LLM game playing bottlenecked by text, replication study finds

Researchers have successfully replicated LLM game-playing capabilities in Riichi Mahjong with significantly reduced computational resources, using models under 2B parameters and a single consumer GPU. Their findings indicate that text is a bottleneck, as a convolutional board encoder outperformed LLMs trained on the same data. Training models to mimic a teacher opponent improved agreement with the teacher more than increasing parameters, but this did not translate to stronger playing performance at the table. The study also identified and corrected a mismatch between the original paper and its code regarding the winner filter. AI

IMPACT Suggests that alternative architectures may be more efficient for game-playing AI than text-based LLMs.

RANK_REASON Academic paper detailing research findings on LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Can Large Language Models Master Complex Card Games?
• Riichi Mahjong

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →