LongCat-2.5 scales an end-to-end agentic RL pipeline across data, environments, algorithms and harnesses for long-horizon tasks, and its MOPD training retains the previous model's expert groups:
Per-cut prefixes and segment-capped student rollouts keep long responses stable, while shared prefill, KV-cache reuse and Tree Attention keep sampling and training cheap.
Meituan LongCat-2.5 scales agentic RL for long-horizon tasks
AIMeituan LongCat says LongCat-2.5 scales an end-to-end agentic reinforcement learning pipeline spanning data, environments, algorithms and harnesses for long-horizon tasks. Its MOPD training retains the previous model's expert groups.
@Meituan_LongCat
Part of a thread · earlier post
Source details
Meituan LongCat · x.comPublished
Original title: LongCat-2.5 scales an end-to-end agentic RL pipeline across data, environments, algorithms and harnesses for long-horizon tasks, and its ...