Jet-Long — dynamic RoPE scaling for long-context LLMs
arXiv paper proposes Jet-Long, a tuning-free method that extends LLM context windows beyond pretraining limits without sacrificing short-context accuracy.
• Uses “bifocal RoPE” — local window stays faithful, long-range window rescales dynamically by sequence length
• Zero-shot, no retraining required; targets RAG, codebase search, agentic workflows
• Avoids fixed-factor trade-off: aggressive rescaling breaks short context, conservative breaks long context