The Gradient Does Not See Rank: Rank-Indifference in Matrix-CODI on ProsQA

arXiv cs.LG Papers

Summary

The study finds that rank truncation in matrix-CODI models does not impact accuracy on reasoning tasks, indicating that rank may not correlate with parallel reasoning paths in latent representations.

arXiv:2609.03090v1 Announce Type: new Abstract: Continuous chain-of-thought models compress reasoning into latent tokens. Matrix-valued variants, which route each latent token through a d x d matrix bottleneck, introduce rank as a single-sample structural observable on the latent matrix Z. If matrix latents carry parallel reasoning paths via superposition, rank should track them, and truncating Z to low rank should hurt accuracy on tasks whose solutions plausibly require multiple components. Across four training regimes of a matrix-CODI model (three on ProsQA, one on GSM8K-Aug below the learning threshold), the rank-k projection ablation curve is flat to within 0.6 percentage points. A three-seed replication yields 81.0 +/- 2.0 percentage points accuracy while the final effective rank of Z spans {4, 12, 13}; the loss does not reward any particular rank. To test whether rank-blindness arises from the flatten-then-project readout alone, we trained four readouts: a bilinear reparametrization, a bilinear-plus-GELU readout nonlinear in Z, an SVD-augmented readout feeding singular values through an MLP, and a quadratic readout in Z Z^T. All four rank-k curves remain flat (Spearman p-values 0.63, 0.14, 0.82, 0.46). The flat curves persist for readouts nonlinear in Z. A linear probe on Z underperforms a raw pretrained hidden state at target prediction (AUC 0.673 vs. 0.846). A negative control on vanilla GPT-2 SFT (no matrix bottleneck, no Z, three seeds, n=500) reproduces a flat rank-k curve under the same intervention paradigm with pooled-mean range 0.20pp, and a random-h sensitivity floor lands at the same accuracy: the rank-k ablation alone conflates rank-blindness with position-irrelevance.
Original Article
View Cached Full Text

Cached at: 09/04/26, 06:20 AM

# The Gradient Does Not See Rank: Rank-Indifference in Matrix-CODI on ProsQA
Source: [https://arxiv.org/abs/2609.03090](https://arxiv.org/abs/2609.03090)
[View PDF](https://arxiv.org/pdf/2609.03090)

> Abstract:Continuous chain\-of\-thought models compress reasoning into latent tokens\. Matrix\-valued variants, which route each latent token through a d x d matrix bottleneck, introduce rank as a single\-sample structural observable on the latent matrix Z\. If matrix latents carry parallel reasoning paths via superposition, rank should track them, and truncating Z to low rank should hurt accuracy on tasks whose solutions plausibly require multiple components\. Across four training regimes of a matrix\-CODI model \(three on ProsQA, one on GSM8K\-Aug below the learning threshold\), the rank\-k projection ablation curve is flat to within 0\.6 percentage points\. A three\-seed replication yields 81\.0 \+/\- 2\.0 percentage points accuracy while the final effective rank of Z spans \{4, 12, 13\}; the loss does not reward any particular rank\. To test whether rank\-blindness arises from the flatten\-then\-project readout alone, we trained four readouts: a bilinear reparametrization, a bilinear\-plus\-GELU readout nonlinear in Z, an SVD\-augmented readout feeding singular values through an MLP, and a quadratic readout in Z Z^T\. All four rank\-k curves remain flat \(Spearman p\-values 0\.63, 0\.14, 0\.82, 0\.46\)\. The flat curves persist for readouts nonlinear in Z\. A linear probe on Z underperforms a raw pretrained hidden state at target prediction \(AUC 0\.673 vs\. 0\.846\)\. A negative control on vanilla GPT\-2 SFT \(no matrix bottleneck, no Z, three seeds, n=500\) reproduces a flat rank\-k curve under the same intervention paradigm with pooled\-mean range 0\.20pp, and a random\-h sensitivity floor lands at the same accuracy: the rank\-k ablation alone conflates rank\-blindness with position\-irrelevance\.

## Submission history

From: Samuel Larson \[[view email](https://arxiv.org/show-email/cae88526/2609.03090)\] **\[v1\]**Wed, 2 Sep 2026 19:03:24 UTC \(118 KB\)

Similar Articles

The Loss Does Not See the Basis, but Adam Does

Hugging Face Daily Papers

This paper investigates why Adam does not exhibit gradient descent's implicit low-rank bias in factored models, showing that coordinate-wise preconditioning breaks the relevant symmetry, while shared-scalar methods like Muon and Shampoo preserve it.

More Thinking, More Bias: Length-Driven Position Bias in Reasoning Models

arXiv cs.AI

This research paper investigates position bias in reasoning models, finding that bias scales with the length of the reasoning trajectory rather than being eliminated by 'more thinking.' The study provides causal evidence and a diagnostic toolkit for auditing this length-driven bias in multiple-choice QA evaluations.