acceptodds
Under review as a conference paper at ICLR 2027

Artificial Entanglement in the Fine-Tuning of Large Language Models

Abstract

Large language models (LLMs) can be adapted to new tasks using parameter-efficient fine-tuning (PEFT) methods that modify only a small number of trainable parameters, often through low-rank updates. In this work, we adopt a quantum-information-inspired perspective to understand their effectiveness. From this perspective, low-rank parameterizations naturally correspond to low-dimensional Matrix Product States (MPS) representations, which enable entanglement-based characterizations of parameter structure. Thereby, we term and measure *"Artificial Entanglement"*, defined as the entanglement entropy of the parameters in artificial neural networks (in particular the LLMs). We first study the representative low-rank adaptation (LoRA) PEFT method, alongside full fine-tuning (FFT), using LLaMA models at the 1B and 8B scales trained on the Tulu3 and OpenThoughts3 datasets, and uncover: (i) *Internal artificial entanglement* in the updates of query and value projection matrices () in LoRA follows a *volume law* with a central suppression (termed as the "Entanglement Valley"), which is sensitive to hyper-parameters and is distinct from that in FFT; (ii) *External artificial entanglement* in attention matrices, corresponding to token-token correlations in representation space, follows an *area law* with logarithmic corrections and remains robust to LoRA hyper-parameters and training steps. Drawing a parallel to the No-Hair Theorem in black hole physics, we propose that although LoRA and FFT induce distinct internal entanglement signatures, such differences do not manifest in the attention outputs, suggesting a "no-hair" property that results in the effectiveness of low rank updates. We further provide theoretical support based on random matrix theory, and extend our analysis to an *MPS Adaptation* PEFT method, which exhibits qualitatively similar behaviors.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Parameter-Efficient Fine-Tuning, Low-Rank Adaptation, Tensor Networks, Entanglement Entropy

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_5XeTOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper tensorizes weight updates, attention matrices, and output Gram matrices into matrix-product-state chains and measures singular-value entropy across their cuts. LoRA updates exhibit concentrated spectra and a central entropy valley, while attention-related entropies change less across adaptation methods. The authors interpret these observations as a no-hair-like explanation of LoRA’s effectiveness, supported by random-matrix calculations and an MPS adapter extension.

Critical strengths
  1. Meaningful controls separate spectral concentration from nominal rank. Appendix J matches trainable Q/V projections and uses three paired seeds. Appendix K truncates all transplant arms to rank 256 and exchanges spectra and singular vectors. Transplanting the LoRA spectrum closes 97.71% of the off-center profile gap and 97.98% of the valley-depth gap. These results go beyond the elementary rank constraint. [Appendices J–K, pp. 51–54, Tables 3–4]

  2. The descriptive observations receive substantial robustness checks. Appendix L tests 21 block-preserving factor orders and eight basis transforms; Appendix M includes all 32 heads at three layers and 100 unpadded nontraining inputs. Appendix I.3 also reports breakdown under extreme scaling. These strengthen the descriptive contribution within the tested regimes. [pp. 41–42, 54–55]

Critical weaknesses
  1. Entropy similarity does not establish functional invariance. At a rank-rr update’s row–column cut, the bound S≤log⁡rS\le\log r follows from rank alone; the matched-rank controls add evidence about spectral concentration, but neither fact explains comparable adaptation performance. Normalized singular-value entropy discards scale and orientation. Even identical output Gram matrices can accompany different downstream predictions under fixed downstream weights. Appendix K evaluates transplanted entropy profiles, and Appendix M compares entropy at loss-matched checkpoints; neither directly demonstrates the proposed suppression of functional differences. [Abstract; Figure 1; pp. 8–9; Appendices K–M]

  2. The theory contains a false bound and does not directly match the empirical observable. In Lemma E.2/F.2, M=I2M=I_2 gives δ=1\delta=1 and an entropy upper bound of zero, although S=log⁡2S=\log 2. The attention asymptotic also concerns full row–column entropy as sequence length grows, whereas the experiments primarily scan tensorized cuts at fixed length. For T=2nT=2^n, there are 2n2n artificial sites, so entropy proportional to log⁡T\log T is extensive in that site count. Moreover, the output theorem concerns XX, whereas Figure 6 tensorizes XX⊤XX^\top. These distinctions undermine the current area-law interpretation without further justification. [pp. 22–28; Theorems F.1/F.3; Figure 6]

Questions for authors
  1. In the existing rank-matched transplantation experiment, how do attention/output distances, predictive KL, and held-out loss change alongside entropy? Evidence connecting the spectral interventions to functional behavior would address the proposed explanation of LoRA's effectiveness.

