Xiaodong Liu

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.

Registers are implemented as dedicated fixed-position tokens whose continuous hidden states are trained to carry reasoning progress across generation chunks, which are especially effective for bounded code generation, where correct programs usually span several chunks.

SetMIR not only boosts conversion rates by over 3% but also streamlines retrieval by cutting down query redundancy by a third.

CAMIE boosts click-through rates by nearly 19% over traditional text-based retrieval methods by aligning item embeddings with user engagement patterns.

Distilling from weaker models can enable a student to outperform its stronger counterparts, challenging the conventional wisdom of model hierarchy in AI training.

Shared descriptive patterns in skill libraries can obscure task-relevant signals, but SkillSight reveals and calibrates these biases to boost retrieval accuracy by over 20%.

GFlowRL achieves unprecedented stability and performance in large language models by eliminating the problematic learned partition function, setting a new standard for GFlowNet-style reinforcement learning.