cs.LGNov 9, 2026

Spectral Rank and Step-Direction Dynamics in Language-Model Post-Training

Authors: Khurram Khalil, Khaza Anuarul Hoque

Organizations: Department of Electrical Engineering & Computer Science, University of Missouri-Columbia, MO, USA

Abstract

Language-model post-training modifies weights through cumulative updates and successive displacements, yet existing studies report seemingly conflicting spectral properties: nearly full-rank updates and low stable-rank changes. We show these observations are compatible and characterize update directions during training. For singular values following σi∝i−ασ_i \propto i^{-α} with 0<α≤1/20 < α\leq 1/2, we prove that stable rank becomes negligible relative to fixed-energy rank as dimension increases. Full singular value decompositions of selected OLMo-2-7B updates confirm this separation: three proximal policy optimization (PPO) attention updates have stable ranks of 9.6-18.3 but require 532-1,222 singular components to capture 80% of their energy. We next compute consecutive checkpoint-displacement cosines from full-model inner products. Across nine publicly released reinforcement learning from verifiable rewards (RLVR) stages, all 131 adjacent pairs have negative cosines on their native checkpoint grids, including PPO and Group Relative Policy Optimization (GRPO) trajectories. Supervised fine-tuning comparisons across four model families show configuration-dependent alignment, including negative late-training cosines for Mistral and FP32 Llama. Finally, a controlled OLMo-2-1B study with three paired seeds and a 2,048-token response cap compares group-relative advantages against a learned critic with generalized advantage estimation (GAE). At matched 60-rollout windows, late-training mean cosines are +0.259 and +0.012, respectively; the paired critic-minus-group difference is -0.247 (seed-level 95% interval [-0.358, -0.137]). These findings distinguish spectral concentration, energy coverage, and trajectory direction, and demonstrate that the advantage system influences directional alignment under matched training conditions.

Figures & tables

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 15, 2026

Rank Measures of Fine-Tuning Updates Need a Noise Baseline

Studies of what fine-tuning does to a language model often summarise the weight update ΔWΔW by its rank. We compare the rank measures used for this against the same measures on a random matrix of the same shape, and against two harder references: the pretrained weights, and matched fine-tunes without alignment content. Measures that count directions put our fine-tuning updates only a little below noise; stable rank, computed from the same singular values, puts them far below. Both are right: an update has a few strong directions on top of a long tail. The pattern holds for thirteen released pairs up to 8B parameters, and it is not a mark of alignment: with training held fixed, instruction data does not stand out from Wikipedia or code, DPO gives the same update whether its preference labels are real or half swapped at random (one model, modest preference strength), and scrambled text gives the most concentrated update of all. Applied to a recent claim that post-training is superficial because the effective rank of the weights barely changes, the baselines show that a random update of the same size leaves that rank almost unchanged while wrecking the model, and that an edit built on the effective rank of the update has no measurable effect in our tests. Rank measures of weight updates need a noise baseline, and claims about alignment need a control without alignment content.
Oct 7, 2026

When Rank Rises as LLMs Degrade

Post-training adapts language models in non-stationary environments. Practitioners monitor representation health with RankMe and related spectral statistics, often assuming that rank falls when representations degrade. We show that this assumption is unsafe for LLM post-training. In a controlled study of Qwen3-0.6B with four degradation modes and three seeds, data duplication worsens held-out loss by 75% relative to healthy while increasing both original and centred RankMe; the latter changes by 13.5 pooled standard deviations. Covariance effective rank rises to nearly twice its healthy value. This failure is spectral dispersion rather than collapse, so a one-sided monitor rates the worst checkpoint as the healthiest. By contrast, a learning-rate misconfiguration lowers centred RankMe and k95, while uncentred RankMe is inconsistent across seeds. Direction is therefore a property of the regime-statistic pair and cannot be fixed by recalibration alone. We also distinguish two often-conflated statistics: RankMe normalises singular values, whereas covariance effective rank normalises eigenvalues. On raw intermediate-layer states in the pretrained model, massive activations pin the latter near 1 out of dimension d while RankMe retains usable range. We then test a two-sided, multichannel sequential monitor with separate calibration and test data. In a pre-registered shared-prefix, leave-one-seed-out evaluation, it detects all three damage regimes in every fold 10 to 60 steps after the fork and separates dispersion from downward-rank damage by firing direction. However, it never precedes held-out probe loss, and calibration with two seeds produces false alarms on the held-out healthy seed. Spectral monitoring can diagnose failure regimes, but it does not warn earlier than held-out loss, and validity claims require held-out healthy data.
Oct 7, 2026

Spectrally Targeted Muon

The Muon optimizer orthogonalizes each update matrix, setting all of its singular values to one, and has proven highly effective for training large language models. It remains unclear, however, whether this success comes from amplifying small singular directions that gradient descent neglects or from suppressing large, degenerate directions that disrupt training. We introduce Spectrally Targeted Muon, which orthogonalizes only the singular values above or below a threshold ττ, so that varying ττ interpolates between normalized SGD and Muon. It isolates the relevant singular subspaces with projections computed by Newton-Schulz iteration on a shifted Gram matrix, so no SVD is needed. We evaluate these variants on the CIFAR-10 and NanoGPT speedruns, tracking the effective rank of gradient, update, and weight matrices and a new metric, the alignment of updates with the tangent space of the weight matrix's isospectral manifold. We find three things. First, the small singular values of the momentum are not noise. Orthogonalizing everything except the few largest singular values of each matrix nearly matches Muon while touching only a small fraction of the momentum, whereas orthogonalizing only the top falls well short even though it holds almost all of it. On language models every momentum singular value is far below one, so targeted orthogonalization can only amplify, and Muon wins by making directions that are too small to train on at their raw scale trainable. Second, shrinking the largest singular values is what keeps the parameter spectrum flat; this is cheap, and it is not what drives the loss. Third, AdamW differs from Muon mainly in how slowly it builds structure, which explains its slower start and why warmup helps AdamW but only hurts Muon.