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

2026-06-12 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A new context compression technique, Telegraph English, has been introduced for multi-hop question answering with small language models. This readable symbolic format rewrites retrieved passages into structured entity-relation statements, aiming to preserve reasoning evidence at a lower token cost. In controlled experiments conducted on MuSiQue, TwoWiki, and HotpotQA datasets, Telegraph English consistently outperformed three matched-budget compression baselines—character-level deletion, truncation, and random sub-sampling—achieving gains of 13 to 20 F1 percentage points across all datasets. It also demonstrated superior performance against a coherent prose summary generated by the same encoder on the most challenging dataset. A pre-registered hypothesis regarding advantage growth with reasoning depth was found to be null. These results suggest that readable symbolic re-expression more densely preserves entity content compared to natural language or coherent summarization within the same token budget.

Key takeaway

For Machine Learning Engineers optimizing context windows in small language models for multi-hop question answering, you should investigate symbolic re-expression techniques. Implementing a structured format like Telegraph English, which converts passages into entity-relation statements, can yield substantial performance improvements, potentially boosting F1 scores by 13-20 percentage points. This approach offers a more token-efficient way to preserve critical reasoning evidence than traditional summarization or truncation methods.

Key insights

Telegraph English's symbolic re-expression significantly improves context compression for multi-hop QA over natural language at matched token budgets.

Principles

• Symbolic re-expression enhances information density.
• Structured entity-relation formats improve context preservation.
• Token budget matching is vital for compression evaluation.

Method

Telegraph English rewrites retrieved passages into structured entity-relation statements, creating a readable symbolic format that preserves reasoning evidence at lower token cost for multi-hop QA.

In practice

• Implement symbolic re-expression for context compression.
• Employ entity-relation statements to reduce token costs.
• Benchmark compression against matched-budget baselines.

Topics

• Context Compression
• Symbolic Re-expression
• Multi-hop QA
• Small Language Models
• Entity-Relation Extraction
• F1 Score

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Beyond Uniform Tokens: Adaptive Compression for Time Series Language Models — Adaptive token compression for time series LLMs improves efficiency and performance by leveraging asymmetric information structures.
• Query-focused and Memory-aware Reranker for Long Context Processing — A new reranking framework uses attention scores for listwise relevance estimation, outperforming existing methods in long context processing.
• Analysing Lightweight Large Language Models for Biomedical Named Entity Recognition on Diverse Ouput Formats — Lightweight LLMs offer competitive biomedical NER performance, with specific output formats proving more effective than multi-format tuning.
• EASE-TTT: Evidence-Aligned Selective Test-Time Training for Long-Context Question Answering — Retrieved evidence can directly supervise query-side attention adaptation for improved long-context QA in smaller LLMs.
• TokenMizer: Graph-Structured Session Memory for Long-Horizon LLM Context Management — TokenMizer uses a knowledge graph to structure LLM session history, significantly reducing context tokens and improving recall.
• Lossless Prompt Compression via Dictionary-Encoding and In-Context Learning: Enabling Cost-Effective LLM Analysis of Repetitive Data — LLMs can learn dictionary encodings in-context, enabling lossless prompt compression without fine-tuning and preserving analytical accuracy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.