Preprint reports agent context compression with up to 50.68% fewer attention pairs

Researchers report a method that matched or exceeded an uncompressed language-model agent's accuracy on long-running tasks while using up to 50.68% fewer end-to-end attention pairs, a measure of attention computation. The arXiv preprint introduces Agentic Soft-Token Compression, which requires no additional training. The method uses a model's internal recurrent processing to carry previously absorbed information forward as compact internal representations while preserving exact values as ordinary text tokens. This reduces computation in later layers and produces reusable caches across the agent's task. The approach uses a hybrid Gated Delta Network and attention architecture; the reported accuracy and computation results come from the tasks tested.