  2. What subsystem and size variable define the claimed area law, and how do the theorems connect to the trained causal-attention and output-Gram observables? Please identify which measured quantities the asymptotic statements actually cover.

  3. Can the entropy bound be corrected and the dependent conclusions rechecked? Correct mathematics, together with functional evidence or a consistently narrower descriptive claim, would support reconsideration of the rating.

2 comments

Thank you very much for your careful and constructive comments. We provide a point-by-point response as follows.

Re: W1 & Q1

  • "Entropy similarity does not ... neither fact explains comparable adaptation performance."

We agree. In the revision, we no longer claim functional invariance. The attention-level result is stated descriptively: attention entanglement entropy is insensitive to the internal differences between adaptation methods. The sentence that this "explains why LoRA performs comparably to FFT" is removed from the Abstract, Section 1 and Section 6 (blue text). We also agree that the normalized entanglement entropy (EE) discards scale and orientation. In our new measurements, the profile of ΔW\Delta W is set by its singular values, whereas the function depends on its singular vectors (see below).

  • "Even identical output Gram matrices can accompany different downstream predictions under fixed downstream weights."

Yes, this is correct, and thank you for pointing it out. The revised Section 5 now states this limitation, and we no longer read the EE of the Gram matrix as evidence about predictions. Table 7 gives a direct example: changing only ΔWV\Delta W_V changes the layer-0 output by 7.8% and the predictions by more than the seed-to-seed KL, while the EE of XX⊤XX^\top stays at 0.559 and 0.558 bits.

  • "In the existing rank-matched transplantation experiment, how do attention/output distances, predictive KL, and held-out loss change alongside entropy?"

We rebuilt the transplanted updates of Appendix K for all 16 layers and measured these quantities together with the EE (Appendix N.2, Table 6). The EE of ΔW\Delta W does not predict function. The update with the LoRA singular values and the dense singular vectors has the LoRA profile but the worst held-out NLL (1.92 instead of 1.57), whereas the update with the dense singular values and the LoRA singular vectors has the dense profile and an NLL of 1.60. Attention and output distances order these updates in the same way as the predictive KL. Updates whose attention EE is within 0.008 bits of LoRA still differ in their predictions by about 1.8 times the seed-to-seed KL, so the attention EE does not capture such small differences in function.

Re: W2, Q2 & Q3

  • "In Lemma E.2/F.2 ... although S=log⁡2S=\log 2."

Thank you very much for raising this point, and we agree. The corrected lemma uses the exact tail mass ε=δ1/(1+δ1)\varepsilon=\delta_1/(1+\delta_1), where δ1=∑i≥2λi/λ1\delta_1=\sum_{i\ge2}\lambda_i/\lambda_1, and gives S(M)≤h2(ε)+εlog⁡(T−1)S(M)\le h_2(\varepsilon)+\varepsilon\log(T-1). For M=I2M=I_2, this gives ε=1/2\varepsilon=1/2 and the bound log⁡2\log 2, so it is tight. The lemma is used in Theorem E.3/F.3 and in the new Corollary F.4. The conclusion S(X)=O(log⁡T/T)S(X)=O(\log T/T) of the theorem still holds, because δ1≤(T−1)η\delta_1\le\sqrt{(T-1)\eta} with η=O(T−3)\eta=O(T^{-3}) gives ε≤δ1=O(T−1)\varepsilon\le\delta_1=O(T^{-1}). The revised PDF includes this correction (blue text).

  • "The attention asymptotic ... log⁡T\log T is extensive in that site count."
  • "What subsystem and size variable ... actually cover."

