Tag: #context-compression5 results
Llama 3.1 8B with structured chain-of-thought and graph context compression matches or exceeds Llama 3.3 70B on multi-hop QA benchmarks without fine-tuning. Retrieval succeeds 77-91% of the time, but reasoning is the bottleneck. Tested on HotpotQA, MuSiQue, and 2WikiMultiHopQA with 12x cost savings on Groq.
Cursor has trained its new AI model Composer using reinforcement learning to enable self-summarization. This feature supports agentic coding for long-horizon tasks by compacting code context with internal tests. It enhances efficiency in handling extended coding sessions.
LangChain added a tool to Deep Agents SDK (Python) and CLI, enabling models to autonomously compress context windows at optimal times. This reduces information in agents' working memory by replacing older messages.
Qwen-Code releases v0.15.9-nightly with fixes for CLI argument validation, OpenAI request logging, and Mistral content filtering. New features include reactive compression on context overflow, QWEN_HOME env var support, and improved slash command discovery. Enhancements also cover VSCode UI for message editing and better handling of long conversations.
Qwen-Code's preview release v0.15.10 includes bug fixes for CLI validation, OpenAI logging, and Mistral content filtering. New features add reactive compression for context overflow and autoSkill for background project skill extraction. CLI enhancements improve slash command discovery and settings preservation.