REVIEW 4 major objections 6 minor 20 references
This paper claims that FFN layers are directional steering fields: only the tangential component of their output moves residual directions, and this control is what preserves model quality and diversity.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
2026-08-04 15:46 UTC pith:XGA2JY75
load-bearing objection Solid theory, honest limitations, but the headline empirical claim is weaker than the abstract suggests—the main diagnostic is explicitly tautological in the body, and the ablation is confounded by relative norms. the 4 major comments →
Feed-Forward Steering in Transformer Residual Dynamics
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central discovery is that the residual stream splits into a direction u and a norm r, and under the projected dynamics only the tangential field PuF changes u. The FFN contributes a tangential component that is needed to account for the realized angular update; attention alone leaves an angular deficit of 0.41 for GPT-2, rising to 0.63 for Llama-3-8B. Ablations confirm: keeping only the tangential FFN output preserves most language-modeling quality, while keeping only the radial output collapses it. The sequential-to-parallel defect DPhiA controls when a block can be parallelized; low-defect layers survive replacement with roughly +0.02 loss in GPT-2-large, while high-defect layers degra
What carries the argument
The central object is the sphere-valued residual-direction ODE dot-u_i = (1/s_i) P_{u_i}(A_i(U) + Phi_i(u_i)), where P_u = I - uu^T projects onto the tangent space. Only the tangential projection moves u; radial motion changes norm and hence angular speed. Critical residual directions are zeros of the projected field, with stability read from the tangent Jacobian. The practical selector is the FFN input-sensitivity defect C_l = ||Phi(x+A(x)) - Phi(x)|| / (||Phi(x)|| + eps), which controls the error of the parallel surrogate x+A(x)+Phi(x).
Load-bearing premise
The load-bearing assumption is that the continuous-depth, direction-only ODE (1) faithfully represents how discrete transformer blocks actually move residual directions; if the discrete-to-continuous idealization or the direction-only reduction is inaccurate, the steering predictions about real models may not hold.
What would settle it
A single counterexample where removing the tangential FFN component leaves loss essentially unchanged, or where attention-only angular alignment equals the full block alignment, would refute the central claim. Concretely, measure the angular deficit on a checkpoint whose residual updates are dominated by norm changes: the theory predicts direction barely moves, while a large directional update would reveal radial-to-tangential coupling the model omits.
If this is right
- If the FFN is primarily a steering field, then FFN ablations that keep only the tangential component should retain most model quality; the paper observes this across GPT-2, Pythia, Mistral, and Llama models.
- Attention-only angular deficit increases from 0.41 to 0.63 across the tested models, implying that attention-only theories miss a growing part of the realized directional motion as models scale.
- FFNs counteract attention-driven contraction: attention usually decreases spherical diversity while the FFN often increases it, supporting an anti-collapse role for the tangential component.
- Layers with small C_l can be replaced by the same-input parallel surrogate with modest loss, yielding an estimated 1.24x speedup on 18 GPT-2-large layers at +0.02 loss, while high-C_l layers degrade quickly.
Where Pith is reading between the lines
- If the steering picture holds, interpretability interventions such as model editing could be refocused on the tangential FFN projection rather than raw value outputs, since only that component changes direction.
- The C_l score could be lifted directly into inference schedulers: layers with low defect can run attention and FFN concurrently, and the score provides a principled selection rule beyond the diagnostic use reported.
- The growing attention-only angular deficit with model scale suggests that larger models may rely more on FFN steering, implying attention-only continuous-depth theories will become less accurate predictors of residual geometry at scale.
- The saddle-dominated local spectra suggest residual flows are transient and transport-like rather than convergent; a testable extension is whether training with explicit diversity regularization pushes the FFN toward stronger tangential steering.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper extends continuous-depth dynamical models of Transformer residual streams by adding the FFN as a local steering field. The state is a token direction u on the unit sphere, and the FFN field Φ enters through its tangential projection P_uΦ, complementing the attention field A. The theoretical part gives well-posedness for the particle ODE, equilibrium and stability criteria for the projected field, a mean-field limit for full attention, a first-order expansion of critical directions in the FFN gain, and commutator/order-defect expansions for sequential versus parallel block maps. The empirical part, on GPT-2, GPT-2-large, Pythia-410M/1.4B, Mistral-7B, and Llama-3-8B, reports an attention-only angular-alignment deficit, tangential/radial FFN ablations, a diversity preservation experiment, and a defect-guided sequential-to-parallel replacement experiment. The paper concludes that FFN layers act as directional steering fields rather than as mere norm corrections.
