← News

ResearchSources: arXiv · 19 Nov 2026

Study finds final network weights carry traces of how training unfolded

An arXiv preprint reports that networks trained on the same task end up with distinct, readable fingerprints of their initialization and task order, though not a full record.

In plain words

When many neural networks learn the same task, they end up with different internal settings even though all perform well. The researchers found those final settings still hint at where each network started and the order in which it learned tasks. However, the earliest settings themselves could not be reliably recovered from the end result.

What was announced

The authors trained 30 networks per dataset and asked whether the finished models could be matched back to their starting points. Despite training reshaping the group so that initial spacing between models no longer predicted final spacing, every trained network was correctly paired with its original starting weights from ten candidates. Networks that began from the same starting point also developed similar-looking internal activity patterns even under different batch schedules, although comparing that activity with the networks' initial activity did not identify them.

Training history was partly readable too. A simple linear probe identified which of two earlier tasks came first with perfect accuracy on 80 held-out networks, and some intermediate weight states could be predicted, though the original starting weights could not be reliably rebuilt. Separately, pairs of finished solutions were mostly linked by low-loss paths, and perturbed models regained performance without retracing their earlier steps. The authors conclude that a shared task steers networks toward related solutions while leaving their individual histories intact.

What it does not show

  • The evidence is an abstract only; full methods, datasets and caveats are not available for verification.
  • The matching and decoding results come from the authors' own experiments and have not been independently confirmed.
  • The abstract reports selective recovery: earlier intermediate states were only partly predictable and initial weights could not be reliably reconstructed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.