Tools Need Not Speak in Text: Latent Context Compression for Agents
Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substantial context overhead and dilute the agent'
- Upvotes6