Efficient LLM Reasoning via Variational Posterior Guidance with Efficiency Awareness

Variational posterior guidance framework reduces LLM reasoning chain length by proving posterior distributions guided by reference answers achieve higher expected utility than prior policies.