acceptodds
Under review as a conference paper at ICLR 2027

The Geometry of Weight-Tying Sensitivity in Transformers

13 views

Abstract

Transpose tying changes both a learned operator and the residual trajectory on which later computation acts. We develop a finite geometric account of these two effects in pre-normalized Transformers. Positive channel scaling yields a unique balanced SwiGLU representation and a weighted-adjoint tying operation that is invariant under function-preserving channel rescaling. Exact gated derivatives identify a rank-two contribution to normalization-induced rotation, while radial–directional coordinates distinguish residual magnitude from the angular change produced by an update. For a fixed pre-RMSNorm suffix at input scale , we derive explicit coefficients for its logit and output-KL response, together with finite-radius bounds and direction-resolved endpoint certificates. Public-model interventions show large propagation differences after local edit-magnitude matching. A complete two-model, six-text strength panel tests the balanced edit itself: its coordinate consistency coexists with substantial functional changes, including a low-amplification regime. A separate two-model reference-precision experiment computes the suffix coefficients without fitting output responses: at , the six cut-level forward-KL ratios to prediction range from to . The theory links a well-defined editing operation to residual geometry and quantitatively testable suffix effects.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Transformers, model editing, parameter sharing, representation geometry, mechanistic analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UnwgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the geometric sensitivity of pretrained Transformers to post-hoc weight-tying interventions, focusing on the interaction between feed-forward weight structure, residual-state geometry, and normalization.

The authors develop three interconnected theoretical results.

First, they characterize the positive channel-scaling symmetry of bias-free SwiGLU feed-forward networks and construct a uniquely balanced representation within each positive-scaling orbit. This construction leads to a weighted-adjoint transpose-tying intervention that is invariant under function-preserving channel rescaling.

Second, they analyze the local Jacobian structure of gated feed-forward networks composed with RMSNorm. They identify a rank-at-most-two contribution to the antisymmetric Jacobian component induced by normalization and introduce a radial-directional decomposition of residual updates.

Third, they derive an asymptotic expansion for the influence of a fixed pre-normalized Transformer suffix when its input is scaled by a factor (c). The analysis predicts an (O(c^{-1})) difference in output logits and an (O(c^{-2})) difference in output KL divergence, with explicit leading coefficients and complementary finite-radius bounds.

The empirical investigation uses two pretrained decoder-only language models, Qwen3-0.6B and SmolLM2-360M. The authors compare balanced transpose-tying interventions against random perturbations matched by weight-change magnitude or local branch-output perturbation. They additionally perform residual-state interventions and evaluate the asymptotic suffix predictions using a separate high-precision computation.

The experiments demonstrate that structured weight-tying interventions can cause substantially greater functional disruption than locally magnitude-matched random edits, and that the theoretical suffix coefficients accurately approximate measured responses in sufficiently large input-scaling regimes.

Overall, the paper aims to establish a geometric framework connecting coordinate-consistent weight tying, residual-state evolution, and downstream functional sensitivity.

Critical strengths
  1. A mathematically well-defined treatment of weight-tying interventions that accounts for parameterization symmetries.

A significant strength is the recognition that naive transpose tying is not invariant under function-preserving reparameterizations of SwiGLU3 hidden channels.

Proposition 1 provides a concrete way to address this issue by constructing a balanced representative within the positive diagonal scaling orbit and defining the corresponding weighted-adjoint tying operator.

This construction makes the intervention independent of arbitrary channel-scale allocations, an important consideration when comparing structural modifications across functionally equivalent parameterizations.

Unlike heuristic weight replacement or global magnitude calibration, the proposed operation has a clearly specified mathematical meaning.

The empirical coordinate-consistency checks in Section 7 provide additional support for the implementation of this construction.

This contribution is consequential because a well-defined intervention is a necessary prerequisite for interpreting the functional effects of structural weight modifications.

  1. Explicit finite-network analysis with independently evaluated predictions.

The fixed-suffix analysis provides a relatively precise mathematical characterization of how residual magnitude affects downstream computation in pre-normalized Transformers.

