acceptodds
Under review as a conference paper at ICLR 2027

The Missing Semantic Reference: Consensus Geometry and Observable Dynamics in Language Models

Abstract

Intermediate states are increasingly used to read, compare, and manipulate the internal computation of language models, yet they are measured in native residual coordinates that carry no scientific reference. An invertible change of hidden basis leaves the model's function intact but changes every Euclidean length, angle, and shrinkage penalty computed from raw activations. We call this the missing reference frame. The model itself supplies the information needed to fix it: under a hidden reparameterization , the input embedding and the output decoder transform reciprocally, and balancing these two lexical interfaces has a unique positive-definite solution, the matrix geometric mean of their Gram operators, which is the one reference in which both interfaces induce the same geometry. Because coordinatewise models, shrinkage, and the direction of a lexical gradient are all defined relative to a reference, this model-induced reference is indispensable for results that describe the model rather than its parameterization. builds on it in three modules. Consensus Geometry fixes this data-free, model-wide reference and reads every state through the decoder's lexical covectors as a vocabulary-indexed Semantic State. Observable Dynamics estimates, without any decoder signal, the part of the terminal state that a source layer predictably supports, which defines Semantic Transitions across depth. Information Projection passes the Riesz-lifted exact forward-KL gradient through the same frozen response operator and allocates three coefficients in a globally convex problem. Across ten models from four families, SmolLM2, MiniCPM5, Granite 4.2, and Ministral 3 (135M to 14B parameters, 322 layers), resetting the Euclidean identity after a chart change or using a lexical covector as a vector distorts the measured quantity by 14% to 92%. lowers macro held-out forward KL from 2.984 for a full-affine Tuned Lens trained on the same data to 2.033, improves every layer and every held-out sequence, and also improves NLL and next-token accuracy on every model. The decoder-free Observable Dynamics alone accounts for most of this gain, and both the lens and its Semantic States remain stable as calibration data grow, while Semantic Transitions need the largest budgets.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Cross-Layer Latent Consistency, Representation Geometry, Semantic Dynamics, Diagnostic Lenses

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_zPENOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

Lexical Lens separates representation analysis into a reference metric derived from the input and output lexical interfaces, decoder-free prediction of terminal states, and a three-coefficient convex forward-KL calibration. The paper argues that this construction permits coordinate-consistent measurement of semantic states and transitions. Across ten language models and 322 layers, it reports macro held-out KL of 2.033 versus 2.984 for Tuned Lens, with most of the improvement already obtained by the decoder-free predictor.

Critical strengths
  1. Useful separation of geometry, prediction, and calibration. The balancing objective for the raw positive-definite lexical Grams has a unique matrix-geometric-mean solution with the stated covariance law. Once the response fields and affine decoder are frozen, the three-coefficient KL objective is convex and contains the decoder-free prediction as a feasible anchor. These are substantive mathematical components with clear roles. [Sections 3.1–3.3; Appendix A.6]

  2. Broad and carefully grouped predictive evaluation. The experiments span ten models, use disjoint calibration and test sequences, and select the response scale by excluding complete sequences. Improvements are reported on all 322 layers and 640 held-out sequences, with secondary metrics and calibration-budget analyses. The decoder-free ablation supports the practical value of the terminal-state estimator. [Section 4; Appendices A.2 and A.5; Tables 2 and 6]

Critical weaknesses
  1. The stated implementation does not establish covariance of the complete estimator. States are represented as X=xUX=xU with orthogonal UU, while the metric is stored separately. Consequently, UU cancels from the ridge response in Eq. (3), leaving an identity penalty in native coordinates; using the metric only to select the scale does not repair this. For X=I2X=I_2 and T=diag⁡(a,a−1)T=\operatorname{diag}(a,a^{-1}), matching the original response after transformation requires both ρ′=a2ρ\rho'=a^2\rho and ρ′=a−2ρ\rho'=a^{-2}\rho, which are incompatible for a>0a>0, a≠1a\ne1, and ρ>0\rho>0. Appendix A.3 also describes an identity-array Riesz lift despite the nonidentity metric, while Appendix A.1's unit-determinant normalization removes the metric scaling required under T=cIT=cI. Undocumented reciprocal coordinates could resolve some discrepancies, but the necessary conventions are absent. Table 1 tests incorrect shortcuts rather than covariance of the refitted full estimator. [Eq. (3), p. 4; Appendices A.1–A.3, p. 13; A.5–A.6, p. 15]

  2. The predictive comparisons do not isolate the contribution of the consensus reference. The proposed shrinkage estimator is compared with an Adam-trained affine translator, without a native-coordinate state-ridge baseline using the same sequence-excluded selection or controls using individual lexical interfaces. In addition, Tuned Lens applies final normalization after its translator, whereas the proposed convex fit uses affine logits in the predicted state. These differences leave regularization and output normalization as alternative explanations for the KL improvement. The results support predictive utility more directly than the claimed necessity of the consensus geometry. [Sections 4.1–4.3; Appendix A.5]

Questions for authors
  1. What exact coordinate transformations implement the ridge operator, diagonal backbone, and Riesz lift? Please reconcile the identity-lift statement and unit-determinant normalization with the metric law, and establish whether this requires clarifying existing code or changing the estimator.

  2. Does refitting the implemented pipeline commute with dilation, determinant-one anisotropic scaling, and shear? Resolving the first concern requires an explicit derivation addressing the counterexample above and numerical agreement of transported held-out predictions and semantic transitions, including any necessary refits.

  3. How much improvement remains with matched state-ridge, single-interface metric, and output-normalization controls? A higher rating would require these controls to substantiate the consensus reference's contribution alongside a consistent implementation of the claimed geometry.

4 comments

Thank you for your detailed review, especially the counterexample you provided and the experimental controls you suggested. We address your concerns point by point below.

W1 / Q1: Exact coordinate transformations and the counterexample.

You are right that UU cancels from the native ridge response. Your two-axis example shows why retuning one scalar cannot repair a penalty reset to identity. The submitted c=xUc=xU fit uses native normalization for the ridge penalty and gradient lift, and the balanced token interfaces for the backbone axes and State-risk metric. Our added code carries these choices into a new chart to refit the same estimator after normalized State extraction (Sections 3.1–3.3 and Appendix A.6).

Under x′=xTx'=xT, the native normalization metric becomes N′=T−1NT−⊤N'=T^{-1}NT^{-\top} and the penalty is ρ(N′)−1\rho(N')^{-1}. In your X=I2X=I_2 example, ρI\rho I becomes ρT⊤T\rho T^\top T, so T(T⊤T+ρT⊤T)−1T⊤=(1+ρ)−1IT(T^\top T+\rho T^\top T)^{-1}T^\top=(1+\rho)^{-1}I. The original ρ\rho now recovers the response on both axes.

The Riesz lift uses the native normalization metric, so its matrix is the identity in native and orthogonal execution coordinates where N=IN=I. For a native decoder covector Γ\Gamma, the lifted State direction in the new chart is (ΓT−⊤)(N′)−1=(ΓN−1)T(\Gamma T^{-\top})(N')^{-1}=(\Gamma N^{-1})T.

The backbone fits one slope per balanced axis, giving a generally non-diagonal map in native coordinates. If AA balances the unit-determinant interfaces, c=xA(A⊤N−1A)−1/2c=xA(A^\top N^{-1}A)^{-1/2} reduces to xUxU in the native chart. Rebuilding the axes gives c′=cSc'=cS and B′=BSB'=BS for the same diagonal sign matrix SS. The raw Consensus metric follows the law for NN. Fixing its determinant multiplies the transported metric by ∣det⁡T∣2/d\lvert\det T\rvert^{2/d}. The square root of this scalar cancels in the normalized State-risk factor, so dilation preserves every candidate's score.

Transporting the decoder preserves logits and KL for every final coefficient triple. Corresponding optima give x^′=x^T\widehat x'=\widehat xT, identical token readings and identical transitions. This assumes exact arithmetic, positive-definite interfaces and a distinct, consistently ordered balanced spectrum. Repeated eigenspaces and nonunique optima require consistent choices.

W1 / Q2: Complete refits under dilation, anisotropic scaling and shear.

