🤖 machine learning

Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology

This paper analyzes the Jacobian eigendecomposition of large language models to reveal that training induces a monotonic spectral gradient from rotation-dominated early layers to symmetric late layers, creating a low-rank bottleneck that dynamically links perturbation propagation and compression to the network's functional topology.

Original authors: Jesseba Fernando, Grigori Guitchounts

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jesseba Fernando, Grigori Guitchounts

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a Large Language Model (LLM) not as a static brain, but as a factory assembly line where a piece of information (a sentence) travels from the beginning to the end. As it moves down the line, it passes through 30 to 40 different workstations (layers). At each station, the information gets tweaked, rotated, stretched, or compressed before being passed to the next one.

This paper, "Dynamics of the Transformer Residual Stream," acts like a high-speed camera and a spectral analyzer for this factory. Instead of just looking at the final product, the authors watched exactly how the information changes as it moves through the line. They used three different factory models (Llama, OLMo, and Gemma) to see if these changes are built into the factory's design or if the factory "learns" how to run itself during training.

Here are the three main discoveries, explained with everyday analogies:

1. The "Spin-to-Straight" Gradient

The Discovery: The authors found that the way information is transformed changes predictably as it moves from the start of the model to the end.

Early Layers (The Spinners): At the beginning of the line, the workstations act like mixers or spinners. They don't just stretch the information; they rotate it. In math terms, about 98% of the changes involve "complex rotations." Imagine taking a piece of clay and spinning it on a wheel while slightly squishing it.
Late Layers (The Straighteners): By the time the information reaches the end of the line, the workstations stop spinning and start acting like straighteners or mirrors. They align the information more directly, making the changes more predictable and symmetrical.
The Takeaway: The factory doesn't treat every station the same. It has a specific "flow": it starts by mixing things up and ends by straightening them out. This isn't just how the factory is built; the factory learns to do this during training.

2. The "Funnel" Effect

The Discovery: As the information travels through the entire factory, it gets squeezed into a tiny channel.

The Analogy: Imagine pouring a gallon of water (the full information) into a wide funnel at the top. As it passes through the layers, the funnel gets narrower and narrower. By the time the water reaches the bottom, it's not flowing through a wide pipe anymore; it's forced through a tiny straw.
The Math: The model starts with thousands of possible directions the information could go. By the end, almost all of those directions are crushed. Only about 7 to 40 directions (out of thousands) actually survive the trip to the end.
The "Why": The authors proved this isn't just because the factory is narrow. It's because the "spinners" (the non-normal parts of the math) learned to actively push information into this tiny straw. If you removed the "spin" and left only the "stretch," the funnel would disappear, and the water would spill everywhere. The factory learned to compress the data.

3. The "Bridge" Workers and the "Team" Structure

The Discovery: The authors looked at how the different parts of the information (units) group together into "teams" or communities. They found that the factory treats the "team leaders" differently than the "team members."

The Analogy: Imagine the information units are workers in a factory. Some workers only talk to people in their own department (a "community"). Others are bridge workers who talk to people in multiple departments.
The Finding:
- In the early/middle parts of the factory (where the "spinners" are), the factory actually suppresses these bridge workers. It tells them to quiet down.
- In the late parts of the factory (where the "straighteners" are), the factory amplifies these bridge workers. It boosts their signal.
The Takeaway: The factory has learned a specific rule: "Don't let the cross-department messengers shout in the mixing zone, but let them lead the way in the final assembly zone." This rule didn't exist when the factory was first built; it was learned during training.

Summary: What is New Here?

Before this paper, scientists mostly looked at the factory when it was empty (randomly initialized) or used blurry, approximate tools to watch it run.

Architecture vs. Learning: By comparing a brand-new, untrained factory to a trained one, they showed that the "funnel" and the "spin-to-straight" gradient are learned behaviors, not just the result of the factory's blueprints.
The Full Picture: They didn't just look at how much the information grew or shrank; they looked at the rotation and the direction. They found that the factory is a dynamic system that actively funnels information and organizes its internal teams in a very specific, learned way.

In short, the paper reveals that these AI models aren't just static calculators; they are dynamic systems that have learned to spin, squeeze, and organize their internal thoughts in a highly structured, non-random way to get the job done.

Technical Summary: Dynamics of the Transformer Residual Stream

Problem Statement
While Large Language Models (LLMs) demonstrate remarkable capabilities, the mechanisms by which computation propagates through their layers remain poorly understood. Existing interpretability tools, such as circuit analysis and sparse autoencoders, identify what components represent but fail to explain how perturbations are amplified, compressed, or redirected across layers. Dynamical systems theory offers a formalism for this, treating depth as discrete time and the residual stream as a dynamical system. However, previous analyses have relied on scalar summaries, approximate linearizations, or studies conducted only at random initialization. Consequently, the full spectral geometry of trained, production-scale Jacobians—specifically their eigenvalue distributions, non-normality structures, and rotational dynamics—remains uncharacterized. Furthermore, the relationship between a network's computational topology (community structure) and its dynamical behavior is untested at scale.

