Does Reasoning Help LLM Agents Play Dungeons and Dragons? A Prompt Engineering Experiment

2025-10-20

AI Summary

This study investigates the capability of large language models (LLMs) to predict player actions and generate executable commands for the Avrae Discord bot in *Dungeons & Dragons* (D&D) gameplay. We conduct controlled experiments on the FIREBALL dataset using two distinct LLMs: DeepSeek-R1-Distill-LLaMA-8B (reasoning-oriented) and LLaMA-3.1-8B-Instruct (instruction-tuned), systematically evaluating how prompt engineering affects output accuracy. Results demonstrate that the instruction-tuned model significantly outperforms the reasoning-oriented one; moreover, fine-grained, single-sentence prompt modifications induce substantial improvements in command generation fidelity—highlighting extreme sensitivity to instruction design. Our key contribution is the first empirical identification of instruction tuning as a critical factor in game-semantic parsing tasks. We further establish that precise instruction modeling—not raw reasoning capacity—is more effective for generating syntactically and semantically correct, structured game commands. This work provides a reproducible, practically viable technical pathway for LLM-driven interactive game systems.

Technology Category

Machine Learning: Large Multimodal Models (LMMs)Natural Language Processing: Prompt Engineering / PromptingHumans and AI: Game Design — Procedural Content Generation & Storytelling

Application Category

Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactionsSearch and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved informationEconomics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

This paper explores the application of Large Language Models (LLMs) and reasoning to predict Dungeons & Dragons (DnD) player actions and format them as Avrae Discord bot commands. Using the FIREBALL dataset, we evaluated a reasoning model, DeepSeek-R1-Distill-LLaMA-8B, and an instruct model, LLaMA-3.1-8B-Instruct, for command generation. Our findings highlight the importance of providing specific instructions to models, that even single sentence changes in prompts can greatly affect the output of models, and that instruct models are sufficient for this task compared to reasoning models.

Problem

• Evaluating reasoning versus instruct models for DnD action prediction
• Assessing prompt engineering impact on command generation accuracy
• Testing LLM performance on Avrae bot command formatting

Innovation

• Using reasoning and instruct models for command generation
• Evaluating models with FIREBALL dataset for DnD actions
• Comparing prompt engineering effects on model outputs