Fisher8: Stabilizing Neural Heteroscedastic Regression via Output-Layer Fisher Geometry

Local loss-landscape stabilization under sample growth is typically measured either pointwise or through isotropic averaging in the full parameter space. Despite practical value, both choices probe directions that contribute little to the dominant local deformation of strongly anisotropic neural landscapes. We recast stabilization as an observational problem and introduce a unified family of criteria parameterized by an aggregation order and a probing distribution; within this family we propose a curvature-aligned criterion $Δ_2^{(D)}$ that probes the loss increment field in the top-$D$ eigenspace of the empirical Hessian near a trained solution. Solely from a local quadratic model, we prove that $Δ_2^{(D)}$ preserves the $O(k^{-2})$ mean-squared rate of the full-space criterion while replacing ambient-dimension curvature dependence with dependence on the subspace dimension $D$; a corollary gives a closed-form spectral expression and a proposition identifies the top-$D$ eigenspace as extremal within the eigenspace-aligned family. We also derive scalable estimators based on Hessian-vector products, subspace Monte Carlo, and a closed-form Gaussian-moment proxy. On a decoder-only transformer, a curvature-aligned probe occupying a tiny fraction of parameter space already reproduces the full-space mean-squared signal to within numerical noise throughout the validated local regime, and the closed-form estimator is orders of magnitude faster than direct Monte Carlo after subspace construction.

Activation steering methods modify intermediate representations of language models to control output behavior, but universally assume the activation space is Euclidean. We show this assumption fails drastically: the local geometry induced by the model's own output behavior -- the Fisher information metric of the softmax layer, pulled back through the Jacobian of subsequent layers -- deviates from the Euclidean metric by over 97% in relative spectral norm on GPT-2, with an effective dimensionality of only 2--17% of the ambient space. From this pullback Fisher metric, we derive a closed-form steering equation that identifies the minimum-distortion direction for any target concept, yielding a closed-form optimal direction at each point that can be applied iteratively without manifold fitting or data-driven geometry estimation. We call the resulting framework FishBack. The metric admits a layer-wise recursive decomposition, which reveals that existing methods -- CAA, ActAdd, ITI, and others -- each implicitly adopt a particular approximate metric, and that their performance gaps are quantitatively predicted by a single spectral diagnostic: the ratio of their implicit metric's cost to the Fisher-optimal cost. On GPT-2, iterative pullback steering consistently outperforms all Euclidean baselines across three verb-morphology concepts and four layers, with off-target KL reductions of $1.3\times$--$2.5\times$ relative to Euclidean gradient ascent and $1.5\times$ relative to CAA at matched concept probability.

The training of neural networks often entails objective functions that are not globally $L$-smooth. For these functions, it is both theoretically and practically difficult to reply to the question: what is the largest possible step size that ensures the convergence of gradient descent (GD)? We address this longstanding open question in deep learning by providing a unifying definition of "large" step sizes that requires only local Lipschitz (or even Hölder) continuity of the gradient. We design first-order adaptive methods that provably yield large step sizes and show that they operate at the edge of stability (EoS) right from the start of the training. In particular, the loss decreases nonmonotonically and the product between the step size and sharpness, i.e., the largest eigenvalue of the Hessian, stays above the EoS threshold of 2 throughout training. Using our method, we are also able to minimize the sharpness all the way down to its global minimum. Contrary to expectation, we find that encountering globally-flat regions too early in the training may both slow down convergence and jeopardize the generalization ability of the network. Exploiting a self-stabilization argument, we allow GD to enter slightly sharper valleys and turn unsuccessful training runs into very successful ones.