Independent refits under dilation, anisotropic scaling, shear and rotation agreed with the native predictions on the original ten models (322 layers). Pooling held-out positions at each layer, the maximum relative error was below 0.019% for Final predictions and 0.066% for Semantic Transitions. These refits supplement the shortcut tests in Table 1.

In each chart we rebuilt the axes from the transformed interfaces and N′N' and refitted the backbone, ridge scale and projection. State and Final denote native-coordinate predictions before and after Information Projection. Semantic Transitions compare adjacent-depth token readings before projection. Entries are maximum relative Frobenius errors in fractions. At a single held-out position, the largest relative Transition error was 0.451%.

Analysis-coordinate changeState predictionFinal predictionSemantic Transition
Dilation (×4)4.53e-051.30e-043.76e-04
Anisotropy (det=1, κ=8)9.62e-051.82e-046.57e-04
Shear (0.5)5.96e-051.61e-046.03e-04
Rotation7.28e-051.51e-044.65e-04

The largest Final discrepancy with transport occurred under anisotropic scaling at layer 12 of Granite 4.2 8B. Against the same native fit and at the same precision, the relative error was 11.42% when we reset the metric to identity and 0.0182% when we carried it. Resetting the penalty changes the relative shrinkage along the two axes in your example. With the metrics carried and the axes rebuilt, fitting before or after the coordinate change recovered the same held-out final State and depthwise change in token readings to the accuracy shown above.

W2 / Q3: Predictive controls and the role of Consensus Geometry.

State-ridge control. Across all ten models, Lexical Lens reduced held-out KL relative to native ridge by 0.13%–4.35% before Information Projection and 0.81%–5.07% after it. When both fits selected shrinkage by Euclidean State error, the balanced-axis backbone still lowered Final KL by 0.75%–5.00% versus plain ridge on every model. These controls use separate fits on the models and evaluation data of Table 2.

Both fits use the native ridge penalty and select its strength by leaving out whole sequences. Native ridge predicts the terminal State directly and scores Euclidean errors. Lexical Lens fits a slope and intercept per balanced axis, applies ridge to the residual, and scores Consensus-weighted errors. Each fit receives its own final KL projection on the same data. Entries are layer-mean held-out KL in nats (322 layers total).

ModelState: Lexical Lens*State: ridgeFinal: Lexical LensFinal: ridge
SmolLM2 135M1.8971.9001.7421.756
SmolLM2 360M2.0102.0291.8011.822
SmolLM2 1.7B2.2772.3021.9822.020
MiniCPM5 1B1.6981.7071.5791.596
MiniCPM5 2B2.1422.1561.9451.978
Granite 4.2 3B2.7942.8452.3192.386
Granite 4.2 8B3.2123.3012.6952.806
Ministral 3 3B2.3282.3682.0202.083
Ministral 3 8B2.4792.5402.1532.242
Ministral 3 14B2.3962.5052.0922.204

All 20 paired 95% intervals favor the submitted fit over plain ridge (held-out sequences resampled with fits fixed). Under Euclidean scale selection, the native-axis backbone lowered mean Final KL by 0.73%–4.95% relative to plain ridge, accounting for most of the balanced-axis backbone's gain on each model. Changing only to balanced axes further lowered mean Final KL on nine of ten, with all changes below 0.30% in magnitude.

Single-interface control. Across all ten models and 322 layers, we changed only the metric for scoring excluded sequences, keeping the balanced axes, native penalty and gradient lift fixed. Negative changes in KL favor the alternative.

State-risk metricSmallest Final-KL changeLargest Final-KL change
Identity II−0.222%+0.205%
Input metric GE−1G_E^{-1}−0.355%+0.052%
Output metric GOG_O−0.592%+0.701%

Changing the scale-selection metric alone moved Final KL by less than 0.71% in either direction. On four models (118 layers), we also rescaled the jointly balanced axes to use one interface for the penalty, lift and scale selection together. Final KL rose on all four, by 0.004–0.152 nats for the input metric and 0.005–0.057 for the output metric.

Output-normalization control. The Lexical Lens readout stayed affine. Switching on the translator's output normalization lowered KL in all eight pairs with raw or normalized inputs on four models (118 layers). Only output normalization varied within each pair. The translator used 56 sequences to fit and eight to validate. Lexical Lens used all 64. Both used the same 64 held-out sequences.

Held-out KLSmolLM2 360MMiniCPM5 1BGranite 4.2 3BMinistral 3 3B
Raw source, output norm on2.5862.4023.1722.676
Raw source, output norm off71.69612.89313.6942.927
Normalized source, output norm on2.5952.4023.1412.679
Normalized source, output norm off3.5894.9027.6144.691

With normalized inputs and output normalization enabled, translator KL was still 0.66–0.82 nats higher than Final Lexical Lens. Removing normalization to match the affine readout widened that gap on all four models.

Role of the Consensus reference. The input metric measures a State displacement relative to the spread of token embeddings. The output metric measures its effect on centered decoder logits. Each raw metric is covariant. Consensus supplies one ruler across depths by minimizing the larger of its affine-invariant distances to the two interfaces, before any predictor is fitted. Appendix A.6 gives the midpoint proof and Table 7 the measurements.

Balancing gives GCGEGC=GOG_CG_EG_C=G_O, with unique positive-definite solution GC=GE−1♯GOG_C=G_E^{-1}\mathbin{\sharp}G_O. For determinant-normalized interfaces a distance LL apart, mean-square distortion is L2/4L^2/4 at Consensus and L2/2L^2/2 at either interface alone.

ModelsIdentity excess distortion over Consensus
4 models with tied embeddings6.08%–15.90%
6 models with untied embeddings22.36%–96.92%

Using only frozen weights, we found greater distortion for native identity on every model. In the submitted fit, joint balancing sets the backbone axes, Consensus weights State errors for scale selection, and native normalization supplies the ridge penalty and gradient lift. The revision gives the per-model distances and distortions.

Thank you for working through the two-axis example. Appendix A.6 in the revised PDF clarifies which metric supplies the ridge penalty and gradient lift and how it transforms, which resolves my covariance concern under the stated assumptions. The numerical refits under dilation, anisotropy, shear, and rotation also confirm that the predicted states and transitions agree in practice. Those refit tables and the predictive control tables are currently only in your comments rather than the PDF, so it would help to include them alongside the proof in the appendix.

The matched-ridge and single-interface controls also clarified what Consensus Geometry contributes to prediction. While the full predictor beats matched ridge on all ten models (and Table 8 shows it remains ahead of the Muon-trained Tuned Lens), your breakdown shows that with Euclidean scale selection, the native-axis backbone already accounts for nearly all of the gain over plain ridge (0.73% to 4.95%). Switching to the balanced Consensus axes changes final KL by less than 0.30% per model, and changing only the State-risk metric to the identity moves final KL by less than 0.25%. The geometric motivation for a common reference is clear, but its actual contribution to predictive performance looks quite small here, so I am keeping my rating at 2.

The remaining disagreement concerns the role of Consensus Geometry within Lexical Lens. The submitted abstract describes it as fixing a “data-free, model-wide reference,” and the introduction explicitly separates reference, dynamics, and readout. Before any layerwise predictor is fitted, the two frozen token interfaces determine one set of fitting axes and one metric for scoring State errors across all depths. Observable Dynamics then estimates what each source layer predicts about the terminal State. The submitted abstract attributed most of the gain over Tuned Lens to this decoder-free prediction stage.

Table 7 shows why the choice of reference matters. Relative to Consensus, choosing either interface alone doubles the mean-square distortion to the pair, and using the native Euclidean metric increases it by 6.08%–96.92% across the ten models. With this common reference, successive State predictions read through the same frozen decoder show how predictable support for each token changes between layers. Your follow-up confirms that the covariance concern is resolved under the stated assumptions. Across all ten models and four tested coordinate changes, complete refits reproduced these transitions. The maximum layerwise relative error was below 0.066% when held-out positions were pooled within each layer.

Both ablations retain a metric and fitting axes shared across layers. Each model’s mean final KL changed by less than 0.25% when only the shrinkage-selection metric was replaced by the Euclidean metric, and by less than 0.30% when only the axes were changed under Euclidean scale selection. These figures show that final KL is relatively insensitive to those substitutions. The separate matched-ridge control directly fitted the terminal State in native coordinates. With each fit given its own final projection, Lexical Lens achieved 0.81%–5.07% lower held-out KL on all ten models. The two ablations alone are therefore insufficient grounds for retaining a 2 for the complete method. We ask you to reconsider the rating and the Area Chair to weigh both the predictive comparison and the verified cross-layer measurements in the meta-review.