Theorem 1 derives explicit leading coefficients for logit, negative log-likelihood, and KL-divergence responses, rather than merely asserting qualitative attenuation or fitting a decay exponent to observed outputs.

Importantly, the coefficients are calculated from frozen branch evaluations without fitting the measured suffix responses.

The high-precision experiments in Section 8 and Appendix F show convergence toward these independently computed coefficients. At (c=32768), the reported forward-KL coefficient ratios are close to one across the six evaluated model/cut configurations.

The authors also retain unfavorable finite-scale results instead of selectively reporting only the asymptotic regime.

The distinction between asymptotic statements, finite-radius bounds, and exact endpoint certificates is carefully articulated.

These results provide a technically meaningful characterization of fixed Transformer computation and represent the strongest theoretical aspect of the submission.

Critical weaknesses
  1. The theoretical components are not sufficiently connected to establish a quantitative explanation of weight-tying sensitivity.

My main concern is that the paper develops several mathematically interesting results without convincingly establishing that they jointly explain the central phenomenon identified in the title.

The balanced weight-tying analysis defines a coordinate-invariant structural intervention, whereas the suffix expansion characterizes the response of an unchanged Transformer suffix to uniformly amplified residual inputs.

These are distinct interventions.

In particular, Theorem 1 predicts the difference between applying a final readout directly to a scaled state and applying the original suffix before that readout. It does not directly predict the difference between the outputs of the original and weight-tied networks.

The empirical experiments likewise investigate these effects largely separately. Section 7 evaluates functional changes caused by balanced tying, while Section 8 validates the asymptotic coefficients using scaled inputs to unmodified suffixes.

The paper does not demonstrate that its theoretical coefficients quantitatively predict the propagation or functional damage induced by balanced weight tying.

The gap is especially important at realistic activation scales. For example, Appendix F reports that at Qwen's early cut, the predicted-to-measured forward-KL coefficient ratio is approximately 0.066 at (c=64), increasing toward 0.991 only at (c=32768). This is fully compatible with the asymptotic theorem, but it leaves the practical relevance of the leading-order approximation at normal residual magnitudes unresolved.

Additionally, the rank-two normalization-induced rotation result is primarily an algebraic decomposition. The paper does not directly measure whether this component explains a substantial fraction of the observed residual rotations or functional sensitivity in the evaluated models.

Several ingredients, including normalization-induced Jacobian asymmetry and the geometric interpretation of structural tying instability, are also closely related to the prior framework of Liang (2026), which the paper explicitly builds upon. The novel balanced construction and suffix coefficients are meaningful extensions, but their integration into a stronger explanatory framework remains incomplete.

This weakness is consequential because the claimed contribution is not merely a collection of correct geometric identities, but an account of weight-tying sensitivity in Transformers.

  1. The empirical evaluation is too narrow to establish generality or isolate the proposed geometric mechanism.

Despite the large number of reported experimental conditions, the main empirical evidence comes from only two pretrained checkpoints: Qwen3-0.6B and SmolLM2-360M.

The balanced intervention panel uses three synthetic passages and three natural-text excerpts per model. The independent suffix-coefficient experiment is restricted to the three synthetic passages.

Although the study reports 2,016 primary intervention cells, these cells represent repeated measurements over a fixed set of models, passages, intervention strengths, and controls. They should not be interpreted as independent evidence across architectures or training instances.

The random perturbation baselines are matched using either the Frobenius norm of the weight change or the RMS magnitude of the resulting local branch perturbation. These are useful controls, but they do not match the structured edits in spectral properties, directional alignment, low-rank structure, or their interaction with the residual stream.

Consequently, the larger effects of transpose tying relative to Gaussian perturbations establish that the tested structured modification behaves differently from these random baselines. They do not isolate the specific role of the proposed geometric quantities.

The distinction matters because a highly structured perturbation can be more disruptive than isotropic noise for reasons not uniquely attributable to weight-tying geometry.

Furthermore, the experiments focus on post-hoc interventions without retraining. They establish sensitivity of fixed checkpoints, not whether balanced tying offers a useful or stable parameter-sharing procedure during optimization.

