Search papers, labs, and topics across Lattice.
2
3
3
0
GFlowNet sampling accuracy is fundamentally limited by the square root of balance loss and a parameterization incompleteness residual, with convergence rates established for diverse backward policy update strategies.
The DeepSeek-V4.1-Flash model, a multimodal Mixture-of-Experts model with 552B backbone parameters and support for contexts of up to one million tokens, is introduced, substantially improving cost efficiency for agentic workloads and pushing the limits of KV cache compression.