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

This paper proposes Lexical Lens, a framework for analyzing intermediate representations in language models through a model-derived geometric reference. The method consists of three components: Consensus Geometry, Observable Dynamics, and Information Projection. It aims to provide coordinate-consistent semantic representations and improve the prediction of terminal token distributions. Experiments across 10 language models from 4 families show improvements over Logit Lens and Tuned Lens in forward KL, NLL, and next-token accuracy.

Critical strengths
  1. This paper addresses an important issue in the geometric interpretation of language model representations, with a clear theoretical motivation.

  2. The experiments cover multiple model families and sizes, showing consistent improvements over the reported baselines.

  3. The paper presents a well-structured methodology with clear and well-articulated motivations.

Critical weaknesses
  1. Potential inconsistency in coordinate covariance. Section 3.1 claims GL(d)GL(d) covariance, while Appendix A.1 normalizes both Gram matrices to unit determinant, potentially introducing an additional scale factor. The authors should clarify how this normalization preserves the claimed covariance under general coordinate transformations.

  2. Insufficient validation of Consensus Geometry. The uniqueness proof establishes the solution to a specific optimization objective but does not demonstrate its superiority over other valid geometric references. Comparisons with alternative references and verification of the complete pipeline's coordinate covariance are missing.

  3. Limited experimental controls. The performance gains may partly arise from stronger structural constraints and regularization rather than Consensus Geometry itself. More comparable baselines, such as regularized or low-rank Tuned Lens variants, are needed to isolate these effects.

  4. Insufficient evidence for semantic interpretability. The experiments primarily demonstrate predictive accuracy and stability rather than the semantic validity of the proposed representations. Further evaluation is needed to establish whether Semantic States and Semantic Transitions capture meaningful semantic information, particularly across models with different tokenizers.

Questions for authors
  1. How is the unit-determinant normalization in Appendix A.1 compatible with the claimed GL(d) covariance, particularly under global scaling transformations?

  2. Can the authors compare Consensus Geometry with alternative valid geometric references and verify the covariance of the complete Lexical Lens pipeline?

  3. Can the authors include regularized or low-rank Tuned Lens baselines to isolate the contributions of geometric reference construction and structured state estimation?

  4. What direct evidence supports the semantic validity of Semantic States and Semantic Transitions, and how are these representations compared across models with different tokenizers?

The necessity of Consensus Geometry remains insufficiently validated. I would consider revising my score if the authors could provide convincing experimental or theoretical evidence supporting its essential role.

2 comments

Thank you for recognizing the theoretical motivation for our approach and for suggesting direct comparisons with alternative geometric references. We address your concerns point by point below.

W1 / Q1: Unit-determinant normalization and global scaling.

The raw metric obeys the tensor law. Fixing its determinant adds the common scale factor you identified, which cancels in ridge selection. Under global scaling T=aIT=aI with a>0a>0, the normalized reference stays fixed while raw GCG_C and the native normalization metric NN scale by a−2a^{-2}. The fitted State scales as x^′=ax^\widehat x'=a\widehat x and decoder rows are divided by aa, preserving every logit and token probability. See Section 3.1 for the metric, Appendix A.1 for its scale and Appendix A.6 for the proof.

For x′=xTx'=xT, the raw law is GC′=T−1GCT−⊤G_C'=T^{-1}G_CT^{-\top}. Normalizing both Grams gives GˉC=GC/det⁡(GC)1/d\bar G_C=G_C/\det(G_C)^{1/d}, hence GˉC′=γ2T−1GˉCT−⊤\bar G_C'=\gamma^2T^{-1}\bar G_CT^{-\top} with γ=∣det⁡T∣1/d\gamma=\lvert\det T\rvert^{1/d}. The square root of the multiplier cancels when we divide the State-risk factor by the dd-th root of its determinant.

