adu2021 Skill Summary
QwenLong-L1.5 introduces a comprehensive post-training system combining three technical contributions: (1) long-context data synthesis generating challenging reasoning tasks requiring multi-hop evidence grounding, (2) stabilized RL methods including task-balanced sampling and adaptive entropy-controlled policy optimization, and (3) memory-augmented architecture handling ultra-long sequences through iterative memory-based processing for sequences exceeding 4M tokens.
When To Use
- Building models that reason over extremely long documents (4M+ tokens)
- Projects requiring multi-hop reasoning across globally distributed evidence
- Scenarios where retrieval-augmented generation would be too slow or complex
- Research exploring memory-augmented LLM architectures for long-context tasks
When NOT To Use
- Inference scenarios with strict latency requirements (long-context processing has inherent delays)
- Short-context tasks where the overhead of memory architecture doesn't pay off
- Projects without computational resources to fine-tune large models with RL
- Domains where simpler retrieval systems already provide sufficient performance
Core Technique
Three interconnected innovations enable long-context reasoning at scale:
1. Long-Context Data Synthesis Pipeline Develops systematic framework generating challenging reasoning tasks. Deconstructs documents into atomic facts and relationships, then programmatically composes verifiable questions. Includes three specialized methods: knowledge-graph-guided multi-hop reasoning, structural tabular data engines for numerical problems, and multi-agent self-evolving framework for general tasks.
2. Stabilized Reinforcement Learning Methods Overcome training instability in long-context RL through:
- Task-balanced sampling with task-specific advantage estimation mitigating reward distribution bias
- Negative gradient clipping targeting high-entropy tokens in failed responses
- Adaptive Entropy-Controlled Policy Optimization (AEPO) "dynamically regulating exploration-exploitation trade-offs" by masking negative-advantage sequences when entropy exceeds target thresholds
3. Memory-Augmented Architecture For sequences exceeding 4M tokens, implement memory agent framework combining single-pass reasoning within 256K context window with iterative memory-based processing through multi-stage fusion RL training.
Implementation Notes
Implement multi-hop data synthesis pipeline starting with document decomposition. Set up stabilized RL training with task-balanced sampling and adaptive entropy control. For ultra-long sequences, implement memory agent framework with iterative processing. Apply progressive training from shorter to longer context windows.
References
- Original paper: QwenLong-L1.5 (Dec 2025)
- Long-context language model training techniques
- Adaptive entropy-controlled policy optimization