For decades, non-player characters (NPCs) have been driven by rule-based systems whose personalities and interactions are scripted before release. Large language models (LLMs) promise NPCs that converse openly, adapt to new situations, and behave in ways developers did not fully author, prompting a wave of research and commercial experimentation. Yet the field lacks a unified picture of what has been built, how well it works, and at what cost. We address this gap with a systematic mapping study of 377 papers from the academic literature, yielding 42 papers analysed against a main research question and ten sub-questions across four thematic axes. We provide the first quantitative characterisation of component adoption, latency, safety coverage, and reproducibility: short-term memory and world-state grounding are near-universal (60% and 69%), whereas long-term memory (31%), reflection (38%), and retrieval-augmented generation (14%) remain rare, and response times span 2.6 to 15 seconds. Finally, we identify six structural barriers to commercial adoption and offer evidence-based recommendations for developers and researchers.
Figures & tables
Fig. 1: PRISMA flow diagram illustrating the systematic identification, screening, eligibility assessment, and final inclusion of studies across the review process.
Element
Definition
Population
Interactive LLM-based NPCs in games or virtual environments
Intervention
LLM-based tools and architectures for generating realistic dialogue and autonomous behaviour, including systems with memory, retrieval grounding, planning, and tool use
Comparison
Scripted NPC AI (behaviour trees, FSMs, GOAP, utility AI), non-LLM dialogue systems, or ablated LLM variants without memory, RAG, or planning components
Outcome
Believability and human-likeness of NPC dialogue and behaviour; dialogue coherence and consistency; behaviour quality and agency; engineering feasibility (latency and cost); safety; and designer controllability
Context
Interactive games and virtual worlds with real or simulated player-agent interaction
TABLE I: PICOC framework for this systematic mapping study.
Criterion
Yes
Partial
No
QA1: Does the paper clearly describe the LLM(s) or generative model(s) used?
23
11
8
QA2: Is the LLM applied to NPCs or autonomous agents in an interactive or simulated environment?
42
0
0
QA3: Does the paper provide sufficient technical detail to understand or reproduce the approach?
12
19
11
QA4: Does the paper evaluate NPC behaviour or dialogue realism using qualitative or quantitative methods?
35
5
2
QA5: Is the study empirically validated (experiments, user studies, benchmarks, or simulations)?
27
11
4
TABLE II: Quality assessment criteria and score distribution ( n=42 ).
Architecture Pattern
# Papers
Paper List
Single-agent (prompt pipeline)
24
[ 6 , 7 , 8 , 2 , 9 , 10 , 11 ]
Multi-agent system
15
[ 4 , 12 , 13 , 14 , 15 , 16 ]
Planner + LLM
17
[ 5 , 17 , 18 , 19 , 20 , 21 , 22 ]
GOAP + LLM
1
[ 14 ]
Tool-using agent
20
[ 23 , 24 , 25 , 18 , 26 , 27 ]
BT/FSM + LLM hybrid
1
[ 28 , 14 ]
TABLE III: Architecture patterns identified in the corpus ( n=42 ). Papers may be coded under more than one pattern.
Component
# Papers
%
Short-term memory (prompt history)
25
60
World-state grounding
29
69
Rules / output constraints
29
69
Emotion / personality model
20
48
Planning module
23
55
Long-term memory (vector / summary)
13
31
TABLE IV: Frequency of grounding and control components ( n=42 ).
Fig. 2: The methods of behavior the LLM-based NPCs/Agents were programmed to perform.
Paper
Condition
Latency
GROMIT [ 23 ]
GPT-4, Sandbox
10,600 ms
GPT-4o, Sandbox
5,200 ms
GPT-4, Escape Room
5,600 ms
GPT-4o, Escape Room
2,600 ms
Player-Driven Emergence [ 24 ]
GPT-4 (game loop)
15,000 ms
Immersive Tailoring [ 29 ]
Tailoring path
7,300 ms
TABLE V: Latency measurements reported in the corpus. Papers that acknowledge latency qualitatively, but provide no figures are excluded.
Control Mechanism
# Papers
%
Personality profiles / persona
26
62
Prompt templates
38
90
Output constraints / rules
27
64
Controllable parameters (sliders,
25
60
weights, budgets)
Lore knowledge base
11
26
TABLE VI: Designer control mechanisms and their frequency ( n=42 ).
Artifact Type
# Papers
%
Source code / repository
13
31
Prompts (full text)
10
24
Dataset
13
31
Benchmark / environment
6
14
Demo video
10
24
Nothing released
18
43
TABLE VII: Artifact availability across the corpus ( n=42 ).
