Telegraph English tests symbolic context compression for multi-hop QA

A new arXiv paper argues that context compression is not a single problem. The authors introduce Telegraph English, a readable symbolic format that rewrites retrieved passages into structured entity-relation statements.

In multi-hop question answering tests, the format preserves reasoning evidence at lower token cost and outperforms matched-budget baselines such as truncation and random subsampling. It also beats a coherent prose summary on the hardest dataset.

The result is relevant for RAG and agent systems that need to fit evidence into limited context while preserving the structure required for reasoning.