fig 4: (a) Responsiveness
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreLimitations
Experiments covered two topical scenarios and a limited set of prompts, so results may not generalize to very different domains or much larger agent populations. Knowledge bases were curated and supplemented using a separate language model, which can introduce bias or artefacts. Evaluation relied on an automated judge model and semantic similarity embeddings, so human evaluation is still needed to validate conversational quality and trust signals.
Methodology & More
Prompts can be treated as lightweight, interpretable policies to control language-model agents. Break each prompt into five parts—task/persona (T), dialogue memory (M), external knowledge (D), a rule template (R), and a weights vector (W)—and let the prompt assembled from those parts act as the agent’s action. Agents retrieve the top-3 relevant passages from their role-specific knowledge each round and include them in the prompt; an optional adaptive weight scheduler changes how much T, M, or D influence the prompt over time. Conversations were run with three agents (each backed by a different model) across two public-policy scenarios, and dialogues were scored each round on responsiveness, rebuttal, non-repetition, evidence usage, and stance shift. For production-like multi-agent deployments, the Orchestrator-Worker Pattern can provide practical structure, and the discussion interfaces with the ReAct Pattern for integrating reasoning and acting steps. Controlling prompts this way produced clear, repeatable differences: enabling or changing rule templates and weight schedules altered how often agents countered each other, how much they cited retrieved evidence, and whether group stances hardened or shifted over rounds. Because no extra model training was required, these controls are cheap to apply and easy to iterate. The approach offers a practical lever for running controlled social simulations, testing agent governance strategies, or shaping agent behaviour in production-like multi-agent deployments—while acknowledging needs for broader domain tests, human evaluations, and safeguards against knowledge-base or judge-model bias.
Need expert guidance?We can help implement this
Credibility Assessment:
ArXiv preprint, no affiliations provided and authors have low h-indices (~4). Lacks top-institution or venue signals—classified as emerging/limited info.