Open Problem
# Papers
Evaluation validity (no benchmark, small n )
33
Hallucination and factual inconsistency
26
Real-time latency
24
Long-term memory consistency
19
Controllability vs. emergent behaviour
23
Dataset and benchmark gaps
16
TABLE VIII: Open problems identified by authors across the corpus, ranked by frequency of mention ( n=42 ).
Large language model-based role-play simulation creates interactive agents that portray people, characters, or users for training, testing, and research.
Scripted and rule-based non-player characters (NPCs) in combat video games often exhibit predictable behaviors that experienced players can exploit, while reinforcement learning (RL) agents typically retain a fixed policy after training and cannot readily adapt their strategy to different opponents. We investigate a runtime strategy-selection framework in which a large language model (LLM) guides a trained RL policy without modifying its underlying behavior. To demonstrate this, we train five NPC agents with a shared PPO policy in Unity and compare a baseline configuration, in which the policy acts independently, with an LLM-augmented configuration in which a locally hosted Mistral 7B model, accessed through Ollama, reads the live game state every five seconds and assigns one of four tactical tags. We evaluate both configurations against three scripted opponent types across 600 episodes and analyze outcomes using the Mann-Whitney U test. Against a Balanced opponent that changes tactics during an episode, the LLM-augmented agents more than doubled their win rate from 11% to 24% and produced significantly longer episodes. Against an Evasive opponent, the augmented agents achieved a higher win rate and faster kills, although their shorter episode duration did not satisfy the strict hypothesis definition. Against an Aggressive opponent, the LLM's near-constant preference for encirclement was counterproductive. Analysis of 2,430 strategy selections showed that Surround was selected in 83.8% of cases regardless of opponent type, indicating limited zero-shot strategic differentiation at this model scale. These results demonstrate both the potential and limitations of LLM-guided runtime strategy selection for adaptive multi-agent game AI.
The rapid advancement of Large Language Models (LLMs) is revolutionizing AI for Games by enabling open-ended and fluid interactive storytelling. However, existing research has largely overlooked the critical challenge of maintaining long-horizon logical consistency and narrative integrity against unconstrained user interventions. To address this, we formulate this challenge as Narrative Commitment Preservation (NCP), and take interactive narrative as our testbed. We introduce NCP-Bench, a benchmark of 100 narrative environments derived from movie synopses. Each environment includes a structured narrative specification (trajectory, commitments, and initial facts) that we can automatically check throughout the interaction between the player agent and the narrator agent. Experiments across state-of-the-art LLMs reveal a substantial long-horizon consistency gap: high linguistic quality does not guarantee commitment preservation; even strong models frequently generate logically conflicting content under adversarial interventions, with the best-performing model (GPT-5.2) achieving only 42% survival rate after 20 turns and fact conflict rates ranging from 40% to 68% across models, and only isolated runs satisfying all achievement commitments within the 100-turn limit.
Yingpeng Ma, Jianhao Yan, Bei Shi +6
NLP2CT Lab, University of Macau, Macau, China · Westlake University, Hangzhou, China · Harbin Institute of Technology, Shenzhen, China +1
On a 300-persona life-simulation benchmark, pcsp achieves compositional zero-shot persona identification up to 17x above chance, Spearman rho approx 0.73 semantic-behavioral alignment, and 22x faster inference than an LLM-as-policy baseline. Life simulation games require hundreds to thousands of non-player characters (NPCs) that behave consistently with distinct personalities while remaining controllable through designer-authored natural language. Existing methods fail on constraints like persona consistency, controllability, or real-time inference. We introduce pcsp (Persona Conditioned Shared Policy), a single reinforcement learning policy conditioned on frozen LLM embeddings of free-form persona descriptions. pcsp combines once-per-NPC persona encoding, low-rank persona projection, neural persona conditioning, and a PPO + InfoNCE consistency + KL diversity training objective. Across three experimental settings, ablations show that the InfoNCE trajectory-consistency objective is load bearing: removing it collapses zero-shot persona identification to chance. External validation on Melting Pot 2.4.0 substrates confirms that our method produces persona-conditioned behavioral divergence in multi-agent strategic environments. We distinguish two senses of held-out evaluation: compositional zero-shot and vocabulary-expansion held-out. Finally, a UE5 deployment reproduces the in-engine persona-conditioning ablation at 64 agents with a low failure rate, showing that the sub-frame inference profile survives in a commercial game engine. These results prove that shared RL policies can support scalable, real-time, persona-conditioned NPC control.