Methodology
The authors perform a full spectral characterization of three production-scale decoder-only transformers: Llama 3.1 8B, OLMo 3 7B, and Gemma 4 E4B.

Jacobian Computation: Exact per-sample Jacobians ( $J_\ell = \partial h_{\ell+1} / \partial h_\ell$ ) are computed at every sub-layer boundary (pre-attention and pre-MLP) across 1,000 WikiText-2 samples.
Spectral Analysis: Full eigendecomposition and singular value decomposition (SVD) are performed on the mean Jacobians. Key metrics include condition numbers, self-alignment (overlap between input/output singular subspaces), Henrici departure (measure of non-normality), and the fraction of expanding modes.
Training vs. Architecture Decomposition: Using OLMo 3 7B, the study compares three checkpoints: Step 0 (random initialization), Step 471k, and Step 1.41M (final pretraining). Since Step 0 shares the architecture and depth but lacks learned weights, it serves as a null model to distinguish architectural priors from training-induced properties.
Topological Analysis: Activation-correlation graphs are constructed for each layer, partitioned into communities using signed Leiden CPM. The study analyzes the coupling between community topology (specifically "boundary nodes" bridging multiple communities) and Jacobian dynamics.
Intervention: Schur decomposition is used to isolate the non-normal component ( $N_\ell$ ) of the Jacobian. By scaling $N_\ell$ while holding the spectrum ( $\Lambda_\ell$ ) and basis ( $Q_\ell$ ) fixed, the authors test whether the observed dimensional bottlenecks arise from the eigenvalue spectrum or the non-normal feedforward structure.

Key Results

Monotonic Non-Normality Gradient:
Trained models exhibit a distinct depth-dependent transition in operator type. Early layers are dominated by non-normal, rotation-heavy dynamics (approx. 98% of eigenvalues appear as complex conjugate pairs), while late layers transition toward near-symmetric operators. This gradient is largely learned; at initialization (OLMo Step 0), layers are already near-symmetric, and training pulls early and mid-layers toward rotation.
Cumulative Dimensional Bottleneck:
Composing Jacobians end-to-end reveals a severe funneling of perturbations. In Llama 3.1 8B, the effective rank of the cumulative Jacobian drops from ~436 (single layer) to ~7 (full depth), meaning only a handful of effective channels survive from embedding to logit. This collapse is not architectural (effective rank remains high at Step 0) but is installed by training. Crucially, Schur surgery experiments show that removing the non-normal component ( $N_\ell$ ) while preserving the spectrum largely dissolves this bottleneck, indicating the funnel is a property of the trained non-normal feedforward structure, not the eigenvalues.
Topology-Dynamics Coupling:
The study identifies a learned coupling between the network's community topology and its dynamics. Units that bridge multiple activation-graph communities (boundary nodes) are preferentially amplified or suppressed by the Jacobian.
- Sign Dependence: The sign of this coupling is determined by the local operator type. In near-symmetric layers, boundary units are amplified. In strongly non-normal layers, they are de-amplified.
- Emergence: This coupling is absent at initialization and emerges monotonically over training. In OLMo, the number of FDR-significant layers increases from 0 (Step 0) to 14 (Step 471k) to 24 (Step 1.41M).

Significance and Claims
The paper claims to provide the first full eigenvalue-level description of production-scale transformer Jacobians after training. Its primary contributions are:

Filling the Gap: It bridges the gap between initialization-time theory (which predicts none of this structure) and approximate linearizations (which cannot resolve the rotational dynamics dominating the spectrum).
Mechanistic Insight: It demonstrates that the "dimensional collapse" observed in LLMs is a learned property driven by non-normality, not an architectural inevitability or a spectral artifact.
Topology-Dynamics Link: It establishes that the network's functional topology (community structure) is not just a static feature but dynamically interacts with signal propagation, with the nature of this interaction (amplification vs. suppression) being a learned function of the layer's operator type.

The authors note that while the mesoscale community structure itself is partly architectural (present at initialization), the specific coupling between boundary nodes and Jacobian amplification is a learned phenomenon. The work suggests that future interpretability efforts must account for the full spectral geometry, particularly non-normality and rotational dynamics, rather than relying solely on singular value decompositions.

1. The "Spin-to-Straight" Gradient

2. The "Funnel" Effect

3. The "Bridge" Workers and the "Team" Structure

Summary: What is New Here?

Technical Summary: Dynamics of the Transformer Residual Stream

More like this