We agree, and we withdraw the area-law interpretation. The subsystem is the row index of AA (n=log⁡2Tn=\log_2 T of the 2n2n sites), and the size variable is TT. As you note, a log⁡T\log T growth is then extensive in the site count. Theorem E.1/F.1 covers only the central row-column cut, the limit T→∞T\to\infty, and the idealized non-causal isotropic model of Assumption 1. It does not cover the off-center cuts at fixed TT in Figures 5, 19 and 20, or trained causal attention. The revised Abstract, Sections 1, 5 and 6, and Appendix E state this scope (blue text).

For trained causal attention, we now test the theorem directly (Appendix N.4). Assumption 1 does not hold there: at layer 8, s1(A)s_1(A) grows from 6.5 to 14.2 as TT goes from 128 to 1024, instead of tending to 1. The EE at the central cut still grows roughly linearly in log⁡T\log T, by 0.22 to 0.50 bits per doubling of TT, which is 2 to 4 times smaller than the coefficient of the theorem. So the logarithmic growth carries over to trained attention, but its coefficient does not.

  • "Moreover, the output theorem concerns XX, whereas Figure 6 tensorizes XX⊤XX^\top."

We agree. The revised PDF adds Corollary F.4 for the measured object O=XX⊤O=XX^\top. At the central cut, its probabilities are qi∝λi2q_i\propto\lambda_i^2, while those of XX are ∝λi\propto\lambda_i. With η=rstable(O)−1\eta=r_{\rm stable}(O)-1 and εG=η/(1+η)\varepsilon_G=\eta/(1+\eta), it gives S(O)≤h2(εG)+εGlog⁡(T−1)S(O)\le h_2(\varepsilon_G)+\varepsilon_G\log(T-1), so η=O(T−1)\eta=O(T^{-1}) implies S(O)=O(log⁡T/T)S(O)=O(\log T/T). This bounds the central cut only and does not cover the off-center cuts of Figure 6.

In summary, following your Q3, we corrected the bound and rechecked the dependent results, added the functional measurements, and narrowed the claims to descriptive ones. We hope we have addressed your concerns. Thank you very much.

Dear Reviewer 5XeT,

Thank you again for your time and careful comments on our paper. We have uploaded the revised PDF, in which all changes are marked in blue. Following your comments, the revision corrects Lemma E.2/F.2, adds Corollary F.4 for XX⊤XX^\top, withdraws the area-law interpretation and states the scope of Theorem E.1, and reports the functional measurements you asked for (Appendix N.2).

We would be grateful to know whether our response has addressed your concerns. Please feel free to reach out to us if you have any further questions or suggestions. We would be happy to discuss them, and we can still include further changes in the PDF before the revision deadline.

Thank you very much!

Best regards,

Authors

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

The paper studies LoRA and full fine-tuning (FFT) with a tool from quantum physics. It reshapes weight updates (ΔW_Q, ΔW_V) and attention matrices into a Matrix Product State (MPS) and computes entanglement entropy at each cut and analysis the evalaution with comparisoin.

Critical strengths
  • Using MPS and entanglement entropy to compare fine-tuning methods seems new. The authors are also honest that this is only math, not real quantum behavior

  • Very good control experiments in Appendix J–M. I belive this is the strongest part of the paper

  • Lemma I.1 gives a clean, correct reason for where the valley sits. The stable-rank lemma (Lemma F.2) is simple and and easy to follow.

Critical weaknesses
  • The entropy at the central cut k₀ is exactly the entropy of the normalized singular values of ΔW. That is a standard effective rank measure. So LoRA update has a more concentrated spectrum / lower effective rank than FFT. We can see from Appendix K. This is already known in Shuttleworth[1]. The paper needs to show what the full MPS profile tells us beyond a simple SVD of ΔW. Please compare against a plain effective-rank or spectral-entropy baseline.

  • Near the edges of the chain, here the entropy at cut k is bounded by min(k, N−k) bits, and almost any generic matrix reaches that bound. Fig. 12 shows a random Gaussian matrix gives the same shape. So the rising edges of Figs. 2–3 seems coming from the reshaping instead of learning. Correct me if i understand wrong.