The historical training experiments in the appendices provide some additional context, but the authors acknowledge unavailable checkpoints, incomplete provenance, and unresolved parameter-count discrepancies. These records cannot substitute for reproducible controlled training experiments.

The limited architectural diversity, small evaluation corpus, and insufficiently structure-matched controls restrict the strength and generalizability of the empirical conclusions.

Given that several mathematical ingredients are general properties of normalized residual networks rather than specifically of weight tying, stronger empirical evidence is needed to establish the distinctive significance of the proposed framework.

Questions for authors
  1. Direct connection between the suffix theory and weight-tying interventions (Critical Weakness 1).

Can the authors demonstrate that the proposed residual-geometric quantities or suffix-response coefficients quantitatively predict the functional effects of the balanced tying intervention itself?

For example, could the authors use the actual edit-induced residual perturbations at intermediate layers to predict downstream logit changes, KL divergence, or NLL degradation, and compare those predictions with complete forward evaluations?

In particular, I would be interested in whether direction-resolved sensitivity provides predictive information beyond the magnitude of the local perturbation and whether the resulting predictions remain informative at naturally occurring residual scales.

Such evidence would materially strengthen the connection between the paper's theoretical analysis and its central empirical phenomenon.

  1. Generalization and stronger structural controls (Critical Weakness 2).

How robust are the reported differences between balanced tying and random edits across additional pretrained architectures, model sizes, and natural-language evaluation samples?

More importantly, can the authors compare balanced tying with structured perturbation baselines that match relevant spectral or directional properties, rather than only Frobenius norm or branch-output RMS?

For example, a low-rank perturbation or a perturbation matched in its alignment with the original residual directions could help distinguish effects specific to the weight-tying construction from more general directional sensitivity.

These experiments could materially change my assessment of whether the proposed geometric explanation identifies a distinctive mechanism.

  1. Practical relevance and validation of the proposed geometric decomposition (Critical Weakness 1).

Can the authors directly evaluate the normalization-induced rank-two rotation component in pretrained models and quantify its contribution relative to the remaining antisymmetric Jacobian terms?

Additionally, can they characterize the usefulness or tightness of the finite-radius bounds and endpoint certificates at the original activation scale, rather than predominantly in artificially amplified regimes?

The exact endpoint certificate is mathematically valid but requires knowledge of the realized suffix endpoint. It would be useful to establish whether the proposed quantities yield a practical diagnostic or predictive advantage over simpler baselines based on residual perturbation magnitude, direction change, or direct logit differences.

Clarifying this would help determine whether the geometric formulation provides substantial explanatory or operational value beyond a mathematically consistent re-description of the network.

Reviewer_QGmvOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the geometric changes and functional sensitivity induced by transpose weight tying in pretrained Transformers. The paper argues that, in a SwiGLU feed-forward network, directly replacing the down-projection matrix DD with the transpose of the projection matrix U\TU^\T does not constitute a well-defined, coordinate-consistent model transformation. To address this issue, the authors first balance the rows of UU and the columns of D on a channel basis and then construct a weighted transpose reference. The paper further analyzes how weight edits propagate through RMSNorm and subsequent Transformer layers by decomposing residual states into their magnitudes and directions. This is my understanding of the paper, as the original presentation is genuinely very difficult to follow.

Critical strengths
  1. The paper carefully distinguishes among the magnitude of a local edit, the magnitude of its propagation through the residual stream, and the resulting change in the model’s final behavior. The authors separately measure propagation statistics, NLL, and output KL divergence.

  2. The paper decomposes residual states into magnitude and direction and argues that these two components play different roles in RMSNorm Transformers. The residual magnitude affects the extent to which subsequent computation can change the state direction, whereas the residual direction directly influences the final normalized readout.

Critical weaknesses
  1. The paper is written in an extremely obscure and difficult to follow manner. Before clearly defining the practical research problem, the abstract and introduction introduce a large number of concepts in rapid succession, including the balanced SwiGLU representation, weighted-adjoint tying, normalization-induced rotation, radial-directional coordinates, suffix coefficients, and endpoint certificates. Even researchers familiar with Transformers and matrix analysis may find it difficult, on a first reading, to understand how these concepts relate to one another, let alone determine the central question that the authors ultimately seek to answer.

