Shorthand for Thought: Compressing LLM Reasoning via Entropy-Guided Supertokens
The episode discusses the paper "Shorthand for Thought," which introduces "supertokens" to compress Large Language Model reasoning. By bundling predictable, low-entropy structural phrases into single tokens via lightweight fine-tuning, the method reduces reasoning token counts by 8.1% and latency by 10.5% while maintaining human-readable logic and accuracy.