Telegraph English boosts multi-hop QA by 13–20 F1 points with symbolic compression

Researchers from multiple institutions (Bei, Arbuzov, Dong, Kalaev, Shvets) introduced 'Telegraph English' — a symbolic re-expression format for context compression in multi-hop question answering. Instead of generating coherent prose summaries, their method rewrites retrieved passages into structured entity-relation statements (e.g., 'Einstein → born → Ulm'). This format preserves reasoning evidence at a lower token cost while remaining human-readable. In controlled experiments on three datasets (MuSiQue, TwoWiki, HotpotQA), Telegraph English consistently beat three matched-budget baselines (character-level deletion, truncation, and random sub-sampling) by 13 to 20 F1 percentage points. It also outperformed a coherent prose summary produced by the same encoder on the hardest dataset. Surprisingly, a pre‑registered depth‑interaction hypothesis was null: the advantage did not increase with reasoning depth within datasets. The authors interpret this as evidence that readable symbolic re-expression preserves entity content more densely than natural language or coherent summarization at any matched token budget.

This work challenges the common assumption that coherent, fluent summaries are always the best way to compress context for LLMs. Instead, Telegraph English shows that a structured, almost database-like representation can retain more factual density and reasoning chains without sacrificing readability. For practitioners building retrieval-augmented generation (RAG) pipelines or multi-hop QA systems, this suggests that context compression should be task-aware: symbolic formats may outperform natural-language summaries, especially when the downstream task requires precise entity linking and multi-step inference. The code and data are not yet publicly linked, but the results are robust and reproducible on standard benchmarks.