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

2026-06-12

AI Summary

This work addresses the limitations of small language models in multi-hop question answering, where constrained context lengths hinder the retention of critical reasoning evidence after compression. To overcome this, the authors propose Telegraph English—a human-readable, symbolic format that rewrites retrieved passages into structured entity–relation statements, enabling denser encoding of inferential information within a fixed token budget. This approach provides the first systematic demonstration that readable symbolic re-expression outperforms both fluent abstractive summarization and conventional compression strategies. Evaluated on MuSiQue, TwoWikiMultihop, and HotpotQA, the method yields substantial performance gains, improving F1 scores by 13–20 percentage points, with the most pronounced advantages observed on the most challenging datasets.

Abstract

We study context compression for multi-hop question answering with small language models. We propose Telegraph English, a readable symbolic format that rewrites retrieved passages into structured entity-relation statements, preserving reasoning evidence at lower token cost. In controlled experiments on MuSiQue, TwoWiki, and HotpotQA, Telegraph English outperforms three matched-budget compression baselines (character-level deletion, truncation, and random sub-sampling) on every dataset, with gains of 13 to 20 F1 percentage point. It also outperforms a coherent prose summary produced by the same encoder on the hardest dataset. A pre-registered depth-interaction hypothesis is null: the advantage does not grow with reasoning depth within datasets. We interpret these results as evidence that readable symbolic re-expression preserves entity content more densely than either natural language or coherent summarization at matched token budget.

Problem

• context compression
• multi-hop question answering
• token budget
• reasoning evidence
• small language models

Innovation

• context compression
• symbolic re-expression
• multi-hop question answering
• entity-relation statements
• token-efficient representation