Significance. If the central claim is correct, the paper would connect two previously separate literatures: continuous attention dynamics and FFN interpretability. It would also give a principled layer-selection score for block-level parallelization. The mathematical core is clean and appears correct: the proofs of well-posedness, the stability spectral criterion, the mean-field contraction, and the Taylor-order defect formulas are complete and self-contained. The paper is also unusually transparent about the status of its own diagnostics, stating explicitly in several places that a given quantity is an attribution diagnostic rather than an independent validation. These strengths deserve credit. The weakness is that the main empirical bridge from the ODE to real transformers is either algebraic or confounded. The reported ``attention-only angular deficit'' is an identity-level rearrangement, the tangential/radial ablation is not norm-controlled, and the parallelization score is essentially equal to the replacement error it is used to predict. Hence the empirical support for the stronger interpretive claim is currently thinner than the abstract suggests.
major comments (4)
- [Abstract and §4 (Eqs. 15–18)] The abstract claims that the extended theory ``improves one-step angular prediction relative to an attention-only baseline,'' but no such independent prediction experiment is reported. Equation (17) shows that the realized tangent displacement is Δu_i = (1/||x^{l+1}||) P_u(A_i + Φ_i), so the full-field tangent direction is exactly collinear with Δu by construction. Consequently cos(P_u(A+Φ), Δu)=1 identically and the attention-only deficit D_A in Eq. (18) measures only the relative tangential mass of A versus Φ. The main text correctly labels D_A as an attribution diagnostic, but the abstract and conclusion use it to support the steering interpretation. Either remove the ``improves one-step angular prediction'' claim or provide a true predictive test of Eq. (1), e.g., integrated over layers or with a finite-step predicted direction compared against the realized next direction.
- [§4, Tangential effect of FFN (Eq. 19)] The radial/tangential ablation is confounded by the relative norms of the two components. Since Φ = ⟨u,Φ⟩u + P_uΦ, a random high-dimensional FFN output will have ||P_uΦ|| approximately sqrt(d−1) times |⟨u,Φ⟩|, so tangential-only is close to the full FFN and radial-only is close to zero-FFN purely by norm, with no need for any directional-steering mechanism. The paper does not report the distributions of ||P_uΦ||/||Φ|| and |⟨u,Φ⟩|/||Φ||, nor a norm-matched control. The claim that the tangential component is ``the useful part'' therefore conflates directionality with vector magnitude. Please report these norms and include a norm-matched ablation (e.g., scale the radial-only update to the same total added norm as the tangential-only update, or compare against a random-tangent control).
- [§4, Sequential-to-parallel defect (Eq. 23)] The score C_ℓ = ||Φ(x+A(x))−Φ(x)||/(||Φ(x)||+ε) is, by Proposition 2, exactly the first-order (indeed one-step) difference between the sequential and parallel residual maps. Selecting low-C_ℓ layers and observing that they are safe to parallelize is therefore a direct measurement of the intervention error, not an independent prediction from the theory. This does not invalidate the practical engineering result, but it overstates the theoretical validation. To support the claim that the commutator or derived defect has predictive power, C_ℓ should be evaluated against an independent criterion, such as loss degradation on held-out layers after fitting the score on other layers, or compared with an unrelated sensitivity proxy.
- [§2/§4, Eq. (1)] The central explanatory model, the continuous-depth direction-only ODE (1), is never directly tested. The experiments decompose one-step updates, run ablations, and measure diversity, but none integrates Eq. (1) or its normalized-time variant over layers and compares the resulting trajectory with actual residual directions. Since the theoretical claims about equilibria, stability, and steering concern the flow of Eq. (1), the gap between the discrete transformer and the ODE idealization is load-bearing. A direct test would be to use the realized A_i^ℓ and Φ_i^ℓ as fields, integrate the ODE over a small number of layers, and report predicted versus actual angular displacements; alternatively, compare the Euler step of Eq. (1) with the realized Δu after accounting for the layer step size. Without this, the paper's explanatory conclusions remain conditional on an unvalidated idealization.
minor comments (6)
- [Abstract] The abstract in the header and the abstract in the full text are inconsistent: one says the extended theory ``improves one-step angular prediction,'' while the other reports only an ``attention-only angular-alignment deficit.'' Please align the two versions and ensure the abstract matches the actual experiments.
