Скачать или смотреть QwenLong-L1.5: Long-Context LLM Training

In this AI Research Roundup episode, Alex discusses the paper: 'QwenLong-L1.5: Post-Training Recipe for Long-Context Reasoning and Memory Management' The episode breaks down how QwenLong-L1.5 proposes a post-training recipe to boost long-context reasoning and memory in large language models, starting from Qwen3-30B-A3B-Thinking. Alex explains the long-context data synthesis pipeline that decomposes documents into atomic facts, builds cross-document knowledge graphs, and generates complex multi-hop tasks with globally distributed evidence. The discussion covers how a multi-agent self-evolve loop and rule-based checks curate a high-quality RL dataset for contexts up to and beyond 64K tokens. Alex also explores their stabilized reinforcement learning approach using Group-Relative Policy Optimization and Adaptive Entropy-Controlled Policy Optimization to keep training stable on ultra-long sequences, including beyond 1 million tokens. Paper URL: https://arxiv.org/abs/2512.12967 #AI #MachineLearning #DeepLearning #LongContext #LLM #ReinforcementLearning #LanguageModels #MemoryManagement

Resources:
GitHub: https://github.com/Tongyi-Zhiwen/Qwen...
Hugging Face model: https://huggingface.co/Tongyi-Zhiwen/...