Telegraph English Beats Summarization at Context Compression by 13-20 F1 Points

Thirteen to twenty F1 percentage points: that's what switching from a coherent prose summary to a structured symbolic format buys you on multi-hop QA, at the exact same token budget.

A new paper on arXiv proposes Telegraph English, a compression strategy that rewrites retrieved passages into terse entity-relation statements. Think "Apple - founded by - Steve Jobs" instead of "Apple Inc. was founded by Steve Jobs in 1976." The key claim: readable symbolic re-expression packs more reasoning evidence per token than any natural language summary can manage.

Three Datasets, Four Baselines, One Winner

The authors ran controlled experiments on MuSiQue, TwoWiki, and HotpotQA, matching token budgets across all methods. Telegraph English faced three blunt compression baselines (character-level deletion, truncation, random sub-sampling) plus a coherent prose summary produced by the same encoder used to generate the symbolic format. On every dataset, the symbolic format won outright. Gains ranged from 13 to 20 F1 points - no overlap, no tie.

A pre-registered hypothesis predicted the advantage would grow with reasoning depth. That hypothesis came back null. The benefit doesn't depend on how many hops a question requires; it's a general property of the representation itself.

Why Symbols Beat Summary

The interpretation is straightforward: readable symbolic re-expression preserves entity content more densely. A coherent summary has to maintain narrative flow, which burns tokens on function words and transitions. Telegraph English skips all that. It keeps the entities, the relations, and the structure, and discards everything else. For a small language model trying to answer a question that requires stitching together facts from two passages, those saved tokens matter.

The paper stops short of claiming this works for large models or open-ended generation. But for retrieval-augmented QA with constrained context windows, the message is clear: if you want to cram more facts into fewer tokens, don't summarize - symbolically re-express.

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