With native N=IN=I carried as N′=T−1NT−⊤N'=T^{-1}NT^{-\top}, the rebuilt coordinates c′=cSc'=cS differ only by axis signs. Each ridge scale gets the same error score, and each final coefficient triple gives the same KL. This holds after normalized State extraction in exact arithmetic, with positive-definite interfaces, a distinct, consistently ordered balanced spectrum and corresponding optima. Repeated eigenspaces require a consistent axis convention.

W2 / Q2, Q5: Alternative references, complete covariance and the essential role of Consensus Geometry.

Either raw interface metric is covariant. Consensus is the unique ruler that minimizes the larger affine-invariant distance to the two interfaces and halves mean-square distortion relative to either interface alone. It weights the terminal-State prediction errors used to select ridge strength at every depth. Refitting preserves token readings when we transport the Consensus and native normalization metrics and rebuild the axes. Appendix A.6 and Table 7 give the proof and interface measurements. The controls and refits below are additional.

The input metric measures a State change relative to the spread of token embeddings. The output metric measures its effect on centered decoder logits. Balancing the two Gram operators gives GCGEGC=GOG_CG_EG_C=G_O, with unique positive-definite solution GC=GE−1♯GOG_C=G_E^{-1}\mathbin{\sharp}G_O. In coordinates where Consensus is the identity, the determinant-normalized embedding and decoder Grams coincide. The two token arrays then have the same directional spread up to overall scale. At interface separation LL, the equally weighted mean-square distortion is L2/4L^2/4 at this midpoint and L2/2L^2/2 at either interface.

After removing overall scale, the interface distance LL was 0.603–1.100. For mean-square distortion JJ, native identity's excess over Consensus is 100[J(I)/J(GC)−1]100[J(I)/J(G_C)-1] percent.

ModelTied embeddingsInterface distance LIdentity excess distortion (%)
SmolLM2 135MYes1.10011.01
SmolLM2 360MYes0.95415.90
SmolLM2 1.7BYes1.0437.11
MiniCPM5 1BNo0.60396.92
MiniCPM5 2BNo0.74695.06
Granite 4.2 3BNo0.98922.36
Granite 4.2 8BNo1.00424.30
Ministral 3 3BYes0.7856.08
Ministral 3 8BNo0.93493.68
Ministral 3 14BNo0.97692.29

The two interface metrics differ even on the four tied models, where shared weights enter through an inverse input Gram and a direct output Gram. Across all ten models, native identity incurred 6.08%–96.92% more distortion than Consensus. This geometric advantage is already present in the frozen weights.

With backbone form and Euclidean scale selection fixed, switching from native to balanced axes lowered mean Final KL on nine of ten models, with all changes below 0.30% in magnitude. With axes, native penalty and lift fixed, changing only the selection metric (GCG_C to II, GE−1G_E^{-1} or GOG_O) moved Final KL by −0.592% to +0.701%.

On four models (118 layers), we rescaled the jointly balanced axes to use one interface for the penalty, lift and scale selection together. Final KL rose on all four, by 0.004–0.152 nats for the input metric and 0.005–0.057 for the output metric.

On all ten models (322 layers), we rebuilt the axes and refitted the backbone, ridge scale and projection under dilation (×4), determinant-one anisotropy (κ=8), shear (0.5) and rotation. Entries are relative Frobenius errors over all held-out positions at each layer, expressed as fractions.

QuantityMaximum relative error
Final prediction1.82e-4
Module-II Semantic Transition6.57e-4

Back in native coordinates, the maximum error was below 0.019% for Final predictions and 0.066% for changes in token readings between adjacent depths before projection.

W3 / Q3: Regularized and low-rank Tuned Lens baselines.

The low-rank translator you suggested improved on the original Tuned Lens on all ten models. Lexical Lens still had lower held-out KL on each. It also had lower KL than regularized Tuned Lens on all four models tested. Appendix B.4 (Table 8) gives the ten-model comparison. The separate four-model control uses its own Lexical Lens fits.

Each comparison used 64 calibration sequences and the same 64 held-out sequences per model. In the four-model test (118 layers), regularized Tuned Lens used 56 calibration sequences to fit and eight to select weight decay and checkpoint. Lexical Lens fitted on all 64. Entries are layer-mean forward KL in nats.

