Distilling Game Code World Model Generation into Lightweight Large Language Models

arXiv 2026 48.3 method

TLDR

Distills game code world model generation into smaller LLMs via SFT and RLVR, improving syntactic correctness and rule adherence.

Reasoning

Strengths include a novel distillation pipeline combining SFT and RLVR with a verification framework, and a curated dataset of 30 games. Weaknesses are the limited scope (only one 3B model, 30 games) and lack of direct comparison to frontier models.

Read-first score

Read-first score 48.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 15%

76.8

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.76787073

Topical relevance 25%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,experiment

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,dataset

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 277.

Keyword Scores

world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

interactive world model

generative world model

video world model

Deep Analysis

Innovations

Distillation of Game Code World Model generation capabilities into smaller LLMs through post-training
Curated dataset of 30 games spanning perfect and imperfect information games
Verification framework that evaluates generated code against structural and semantic game properties
Post-training pipeline combining Supervised Fine-Tuning (SFT) with Reinforcement Learning with Verifiable Rewards (RLVR)

Methodology

The authors introduce a curated dataset of 30 games, a verification framework for structural and semantic properties, and a post-training pipeline that combines Supervised Fine-Tuning (SFT) with Reinforcement Learning with Verifiable Rewards (RLVR). They experiment with the Qwen2.5-3B-Instruct model, evaluating syntactic correctness and execution-level adherence to game rules.

Key Results

SFT increases syntactic correctness, while RLVR improves execution-level adherence to game rules, making Qwen2.5-3B-Instruct more capable of generating valid Game Code World Models in both perfect and imperfect information games.

Resources

Paper