Physics-IQ 官方协议 · 涨分最高的场景
官方头条 benchmark(base 31.75 → a14b_2x 37.04)。fmdual_buggy = 与 a14b_2x 同 buggy 数据、仅去掉 V-JEPA 对齐,用于隔离对齐贡献。
two-balls-pass
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
lit-candle
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
marble-run-x
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
roll-behind-box
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
ball-train
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
ball-rolls-on-glass
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
weight-on-ceramic
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
rolling-reflection
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
magnet-wrench
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
roll-front-box
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
napkin-soak
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
glass-stays-same
Ground truth(真实)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
PhyCo 域内 test set · continuation(合成物理,带真实 GT)
PhyCo held-out 测试集:给定首帧续帧 vs 真实后续片段。域内 ST-IoU 总体 base 0.10 → a14b_2x 0.31(3.2×)。按场景标注(合成数据不显示 caption)。
ball
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
ball
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
ball
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
ball
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
friction
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
friction
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
pool
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐
pool
Ground truth(真实续帧)
Base(未训练)
fmdual_buggy
a14b_2x ⭐