📄 Telegraph English: Semantic Prompt Compression via Structured Symbolic Rewriting

Summary

Prompt compression usually works by deleting tokens. LLMLingua-2, for example, trains a classifier to remove low-importance tokens at a fixed ratio. That approach can break coreference and logical links, and it leaves no structure behind. The authors propose Telegraph English (TE), an LLM-driven rewrite of a passage into a terse, upper-case dialect. Each line holds one atomic fact. Relations are written with about 40 fixed symbols, such as causation, implication, contrast and increase. Tags cover time, modality, semantic roles and scope. Because every line is a single fact, the compressed output also works as a set of semantic chunks for retrieval and context management. The authors test TE on 4,081 chunks drawn from LongBench-v2, using a multiple-choice QA protocol across five OpenAI models. TE beats LLMLingua-2 at comparable compression on every model and task. With GPT-4.1 it keeps 99.1% key-fact accuracy, and its lead grows on small models and on fine details. This matters for cheap production pipelines and multi-step agents, where TE text can stay compressed from stage to stage.

Key points

• Grammar: Version 5 of the grammar is a 430-line spec that also serves as the compressor's system prompt. Its priorities, in order, are:

  • fidelity first;
  • one claim per line;
  • upper case unless case carries meaning;
  • about 5× compression only when feasible.

  At most three symbols may appear in a row.

• Self-checking: A 12-point quality checklist and a six-pass reasoning order run inside a single call. The passes are: concepts → claims → relations → redundancy → numbers → citations.

• Data and compression: The authors kept 339 of 503 LongBench-v2 documents, from the Single-Doc QA, Multi-Doc QA and Long-Dialogue categories. They split these into chunks of at most 1,000 words using NLTK, giving 4,081 key-fact and 801 fine-detail QA items. Compression used o4-mini.

  • The mean ratio was 0.585, a 41.5% reduction, with a range of 0.13–1.57.
  • Dense short inputs may expand rather than shrink.
  • LLMLingua-2 was run at retention rates of 0.50 and 0.33.

• Results: On fine-detail questions with smaller models, such as GPT-4.1-nano and GPT-4o-mini, TE leads by up to 11 points. In a modeled five-step agent pipeline, cost per 1K calls is about $16 for TE, $33 for LLMLingua-2 and $40 uncompressed.

• Limitations:

  • English only.
  • Only OpenAI models, which also generated the QA.
  • Dynamic context management is not benchmarked.
  • The only baseline is LLMLingua-2.

Interesting ideas

• I1. Writing exactly one fact per line makes compression and indexing the same step, so each line can be retrieved, updated or pruned on its own without re-running the compressor.
• I2. Tags such as AGENT:, PATIENT:, INSTRUMENT:, PAST:/NOW:/FUTURE: and CONF= reduce verbose framing to one marker while staying machine-readable.
• I3. The compression ratio is allowed to follow the information density of each input, even expanding dense inputs, instead of being forced to a fixed target.
• I4. Key facts are evaluated separately from fine details, where a question can hinge on a single number such as 4.8% versus 4.3%.
• I5. Surface forms are fixed at the grammar level, such as VAR=VALUEUNIT, with numbers never abbreviated and always tied to their unit, which removes a whole class of factual errors.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (back): Our text → graph → text compressor could serialize the meaning graph as one addressable fact per line, so parts of it can be dropped or edited without regenerating the whole text.
• I2 (frames): This suggests a compact, human-auditable linear format for frames, combining VerbNet roles with tense and modality, that is useful for debugging and for feeding frames to LLM teachers.
• I3 (back): We should evaluate our compressor on fidelity first and treat expansion on dense inputs as acceptable, not as a bug.
• I4 (learn): We could build negative-control distractors that change only one number, unit or attribution, to check whether our pipeline really preserves details.
• I5 (world): The measures and values layer could normalize quantities into one canonical form attached to their entity, which would make holdings, maths word problems and comparisons more reliable.