Text is the bottleneck: a small-scale replication of LLM game playing in Riichi Mahjong
What happened
Mingge Xu and co-authors replicate a study of LLMs playing Riichi Mahjong (Japanese Mahjong). Their conclusion is that, when data and labels are matched, the text prompt is the main bottleneck. According to the paper, the replication was run on a very small budget: models with no more than 2B parameters, a single 16 GB consumer GPU, and about five GPU-days of compute. The study being replicated is the Riichi Mahjong section of arXiv:2509.01328, "Can Large Language Models Master Complex Card Games?" The paper releases its code, configurations and evaluation protocol, but not the trained model weights. The authors say this arrangement follows Tenhou's terms.
AI-written from the coverage · updated 26 d ago
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Board & Classic Game AIPickText is the bottleneck: a small-scale replication of LLM game playing in Riichi Mahjong
A small-budget replication of LLM Riichi Mahjong gameplay research