Aug 2026· International Conference on Foundations of Digital Games· pp. 1-11· 0 citations· 3 references
Computer Science
TL;DR
An LLM-driven Hearthstone agent is developed using the Sabberstone framework to evaluate several models, including GPT-4o, GPT-4o-mini, o3-mini, and GPT-5-mini, across multiple decks and prompting strategies, and results indicate that all evaluated LLMs outperform the random baseline, and GPT-5-mini achieves win rates close to the strongest numerical agents under the evaluation setting.
Abstract
This paper investigates the performance of Large Language Models (LLMs) as autonomous agents in the Hearthstone digital collectible card game. Although traditional numerical agents have demonstrated strong results in competitive game environments, the reasoning capabilities of LLM-based agents remain largely unexplored in this context. To address this gap, we developed an LLM-driven Hearthstone agent using the Sabberstone framework to evaluate several models, including GPT-4o, GPT-4o-mini, o3-mini, and GPT-5-mini, across multiple decks and prompting strategies. Our experiments compare their win rates against established numerical agents and analyze the impact of different prompting techniques, such as Chain-of-Thought (CoT), Reverse Chain-of-Thought (RCoT), ReAct, and Directed Acyclic Graph (DAG) prompting. The results indicate that all evaluated LLMs outperform the random baseline, and GPT-5-mini achieves win rates close to the strongest numerical agents under our evaluation setting. Among the prompting strategies, DAG achieved the highest win rate, while ReAct performed significantly worse. These findings highlight both the current limitations and the promising potential of LLMs for strategic reasoning and decision-making in complex, rule-based digital games such as Hearthstone. Leveraging their natural language understanding capabilities, LLMs have the potential to act as intelligent virtual companions, guiding and supporting players through the initial phases of gameplay.
A novel RAIE taxonomy along four scaling dimensions is proposed, which optimizes the entire thought process through search algorithms and self-verification, and introduces a task-oriented guideline for choosing the best TTS strategy.
Jia-Yu An, Zheng Chen, Yongcheng Jing et al.· 0 citations
Efficiency is increasingly important for Large Language Model (LLM)-based multi-agent systems (MAS), as larger models and more agents introduce substantial execution costs. Recent methods aim to make MAS cheaper by pruning agents, removing communication edges, or searching for compact structures. However, we argue that existing evaluations may overestimate their true ability to improve MAS efficiency. Reported gains are often measured under method-specific prompts and starting topologies, making them difficult to attribute to the proposed structural changes. Moreover, many reported successes appear in non-MAS-demanding settings, where a single agent or a randomly pruned system can already preserve strong performance. To study these issues, we introduce a controlled and MAS-demanding diagnostic benchmark for representative MAS efficiency methods. We evaluate methods under a shared backbone model, agent registry, and runtime, across controlled variations in topology, scale, depth, and tool use. Our analysis shows that many reported gains are setup-dependent and may arise from structural collapse, disabled tool pathways, or starting systems where random pruning already preserves accuracy, rather than robust improvements in MAS efficiency.
Jia-Mu Zhang, Ling-Xi Zhang, Peng-Jun Lu et al.· 0 citations
LLMs are increasingly deployed as orchestrators that coordinate specialized subagents to solve complex tasks through natural language. However, in many important domains like game playing and robotics, the strongest available agents are not language models. Integrating non-language agents with LLMs would require \emph{verbalization}: compressing their rich continuous representations into sparse textual summaries at each interaction step. To study whether verbalization constitutes a bottleneck, we introduce \textsc{LLAMIA-Bench}, a suite of six diverse collaborative chess tasks spanning three facets: behavioral imitation, state assessment, and natural-language explanation. Each task instantiates a well-established chess problem that neither the LLM nor the chess engine can solve alone. To solve LLM collaboration with non-language agents, we introduce \emph{latent state internalization}, which projects the subagent's continuous representations directly into the LLM's token stream as learned state tokens, with dynamic re-encoding as actions advance the environment state. Comparing internalization to verbalized integration, our experiments reveal a consistent \emph{verbalization debt}: the performance gap widens throughout training and persists as the LLM scales from 4B to 14B parameters. A single 14B model, \textsc{LLAMIA}, trained with latent state internalization, matches or exceeds task specialists and frontier models including GPT-5.1 with tool access across all benchmark tasks, and generalizes out-of-distribution where task-specific finetunes collapse
Large Language Models (LLMs) are increasingly used in simulations, either in academic settings for research studies or in industry for prototyping. Previous research has investigated the extent to which agents can mimic human behavior in socioeconomic settings; however, there is limited research on greedy decision-making by agents in simulated resource allocation environments. Furthermore, there is limited work on cross-model evaluation. Our research investigates the decision-making of ten agents across two different experimental conditions: one in which agents are able to communicate with other agents, and one in which emotional contexts are directly injected into the prompt. Based on a conceptual framework of greed, we find that agents predominantly exhibited greed-like behavior across all conditions. Interaction and self-reported social connection did not meaningfully influence the agents’ decision-making. We evaluated four different models: gpt-5-mini, gemini-3-flash-preview, claude-haiku-4-5-20251001, and grok-3-mini-fast-beta, and observed that while models differed in their self-reported connection scores, they did not differ significantly in greediness scores. The code and experimental artifacts are available at https://github.com/tiaL-ops/simCo.
Landy Rakotoarison, Fanamby T. Randriamahenintsoa· International Conference on...· 0 citations
This work extends the classical Schelling segregation model with a hybrid population: ordinary agents classify neighbors using the standard symbolic rule, while one agent delegates this task to an LLM through tool calls, providing a minimal but controlled setting where the semantic, operational, and computational behavior of LLM-based decisions can be studied inside an otherwise standard ABM.
Stefano Blando, Emanuele Guerrazzi, R. Porcedda et al.· arXiv.org· 0 citations
The growing rate at which LLM agents interact with one another raises key questions about language evolution in multi-LLM-agent settings, with implications for safety and monitorability as well as for linguistic accounts of LLMs. To address these questions, we introduce GlossoGen, a novel platform for studying multi-agent language evolution in complex scenarios. Within GlossoGen, we build the SaveVeyru scenario, which requires agents with partial information to communicate under pressure. We find that language evolution does occur between LLM agents, that the resulting languages are compositional and morphologically productive, and that they deviate from the LLMs'English prior in ways that render them incomprehensible to humans. Moreover, we identify several qualities essential to this evolution: pressure towards efficiency; the strength of the models backing the agents; and access to a"postmortem"stage in which agents can agree on linguistic conventions. Importantly, we observe that different conditions govern the transmission of language to new agents. Specifically, we find that agents learn new languages from usage alone, take an active role in this learning, and that while stronger models are required for novel language emergence, weaker models can learn an existing language once it has emerged. Taken together, our results indicate that current LLMs have the potential for cumulative cultural evolution -- previously attested only in humans -- with mixed populations of agents developing capacities that go beyond their lowest common denominator.
Elias Stengel-Eskin, Newton Sander, Carlos Bonetti et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