ModelRegularized Tuned LensLexical Lens
SmolLM2 360M2.5641.801
MiniCPM5 1B2.3791.579
Granite 4.2 3B3.1082.319
Ministral 3 3B2.6552.020

On the ten models (322 layers), the low-rank variant adds a rank-64 update to identity under the original training settings. The Muon variant fits a full-affine translator. Both retain output normalization.

ModelTuned LensLow-rank variantMuon variantLexical Lens
SmolLM2 135M2.3722.2692.1701.742
SmolLM2 360M2.5732.4191.9851.801
SmolLM2 1.7B2.9542.7942.4301.982
MiniCPM5 1B2.6782.0661.8401.579
MiniCPM5 2B3.1172.5282.1531.946
Granite 4.2 3B3.2602.9162.7222.322
Granite 4.2 8B3.7613.3273.0312.695
Ministral 3 3B2.7082.4162.1452.020
Ministral 3 8B3.1462.6252.3012.153
Ministral 3 14B3.2672.7232.2322.092

The Muon variant was the strongest translator on each model after 100 passes over the same 8,128 calibration positions, about 6.1 times the original Tuned Lens mean of 16.3 passes. Lexical Lens retained lower held-out KL on all ten, averaging 2.033 nats versus 2.301 for Muon with equal model weights. Against both strengthened translators, the closer predictions came from fitting the terminal State first and then calibrating three readout coefficients. The fixed-backbone comparison in Q2 isolates the effect of the axes.

W4 / Q4: Semantic validity and different tokenizers.

A State contrast gives the predicted log odds of one next token relative to another. A Transition contrast gives their change across depths. Across models with different tokenizers, we compare fidelity and stability scores computed within each model's own vocabulary. A direct comparison of individual token readings would require matched inputs and aligned tokens and positions. See Section 3.1 for lexical meaning, Section 4.2 (Table 2) for fidelity, and Table 3 and Figure 8 for stability.

The same frozen decoder rows identify tokens vv and ww at every depth. In the decoder distribution pcp_c, the common softmax denominator cancels, giving s(c)v−s(c)w=log⁡[pc(v)/pc(w)]s(c)_v-s(c)_w=\log[p_c(v)/p_c(w)] exactly. If the Transition contrast is log⁡2\log 2, the deeper source predicts twice the odds of the first token relative to the second at the same position.

We decoded the States used for Semantic Transitions and compared their token probabilities with the model's terminal probabilities at the same held-out positions. Mean forward KL across ten equally weighted models was lower for Lexical Lens* (2.323 nats) than for Tuned Lens (2.984 nats). These are the token preferences we interpret across depths, and they already track the model's output more closely before Information Projection fits any coefficient to token probabilities.

To test stability, we fitted each lens on 32 sequences and on the same set extended to 64. At fixed held-out positions we correlated the two fits' vocabulary readings for States and for adjacent-depth changes. Each range spans the models in one family.

Model familySemantic State correlationSemantic Transition correlation
SmolLM20.972–0.9800.759–0.824
MiniCPM50.974–0.9760.798–0.818
Granite 4.20.987–0.9880.822–0.829
Ministral 30.983–0.9850.786–0.800

On every model, the two fits recovered a similar pattern of which tokens gained or lost support across a block, with closer agreement on the States themselves. The correlations let us compare reproducibility across tokenizers while interpreting each token-pair contrast within its own vocabulary. The revision gives per-model fidelity scores and tracks both correlations across all four calibration sizes.

Reviewer_AcH9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors argue that residual connections do not yield a uniquely fixed (canonical) basis because the representations allow for arbitrary linear reparameterizations (invariance), which leave the network's function unchanged and invalidate coordinate-dependent analyses. To address the limitations of current analysis methods, the authors propose "Lexical Lens," consisting of three modules.

The first module builds the reference basis through the model's input embeddings and output distribution. Next, the second module predicts the terminal state from intermediate layers within this reference to capture observable dynamics. Finally, the third module provides a lens-specific addition that projects the fixed fields to the final distribution.

Experiments demonstrate that Lexical Lens achieves significantly lower predictive KL across layers compared to the Logit Lens and Tuned Lens.