- [§B, Evaluation metrics] The metric called ``one-step faithfulness'' is defined verbally as cos(A+Φ,actual)−cos(A,actual). Since the first term is identically 1 by Eq. (17), this metric reduces to D_A. This should be stated explicitly to avoid the impression that a separate prediction experiment was run.
- [§C.4] Section C.4 reports ``91/111 candidates are saddle-like'' for the full field, whereas Table 2 reports 116/140 for the full field. The counts may come from different candidate sets or thresholds, but the discrepancy should be reconciled or explained.
- [§4, Fig. 3 and Table 1] The text reports numeric loss ranges for the ablation (0.07–1.72 tangential-only, 7.19–12.96 radial-only), but Figure 3 appears to show only qualitative curves and Table 1 reports gate diagnostics. Please add a table with the exact per-model loss increases for the five forward-pass variants.
- [§B, Statistical significance] The paper states that formal significance tests are not applied. Given that the cross-model claims are central, at least a paired nonparametric test across the six models or a per-model confidence interval would strengthen the robustness claims.
- [Eq. (4) and surrounding text] The FFN proxy notation in Eq. (4) is clearly disclaimed as non-literal for gated architectures, which is good. However, the main text occasionally refers to ``key--value memories'' without the caveat; please add a one-sentence reminder where the experimental models are introduced.
Circularity Check
The headline one-step angular prediction is an algebraic identity (Eq. 17), so the reported attention-only deficit cannot validate the steering ODE; the remaining experiments are empirical and not circular.
specific steps
-
self definitional
[Abstract; Section 4 'Attention-only angular deficit', Eq. (17); Appendix B 'One-step faithfulness']
"Because P_{u^ℓ_i} x^ℓ_i = 0, the realized full branch sum satisfies Δu^ℓ_i = 1/∥x^{ℓ+1}_i∥ P_{u^ℓ_i}(A^ℓ_i + Φ^ℓ_i). (17) Thus the full-update tangent direction is algebraically collinear with the observed displacement and is not an independent prediction."
The 'full-field' one-step prediction is the observed displacement by construction. Δu_i is defined as the projection of the actual residual update, and Eq. (17) shows it equals the normalized projection of A+Φ. Hence cos(P_u(A+Φ), Δu) = 1 identically. The abstract's claim that 'the extended theory improves one-step angular prediction relative to an attention-only baseline' therefore reduces to the identity in Eq. (17). The Appendix B metric 'one-step faithfulness', cos(A+Φ,actual) − cos(A,actual), is just 1 − cos(P_u A, Δu) = D_A, since the first cosine is identically 1. This measures the tangential mass of Φ, but does not test the ODE (1), its equilibria, or stability. The paper itself labels the result an 'attribution diagnostic', but the abstract's predictive claim is still circular.
full rationale
The paper's central mathematical framework (Eqs. (1)-(14), Theorems 1-3, Propositions 1-2) is developed from external prior work (Geshkovski, Karagodin, etc.) and is not itself circular. The ablation, diversity, and parallelization experiments are genuine empirical tests: their outcomes could have been different, and they do not reduce to definitions. However, the headline empirical claim in the abstract—'improves one-step angular prediction relative to an attention-only baseline'—is supported only by the attention-only angular deficit D_A. Eq. (17) makes the full-field tangent direction identically collinear with the realized displacement, so the 'full-field prediction' is the observation itself. The paper acknowledges this in the main text ('not an independent prediction'), but the abstract still presents it as predictive validation. This is a partial circularity: one of the four headline results reduces by construction. The other three experiments (tangential/radial ablation, diversity under aggregation pressure, defect-guided parallelization) provide independent, non-circular evidence for the steering interpretation, so the score is not higher.
Axiom & Free-Parameter Ledger
axioms (5)
- domain assumption Transformer residual updates can be modeled as a continuous-depth ODE ẋ_i = F_i.
- domain assumption Residual direction u = x/||x|| is the natural state, with norm r controlling angular speed through 1/r.
- domain assumption The FFN field Φ is Lipschitz and the single-particle field g(u) = Pu(Vu + αΦ(u)) approximates dynamics near a coherent cluster.
- domain assumption The mean-field limit applies to full attention; causal decoder-only models follow a prefix-indexed Volterra flow.
- standard math C2 smoothness of fields for stability and Taylor expansions in Proposition 2 and Theorem 3.
