Tag
该论文提出将 Transformer 网络解释为混合概率推理过程的可学习近似：自注意力对应在输入条件化、动态结构的马尔可夫随机场上执行 max-product 信念传播，而层间深度演进则对应隐式的期望最大化（EM）过程。基于此视角，作者提出 meta-model 范式，将自注意力、混合专家和门控单元等机制统一起来。
This paper connects score matching, maximum likelihood, and expectation-maximization in mixed linear regression, providing theoretical guarantees and analyzing gradient behavior under diffusion noise levels.
Introduces Cluster-Weighted EDMD, a data-driven method that jointly learns a partition and per-cluster Koopman operators via expectation-maximization, improving prediction accuracy over standard EDMD on classical dynamical systems.
This paper proposes a framework for Markov chain choice models with panel data, including estimation via novel EM algorithms that leverage partial-ordering preference information, personalized choice prediction, and assortment optimization. Experimental results on synthetic data and the sushi dataset show improvements over traditional methods.
This paper identifies a mismatch between training (winner-take-all loss) and inference in trajectory forecasting models, leading to uninformative mode probabilities. It proposes post-hoc treatments using posterior-weighted merging and a one-step EM update to improve mode ranking and forecasting accuracy without retraining.
TEMPO introduces a test-time training framework that alternates policy refinement with critic recalibration to prevent diversity collapse and sustain performance gains in large reasoning models, boosting AIME 2024 scores for Qwen3-14B from 42.3% to 65.8%.