arXiv:2611.23245v1 Announce Type: new Abstract: Response-relevant information is not sufficient evidence that two neural networks are in the same training state. We define training-state equivalence operationally: two checkpoints are equivalent, relative to an intervention family and horizon, if they respond alike when training is perturbed and continues. On fresh runs of a one-layer transformer learning modular addition, response to a weight-decay pulse is highly reproducible relative to between-checkpoint differences (rho_state = 0.934 at 500 steps), establishing the target as measurable. Fifteen descriptors of weight spectra, representation geometry and optimiser geometry carry within-run response-relevant information beyond weight norm and loss on a development dataset, predicting held-out response deviation across seed blocks. But their apparent 16.1% advantage at matching checkpoints by future response does not separate from a conditional placebo preserving each descriptor's fitted dependence on the conventional observables, and a calibrated positive control makes simple insensitivity an unlikely explanation. Information a decoder can extract need not be organised in the pairwise geometry that similarity judgments require.
Response-Relevant Information Is Not Sufficient for Neural Training-State Equivalence
A study posted to arXiv (2611.23245v1) finds that response-relevant information alone is not sufficient evidence that two neural networks occupy the same training state. On fresh runs of a one-layer transformer learning modular addition, the authors measured highly reproducible responses to a weight-decay pulse relative to between-checkpoint differences (rho_state = 0.934 at 500 steps), but the apparent 16.1% advantage of fifteen weight-spectrum, representation-geometry and optimiser-geometry descriptors at matching checkpoints by future response did not separate from a conditional placebo preserving each descriptor's fitted dependence on conventional observables. The authors conclude that information a decoder can extract need not be organised in the pairwise geometry that similarity judgments require.
When using representational or weight-spectrum descriptors to match checkpoints, test them against a conditional placebo that preserves each descriptor's dependence on loss and weight norm before relying on the apparent advantage.
more in playbook: Evals and testing for LLM appsRun your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.