Tag
This paper proposes GittinsEval, a cost-aware Bayesian bandit framework for efficient LLM evaluation that significantly reduces costs while maintaining high performance by adaptively selecting configurations.
This paper demonstrates that volatility and stochasticity, both sources of uncertainty, drive optimal exploration in opposite directions: volatility increases exploration while stochasticity suppresses it. The authors extend the Gittins index framework to Gaussian state-space bandits and introduce CAUSE, a closed-form exploration bonus that outperforms standard strategies.