Introducing LongCat-2.0: Trillion-Parameter MoE Large Language Model
1.6 trillion parameter MoE architecture with 1M long context, trained on domestic superpod accelerators, excelling in coding and agent tasks.
LongCat Research
GUI-CIDER: GUI Agent Mid-Training via Causal Internalization and Density-Aware Resampling
Proposes GUI-CIDER mid-training method that explicitly internalizes GUI world knowledge into models through causal internalization and density-aware sample reselection.
Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments
Identifies user noise and tool noise as two major sources of interaction noise, proposing training in noisy environments to enhance LLM agent robustness.