Context Compression Is Not One Thing: Readable Symbolic Re-expression vs. Coherent Summary at Matched Budget

The paper presents a novel approach to context compression for multi-hop question answering using a symbolic format called Telegraph English, which rewrites passages into structured entity-relation statements. In experiments on datasets like MuSiQue, TwoWiki, and HotpotQA, this method achieved 13 to 20 F1 percentage point improvements over traditional compression techniques and coherent prose summaries, demonstrating that symbolic re-expression can retain more entity content at a lower token cost. This has implications for practitioners seeking efficient methods to enhance reasoning in language models without increasing computational overhead.