This paper introduces Semantic Soft Bootstrapping (SSB), a novel self-distillation technique for enhancing long-context reasoning in large language models without relying on reinforcement learning. SSB uses the same base LLM as both teacher and student, leveraging semantically rich contexts to generate robust explanations and distilling logit-level supervision to improve reasoning capabilities efficiently.

This paper introduces Arbitrage, a novel step-level speculative generation framework designed to enhance the efficiency of Large Language Models (LLMs) in reasoning tasks. It dynamically routes between a fast draft model and a more capable target model based on the expected quality advantage, significantly reducing computational waste and inference latency while maintaining accuracy.