LLM game characters: a research map reveals gaps in speed, safety and memory

Language models offer game developers a way to move beyond prewritten dialogue branches. But generating a conversation is different from delivering a reliable character inside a running game. A systematic mapping study by Dilan Mian, Jeremy Bradbury, Yann-Gaël Guéhéneuc, Fabio Petrillo and Cristiano Politowski surveys that gap. Posted to arXiv on November 2, 2026, the source is an accepted manuscript; its landing page identifies it as published in IEEE Transactions on Games Early Access.

The authors find considerable architectural experimentation, alongside weaknesses in evaluation, reproducibility and runtime safety. Their results describe the selected research literature, not a comprehensive audit of commercial games or proof that language-model characters outperform scripted alternatives.

What the review covers

The researchers searched ACM Digital Library, IEEE Xplore and Scopus for publications from January 2022 through February 2026. They screened 377 records, assessed 50 candidate papers and retained 42 after applying a quality threshold. Two reviewers independently screened records and scored quality; one extracted study information and another checked it.

The resulting corpus spans 2023 to early 2026 and uses a broad definition of non-player characters. Only 20 papers concern player-facing NPCs; the other 22 examine broader simulation or virtual-agent systems. These include Minecraft task agents and social simulations whose architectures may inform game-character design, even when their immediate users are researchers rather than players.

That distinction matters. Completing an open-world task and sustaining a convincing conversation are different objectives. The review separates these groups in a sensitivity analysis rather than treating every result as evidence about conversational game characters.

Memory and grounding are unevenly developed

The most common architecture is a single-agent prompt pipeline: a model receives information such as a persona, recent conversation, world state and rules, then generates dialogue or an action description. Other approaches add planners, multiple agents, specialised model modules or conventional game-AI controllers.

Across the 42 papers, 25 use short-term memory and 29 use world-state grounding—60% and 69%, respectively. These are majority practices, not universal capabilities. Long-term memory appears in 13 papers, reflection or self-critique in 16, and retrieval-augmented generation or skill retrieval in six.

The authors identify persistent memory and identity as major unresolved problems. Summaries can lose detail, context windows fill, and characters can contradict earlier interactions. The review reports no solution that achieves sustained cross-session coherence. It also notes that text inventories of nearby objects often omit spatial relationships needed for richer world understanding.

Response time is not consistently measured

Although 32 papers discuss real-time feasibility, only seven provide numerical latency measurements. Reported end-to-end responses for cloud-hosted frontier-model configurations range from roughly 2.6 to 15 seconds. That is not the range for every system: the latency table also includes faster specialised inference or selection paths and slower behaviour-tree generation.

Researchers attempt to reduce delays through smaller models, local deployment, caching, parallel processing and offline generation. One approach generates dialogue variants before play and selects among them at runtime, removing live model inference from that interaction loop.

These choices introduce trade-offs. Smaller models may reduce resource demands while weakening characterisation or task performance. The authors call for controlled measurements of that balance. They also find no reviewed paper that quantifies operating costs at player scale, so commercial affordability remains unestablished by this corpus.

Safety and evidence remain limiting factors

Only two papers implement runtime guardrail components as system elements. Persona prompts and role constraints are much more common, but the review describes them as vulnerable to adversarial input. Player jailbreaking is documented in the reviewed literature, making prompt-based control an incomplete answer to deployment safety.

Comparisons are also difficult. Sixteen papers include no baseline, and studies use incompatible measures of task success, usability, player experience and believability. User studies typically involve small, demographically narrow samples. Individual reported gains therefore do not establish a general ranking of architectures.

Reproduction is another obstacle: 13 papers release source code, ten provide full prompts, and 18 release no artifacts. Changing or retired model APIs further complicate replication.

The authors recommend shared benchmarks, clearer reporting and stronger cross-session memory, while favouring hybrid designs as a promising commercial direction—not a demonstrated general solution. Their conclusions remain bounded by database coverage, interpretive coding and a fast-changing technology landscape. The review maps where evidence is missing as much as where prototypes show promise.