一位独立研究者针对Yann LeCun团队发布的基于联合嵌入预测架构(JEPA)的世界模型LeWM进行了600次系统性实验 1。研究发现,在训练数据覆盖的路径上,LeWM的预测能力已接近完美的物理模拟器,其性能瓶颈并非在于预测精度,而在于长时程任务的规划方法 1。数据显示,原始的Vanilla LeWM模型在处理75步的长时程任务时,成功率仅为22% 1。为了验证这一结论,研究者尝试使用完美物理替换原有的预测器,结果发现这仅使成功率提升了约20个百分点(达到77-78%左右),进一步证实了预测误差并非主要限制因素 1。
通过引入无需重新训练的优化规划策略,该研究显著提升了模型在实际任务中的表现。第一种方法是实施分层规划,将长任务拆分为25步的子目标,因为研究发现LeWM的有效规划视界约为25步 1。结合其他优化措施后,这种方法成功将任务完成率从22%提升至85% 1。第二种方法则是构建一个包含46万个嵌入状态的潜在记忆图,并为跨演示跳转添加成本约束,实验显示嵌入距离排名与物理可达性一致的概率高达87% 1。此外,一项关于目标评分的单次更改单独贡献了60个百分点的成功率提升 1。综合这些改进后的规划方法,在不改变模型本身的情况下,将成功率从基准的62%进一步提升至85%,达到了专家级子目标的表现水平 1。
An independent researcher conducted 600 experiments on LeWM, a world model based on the Joint Embedding Predictive Architecture (JEPA) developed by Yann LeCun’s team, to evaluate its capabilities in long-horizon tasks 1. The study found that vanilla LeWM achieves only a 22% success rate on 75-step tasks, but replacing its predictor with a perfect physics simulator boosts performance by approximately 20 percentage points, reaching roughly 77–78% 1. This marginal improvement indicates that prediction accuracy is not the primary bottleneck; rather, the limiting factor lies in how the model plans over extended sequences 1.
To address this planning limitation without retraining the model, the researcher introduced two distinct strategies: hierarchical subgoal segmentation and a latent memory graph 1. By breaking long tasks into smaller segments of about 25 steps—which aligns with LeWM’s effective planning horizon—and combining this with other optimizations, the success rate increased to 85% 1. Additionally, constructing a latent memory graph containing 460,000 embedded states, where embedding distance rankings corresponded to physical reachability with 87% consistency, further refined navigation across demonstrations 1. One specific adjustment to the goal scoring mechanism alone contributed to a 60-percentage-point jump in success rates 1. Ultimately, these improvements raised the success rate from an intermediate baseline of 62% to 85%, matching the performance of expert-defined subgoals and demonstrating that optimizing usage methods yields greater gains than enhancing the model architecture itself 1.
评论
还没有评论,欢迎留下第一条。