SpireBench: Evaluating Long-Horizon Strategic Play of Agentic Models with Slay the Spire
Quick Answer
SpireBench introduces a long-horizon benchmark for evaluating agentic large language models (LLMs) in the game Slay the Spire, revealing that while frontier models perform comparably to average human players, they lag behind experts, particularly in risk assessment over extended play.
Quick Take
The study highlights that often misjudge health preservation versus strength growth, indicating a need for improved strategic reasoning in complex tasks.
Key Points
- SpireBench tests 20 LLMs at the highest difficulty of Slay the Spire.
- Frontier models match average human player performance but fall short of experts.
- Models misjudge risk, prioritizing health preservation over strength growth.
- The benchmark reveals gaps in strategic reasoning for long-horizon tasks.
- SpireBench provides tools to analyze agent failures in complex scenarios.
DeepSignal Analysis
What happened
SpireBench has been introduced as a benchmark for evaluating long-horizon strategic play in the game Slay the Spire using large language models (LLMs). The study found that while these models perform similarly to average human players, they significantly underperform compared to expert players, particularly in risk assessment during extended gameplay.
Key evidence
- SpireBench evaluates 20 large language models at the highest difficulty of Slay the Spire, revealing that frontier models match average human performance but fall short of expert levels.
- The research indicates that LLMs misjudge risk, often prioritizing health preservation over strength growth, which affects their long-term success in the game.
- The study highlights that models do not improve from their experiences across runs, as their recorded memories are often incorrect and overly cautious.
Why it matters
The introduction of SpireBench addresses a gap in evaluating the strategic reasoning capabilities of LLMs over long horizons, which is crucial for developing more effective AI systems. Understanding the limitations of current models in complex tasks like Slay the Spire can guide future research and improvements in AI decision-making processes.
Paper Resources
📖 Reader Mode
~2 min readAbstract:Agentic large language models now tackle complex tasks, yet strategic reasoning over long horizons is under-studied: most agentic benchmarks are short and episodic, and performance on long tasks is hard to dissect into stored knowledge, reasoning and luck. We introduce SpireBench, in which models play the complete commercial deck-builder Slay the Spire (StS), where early decisions decide long-term success. Across 20 large language models (LLMs) at the game's highest difficulty, frontier models beat heuristic baselines and match the average human player but remain far behind experts. This gap widens as the run lengthens: the strongest models win the first act as often as experts, then fall behind. What separates strong models from weak is how well they play each fight; all of them misjudge risk over a run, preserving health rather than spending it to grow stronger. A comprehensive, self-consistent renaming of every game term shows that part of their performance rests on prior knowledge of the game, and models do not improve from the memories they write across runs, which are often wrong and overly cautious. Our harness is the first to give a model the complete original game, and we compare it with existing harnesses. Overall, SpireBench offers an unsaturated long-horizon benchmark with tools to locate where and why agents fail.
| Subjects: | Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.12596 [cs.AI] |
| (or arXiv:2610.12596v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.12596 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Scott Lowe [view email]
[v1]
Thu, 8 Oct 2026 18:18:19 UTC (2,011 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →VEPA: Verified Evolving Formal Policy for Lifelong Agent Guardrails
VEPA introduces a framework for verified evolving formal policies for LLM agents, enhancing safety by adapting guardrails based on deployment feedback. It outperforms existing methods across benchmarks like ASSEBench-Safety and ASSEBench-Security, ensuring deterministic policy enforcement even under challenging conditions.