Cite this review
Pith. "Pith review of Feed-Forward Steering in Transformer Residual Dynamics." pith.science (2026). https://pith.science/paper/XGA2JY75
@misc{pith2026260802071,
author = {Pith},
title = {Pith review of: Feed-Forward Steering in Transformer Residual Dynamics},
year = {2026},
howpublished = {\url{https://pith.science/paper/XGA2JY75}},
note = {Machine review of arXiv:2608.02071}
}
read the original abstract
Attention-only dynamical theories model Transformer residual directions as particles aggregating on a sphere. We extend this framework by incorporating the feed-forward network (FFN) term as a local steering field acting on each token state. The resulting theory predicts that the tangential component of the FFN field is necessary for motion in residual-direction space, that critical residual directions correspond to nonlinear projective equilibria, and that a commutator defect determines when a finite attention--FFN block can be accurately approximated by a parallel, additive flow. Across GPT-2, Pythia, Mistral, and Llama models, the extended theory improves one-step angular prediction relative to an attention-only baseline, with the contribution of the FFN increasing from GPT-2 to Llama-3-8B. Intervention experiments show that retaining only the tangential FFN component preserves most model quality, whereas retaining only the radial component causes performance to collapse. The tangential component also preserves output diversity under aggregation pressure. As a practical application, layers with small commutator defects can be approximately parallelized with only a modest increase in loss, whereas layers with large defects degrade rapidly. These findings support the interpretation of FFN layers as directional steering fields that shape Transformer residual geometry and govern the feasibility of block-level interventions.
Figures
Reference graph
Works this paper leans on
-
[1]
arXiv preprint arXiv:2312.10794 , year=
A Mathematical Perspective on Transformers , author=. arXiv preprint arXiv:2312.10794 , year=
-
[2]
Advances in Neural Information Processing Systems , year=
The Emergence of Clusters in Self-Attention Dynamics , author=. Advances in Neural Information Processing Systems , year=
-
[3]
Advances in Neural Information Processing Systems , year=
Clustering in Causal Attention Masking , author=. Advances in Neural Information Processing Systems , year=
-
[4]
arXiv preprint arXiv:2510.22026 , year=
Normalization in Attention Dynamics , author=. arXiv preprint arXiv:2510.22026 , year=
-
[5]
Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=
Transformer Feed-Forward Layers Are Key-Value Memories , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=
2021
-
[6]
Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=
Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=
2022
-
[7]
2021 , howpublished=
A Mathematical Framework for Transformer Circuits , author=. 2021 , howpublished=
2021
-
[8]
International Conference on Machine Learning , pages=
Attention is Not All You Need: Pure Attention Loses Rank Doubly Exponentially with Depth , author=. International Conference on Machine Learning , pages=
-
[9]
Advances in Neural Information Processing Systems , year=
Locating and Editing Factual Associations in GPT , author=. Advances in Neural Information Processing Systems , year=
-
[10]
International Conference on Learning Representations , year=
Mass-Editing Memory in a Transformer , author=. International Conference on Learning Representations , year=
-
[11]
International Conference on Machine Learning , pages=
On Layer Normalization in the Transformer Architecture , author=. International Conference on Machine Learning , pages=
-
[12]
Advances in Neural Information Processing Systems , year=
Root Mean Square Layer Normalization , author=. Advances in Neural Information Processing Systems , year=
-
[13]
Geometric Numerical Integration: Structure-Preserving Algorithms for Ordinary Differential Equations , author=
-
[14]
Differential Equations with Discontinuous Righthand Sides , author=
-
[15]
2019 , howpublished=
Language Models are Unsupervised Multitask Learners , author=. 2019 , howpublished=
2019
-
[16]
Proceedings of the 40th International Conference on Machine Learning , year=
Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling , author=. Proceedings of the 40th International Conference on Machine Learning , year=
-
[17]
arXiv preprint arXiv:2310.06825 , year=
Mistral 7B , author=. arXiv preprint arXiv:2310.06825 , year=
-
[18]
arXiv preprint arXiv:2407.21783 , year=
The Llama 3 Herd of Models , author=. arXiv preprint arXiv:2407.21783 , year=
-
[19]
2019 , howpublished=
OpenWebText Corpus , author=. 2019 , howpublished=
2019
-
[20]
arXiv preprint arXiv:2002.05202 , year=
GLU Variants Improve Transformer , author=. arXiv preprint arXiv:2002.05202 , year=
Pith/arXiv arXiv 2002
This paper was first reviewed by deepseek-v4-flash on August 4, 2026.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.