Parallel Context Compaction for Long-Horizon LLM Agent Serving

In three linesPaper introduces parallel context compaction for long-horizon LLM agents to address latency and unpredictability of sequential summarization. Enables fine-grained control over summary volume and targeted prompt engineering per block. Evaluated on HotpotQA and LoCoMo benchmarks across 8B-120B models (dense and MoE architectures).

Summary generated by Claude — human-verified