Study of output recycling

[hanaol]

Goal

Evaluate whether the model can refine its predictions through output recycling.

Setup

We test two scenarios:

• Single-step training + inference-only recycling
• Multi-step training (iterative feedback of model outputs)

Results (Single-step + Inference Recycling)

NMAE (~3K subset training) across structures over successive recycling iterations:

Task ID      Iter 1  Iter 2  Iter 3  Iter 4  Iter 5
mp-1877980   15.33   21.60   33.71   47.04   59.71
mp-1895405    8.45   15.59   24.90   33.84   42.68
mp-2706573    8.87   18.83   30.43   41.32   51.02
mp-1843254    8.52   14.55   23.16   31.87   40.26
mp-2503753    8.84   15.65   28.01   40.36   52.33
mp-1828114    0.80    6.87   13.39   19.74   25.96
mp-1920881    0.67   13.36   24.38   34.77   44.63
mp-1792559    1.09   11.28   20.85   30.08   38.55
mp-2457687    2.19   10.10   17.28   24.34   31.23