Universal YOCO: Recursive Depth Scaling for Efficient Test-Time Compute in Reasoning LLMs
arXiv·medium signal
Combines YOCO decoder-decoder architecture with recursive computation to address the KV cache inflation problem as model depth scales during test-time reasoning. Standard Transformers struggle to scale inference compute efficiently — conventional looping inflates KV cache proportionally. YOCO-U achieves a synergistic effect where recursive depth is added without proportional memory cost, directly addressing the compute bottleneck in reasoning model deployment.