LLMs Relearn Telegraphese: Compression Boosts Accuracy

1 reports archived1 independent sourcesupdated 10/7/2026, 12:04:36 PM

Synthesis & Latest Updates

1 Sources Cross-Validated

A benchmark demonstrates that forcing LLM outputs into 'telegraphese' significantly reduces token usage while improving cross-model reading accuracy. The study provides open-source tools to reproduce these findings, highlighting the efficiency of minimalist styles over natural language verbosity.

LATEST/A benchmark demonstrates that forcing LLM outputs into 'telegraphese' significantly reduces token usage while improving cross-model reading accuracy. The study provides open-source tools to reproduce these findings, highlighting the efficiency of minimalist styles over natural language verbosity.

TIMELINECoverage timeline

Total 1 reports · Latest first

Hacker News AIT2·Oct 7, 12:04 PM68 pts
LLMs Relearn Telegraphese: Compression Boosts Accuracy
Original: Write Like It's 1866: LLMs Relearn Telegraphese
- Telegraphese compression reduces token consumption by ~30%, with significant variance across models
- Cross-model recovery ratios exceed 1.0, indicating higher accuracy than plaintext controls
- Some models (e.g., GPT-5-mini) incur double costs due to inability to disable reasoning chains during compression
View Evidence Original Source