  2. The paper does not adequately explain why transpose weight tying should be forcibly applied to an already trained Transformer or what practical application this operation is intended to serve.

  3. The experimental scope is relatively limited. The main results are evaluated only on two small language models, Qwen3-0.6B and SmolLM2-360M, using primarily three synthetic passages and three WikiText passages. These experiments are insufficient to establish that the reported phenomena generalize reliably across different model scales or real-world tasks.

Questions for authors

See the weaknesses

Reviewer_yZvQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper mainly studies why transpose weight tying in pretrained Transformers can have a large impact on the model, and tries to explain this problem from a geometric perspective. First, the authors point out that ordinary transpose weight tying is affected by parameter scaling. Therefore, they propose a weight-tying method based on channel balancing to make the editing operation consistent under equivalent parameter representations. Second, the authors analyze how gating and normalization affect residual states, and explain how perturbations propagate through later layers by considering both the magnitude and direction of residual vectors. Based on these analyses, the paper further derives the asymptotic behavior and corresponding coefficients of the influence of the remaining Transformer layers on the final output as the input scale increases.

For the experiments, the authors evaluate their method on two publicly available pretrained language models. The results show that even when the local weight perturbations have similar magnitudes, transpose weight tying and random perturbations can have quite different effects on the internal states and final performance of the models. In addition, the experiments verify the coordinate consistency of the proposed weight-tying method and support the theoretical asymptotic results through tests under different input scaling conditions.

Critical strengths

1. The paper provides a comprehensive and relatively in-depth analysis of the problem.

The authors analyze why transpose weight tying in pretrained Transformers may lead to a significant drop in model performance. They study this problem from several perspectives, including parameter scaling, changes in residual states, and perturbation propagation. The paper not only focuses on the effects of weight editing itself, but also discusses the different roles of the magnitude and direction of residual vectors in later computations. This analysis follows a clear process from model parameters to internal states and then to final outputs. It helps us better understand the sensitivity of Transformers to structural weight editing and also provides a basis for the following theoretical derivations.

2. The theoretical derivations are complete and supported by experimental validation.

Paper provide mathematical derivations, including a weight-tying method with coordinate consistency, local rotation caused by normalization, and the asymptotic behavior and corresponding coefficients of the influence of a fixed Transformer suffix on the final output. The paper not only discusses these problems from a theoretical perspective, but also conducts experiments on two pretrained language models. Through comparisons with random weight edits, coordinate consistency tests, and prediction tests under different input scales, the authors evaluate the main theoretical results.

Critical weaknesses

1. Limited analysis of the underlying mechanisms.

Although the paper uses several control experiments to show that transpose weight tying and random weight edits can have significantly different effects on the models, the experiments do not fully distinguish the contributions of factors such as weight spectral structure, gating, and normalization. Therefore, although the paper provides geometric theoretical analysis, the relationship between these theoretical factors and the observed experimental results still needs to be further clarified.

2. The organization and explanation of the theoretical parts could be improved.

The paper introduces several theoretical results, including channel balancing, local rotation caused by normalization, and the asymptotic analysis of Transformer suffixes. Although the authors provide mathematical definitions and derivations, with detailed proofs in the appendix, the main text gives relatively limited explanations of the intuitive meaning of some key formulas and the connections between different theoretical results. Adding more explanations or concrete examples would help readers better understand the theoretical analysis and the main contributions of the paper.

Questions for authors

1. The experiments show that transpose weight tying and random weight edits can have different effects even with similar local perturbation magnitudes. Can the authors clarify which geometric factors mainly contribute to these differences, and what evidence supports this explanation?

2. The paper presents several theoretical results, including coordinate-invariant weight tying and the asymptotic analysis of Transformer suffixes. Can the authors further clarify how these results are connected and how they jointly contribute to understanding weight-tying sensitivity?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.