Critical strengths
  1. Lexical Lens outperforms existing baselines such as Tuned Lens while requiring fewer computational resources for fitting.
  2. The KL curves of Lexical Lens fall steeply over the last third of the layers, demonstrating its stronger ability to approximate the final output distribution across layers.
Critical weaknesses
  1. The paper claims to make model dynamics observable; however, these dynamics are observed only indirectly by mapping intermediate states to terminal-state predictions (ZCZ_C) rather than tracking true layer-to-layer transitions.
  2. Although RQ1 demonstrates the harm of analyzing raw activations directly, it doesn't show that the method for Lexical Lens (Module I) is helpful to reduce this issue.
  3. Owing to Massive Activations [1], the model's dynamics may collapse at specific sequence positions. The extreme magnitude of hidden states at these tokens risks corrupting the Gram matrix estimation in Module II.

[1] Sun, M., Chen, X., Kolter, J. Z., & Liu, Z. (2024). Massive Activations in Large Language Models. First Conference on Language Modeling. Retrieved from https://openreview.net/forum?id=F7aAhfitX6

Questions for authors

Please see Weaknesses.

1 comment

Thank you for recognizing both the predictive accuracy of Lexical Lens and its low fitting cost. We address your concerns point by point below.

W1: Observable Dynamics and layer-to-layer transitions.

You are right that we observe the transition indirectly. At a fixed input position we use the States entering and leaving a block to predict the same terminal State. The change between the lexical readings of those two predictions is the transition we measure (Section 3.2).

Reading both predictions with the same frozen decoder gives τℓ=s(ZC,ℓ+1)−s(ZC,ℓ)\tau_\ell=s(Z_{C,\ell+1})-s(Z_{C,\ell}). A positive Transition contrast for a token pair means the deeper source predicts higher terminal odds of the first token relative to the second. This lets us locate the depth at which a predicted preference strengthens or weakens, even when the most likely token stays the same on both sides of the block.

W2: Positive evidence that Module I addresses the reference problem.

Carrying the Module I reference into the new coordinates preserved the measured source-to-terminal State differences on all ten models. After accounting for the common length unit, maximum relative error in the reference-weighted differences was below 1.27e-15. These positive checks supplement the shortcut tests in Table 1. Table 7 compares reference choices.

At each calibration position we also took the State's dot product with the decoder rows averaged under the terminal token probabilities. We repeated it after transforming State and covector reciprocally.

Positive checkMaximum relative error
Interface balance4.11e-14
Transported State geometry1.27e-15
Preservation of State–lexical covector pairings6.40e-16

To test the choice of ruler itself, we also compared distortion to the two frozen interfaces before fitting. With equal weight on both, either interface alone incurs twice the mean-square distortion of Consensus.

ModelsIdentity excess distortion over Consensus
4 models with tied embeddings6.08%–15.90%
6 models with untied embeddings22.36%–96.92%

Carrying the reference preserves the source-to-terminal separation, and reciprocal transport of the decoder covector preserves the lexical pairing. Consensus also gives lower mean-square distortion across the two interfaces than native identity on every model. Module I thus supplies one ruler, fixed from token weights, for comparing depths in either chart.

W3: Massive activations and Gram estimation.

Normalization bounds each token's contribution to the source Gram even when its raw activation is extreme. Module II uses the normalized State before the learned per-channel gain, which we absorb into the frozen decoder (Appendix A.2).

Under RMSNorm or LayerNorm with positive epsilon, each row xx contributes x⊤xx^\top x to X⊤XX^\top X with operator norm ∥x∥2≤d\lVert x\rVert^2\le d, where dd is the State dimension. Several normalized States can still line up and concentrate the Gram in one direction. The revision gives the bound and cites the study you recommended.

Comments on the paper

1 comment

We thank the reviewers for their careful reading and constructive suggestions. The revision retains the submitted approach and main results. We have also carefully revised the exposition to clarify key definitions, assumptions, and methodological choices. It clarifies coordinate consistency for the complete estimator (Sections 3.1–3.3 and Appendix A.6) and adds comparisons of lexical references (Appendix B.3, Table 7) and stronger baselines (Appendix B.4, Table 8). We respond to each reviewer separately, addressing each weakness and question. The corresponding replies present the relevant results, including additional analyses beyond those reported in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.