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

Telegraph English, a structured entity-relation rewriting format for context compression, outperforms character deletion, truncation, and random sub-sampling baselines by 13–20 F1 points on MuSiQue, TwoWiki, and HotpotQA using small language models. It also beats coherent prose summaries from the same encoder on the hardest dataset, suggesting symbolic re-expression preserves entity content more densely than natural language at equal token budgets. The reasoning-depth interaction hypothesis was pre-registered and found null.