Tag
This paper systematically studies LLM-based verification of Olympiad-level mathematical proofs, introducing a human-curated USAMO 2025 dataset (Trojan-Bench), an adversarial proof generation pipeline (TROJAN), and MABGrader, a Multi-Armed Bandit grading framework that outperforms ProofGrader by 26.4% in Quadratic Weighted Kappa across 20 evaluated models.
该论文提出将战略选择建模为叠加在基础博弈之上的元博弈,评估了5种博弈和10个大模型,发现多臂老虎机方法比基于嵌入的选择器更可靠,但仍未超过固定模型基线,揭示了当前方法应对动态对手的不足。
The paper proposes a cost-aware algorithm for identifying the best large language model using dueling feedback in a multi-armed bandit framework, demonstrating optimal cost and improvements over existing methods.
The paper introduces Odds-Ratio Thompson Sampling (OR-TS), a method for batched multi-armed bandits that uses contrast-based updates to handle varying common levels, showing improved regret over absolute-rate memory in simulations and real experiments.
This paper proposes a cost-aware framework using a multi-armed bandit controller to adaptively select prompting strategies for LLM essay scoring, reducing LLM calls by 78.4% while maintaining accuracy.
This paper formalizes multi-model human evaluation as a best-arm identification problem in a multi-armed bandit setup, adaptively allocating annotation effort to focus on competitive models and improve ranking discrimination.
MiMo-V2.5-Pro-UltraSpeed is a fast multimodal model that can generate an illustrative video of reinforcement learning based on multi-armed bandits.