Yifan Song

Yifan Song, Keyang Yu, Seth Young

This work leverages the U.S. Federal Aviation Administration's Traffic Flow Management System dataset and DV8, a recently developed tool for highly interactive visualization of air traffic data, to develop clustering algorithms for categorizing air traffic by their varying flight paths. Two clustering methodologies, a spatial-based geographic distance model, and a vector-based cosine similarity model, are demonstrated and compared for their clustering effectiveness. Examples of their applications reveal successful, realistic clustering based on automated clustering result determination and human-in-the-loop processes, with geographic distance algorithms performing better for enroute portions of flight paths and cosine similarity algorithms performing better for near-terminal operations, such as arrival paths. A point extraction technique is applied to improve computation efficiency.

Wenhan Ma, Hailin Zhang, Liang Zhao et al. · pku

Reinforcement learning (RL) has emerged as a crucial approach for enhancing the capabilities of large language models. However, in Mixture-of-Experts (MoE) models, the routing mechanism often introduces instability, even leading to catastrophic RL training collapse. We analyze the training-inference consistency of MoE models and identify a notable discrepancy in routing behaviors between the two phases. Moreover, even under identical conditions, the routing framework can yield divergent expert selections across repeated forward passes. To address this foundational inconsistency, we propose Rollout Routing Replay (R3), a method that records routing distributions from the inference engine and replays them during training. R3 significantly reduces training-inference policy KL divergence and mitigates extreme discrepancies without compromising training speed. Extensive experiments on various settings confirm that R3 succeeds in stabilizing RL training, preventing collapse and outperforming methods such as GSPO and TIS. We believe this work can offer a new solution for stabilizing RL in MoE models.