Search papers, labs, and topics across Lattice.
2
38
6
9
Even massive LLMs like GPT-5.4 hit a surprisingly low "depth ceiling" of just 5-7 steps when discovering multi-step planning strategies latently, suggesting that complex reasoning may require explicit training or externalization.
Command A shows how to build an enterprise-grade LLM that balances performance, efficiency, and multilingual capabilities using decentralized training and model merging.
