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

Short summary

Researchers introduce Telegraph English, a symbolic rewriting format that compresses context for small language models while preserving reasoning evidence. In experiments on multi-hop QA benchmarks, this format outperforms natural language and summary baselines by 13-20 F1 points, showing symbolic representation preserves entity information more densely than prose.

• •Telegraph English converts passages into readable entity-relation statements, reducing tokens while preserving reasoning structure
• •Outperforms deletion, truncation, and random-sampling compression baselines by 13-20 F1 points across three benchmarks
• •Symbolic format preserves entity content more densely than coherent summaries at matched token budget

Generated with AI, which can make mistakes.