  • The attention matrix A = softmax(QKᵀ) does not depend on W_V at all. Since LoRA is applied only to Q and V, half of the update cannot affect A. Can you test Q, k?

  • In Figs. 11 and 20, the curves at step 0 already match later steps. This suggests fine-tuning barely changes A migh be following NTK phenomena, so the “invariance” may only mean that ΔW is small compared to W₀.

  • Assumption 1 comes from isotropic random initialization, inputs with orthonormal rows, and no causal mask. Trained Llama attention with RoPE and a causal mask is very different.

Typo

  • Line 1591: “learning rate to be 2 × 10²”. Probably 2 × 10⁻².
  • Fig 10 caption says the MPS optimal learning rate shifts to larger values, but the text (line 1598) says it shifts smaller as α increases. Please check
  • Line 217: “provided in Sec. C” should be “Appendix C”.
  • Line ~1240: “under a asymptotic model” should be “an asymptotic model”.
  • Line ~1405: “Here we provides” should be “provide”.

[1] LoRA vs Full Fine-tuning: An Illusion of Equivalence

Questions for authors
  • Can you repeat the no-hair test for attention outputs X while changing only ΔW_V? A does not depend on W_V.
  • Is there any setting where attention entropy changes and test loss also changes? This would show the metric is useful.
3 comments

Thank you very much for your constructive comments. We provide a point-by-point response as follows.

Re: W1

  • "So LoRA update has a more concentrated spectrum / lower effective rank than FFT ... This is already known in Shuttleworth et al. 2024."

We agree, and thank you for this pointer. The spectral concentration of LoRA is their finding, and it is not the main finding of our paper (see the last paragraph of this section). The revised Section 2 acknowledges this, and in Section 5 we removed the sentences that presented it as our own proposal (blue text). Our paper adds two controls on this point. We compare LoRA with a dense update trained on the same projection matrices WQW_Q and WVW_V (Appendix J), and with dense updates truncated to the same rank 256 (Appendix K). The concentration persists in both comparisons, so it is not caused by the rank constraint alone.

  • "The paper needs to show ... a plain effective-rank or spectral-entropy baseline."

As you note, the entanglement entropy (EE) at the central cut is exactly the spectral entropy of ΔW\Delta W, and 2S(k0)2^{S(k_0)} is the corresponding effective rank. To test what the profile adds beyond the singular values at the other cuts, we added a simple control (Appendix N.1, Table 5): we keep the learned singular values of ΔW\Delta W and replace its singular vectors with random ones. The profile at the other cuts then changes by at most 0.12 bits (RMSE). For comparison, a random Gaussian matrix of the same shape differs from the LoRA profile by 2.8 to 3.0 bits.

Hence, for ΔW\Delta W, the profile is close to what the singular values alone predict, so the entanglement valley is mainly a property of the spectrum. Section 5 already states this, Appendix N.1 now adds the control above, and the revised Section 2 states that we do not claim the EE is a better measure than the spectral entropy or the effective rank (blue text).

We would like to clarify that our main finding is neither a new measure nor the spectral concentration itself. Our main finding is that the fine-tuning methods differ a lot in their weight updates, but hardly at all in their attention matrices. In ΔW\Delta W, the entanglement valley of LoRA is 3.5 to 4.2 bits deeper than that of a dense update trained on the same Q/V projections (Appendix J). In the attention matrices, the EE of LoRA and that of the dense updates differ by about 0.01 bits or less (Appendices M and N.3), and both stay far below the capacity at every cut (Figure 5).

Re: W2

  • "Near the edges of the chain ... instead of learning."

Your understanding is correct. The rising edges (the boundary regions in our paper) come from the reshaping and are not learned. The EE at cut kk is capped by the capacity min⁡(k,N−k)\min(k,N-k), and both a random Gaussian matrix (Figure 12) and ΔW\Delta W nearly reach this cap near the edges of the chain. The revised Section 5 now states this (blue text).

The edges therefore serve only as a reference. Attention matrices go through the same reshaping, yet their EE stays far below the capacity at every cut (Figure 5), so reaching the capacity is not automatic.

The learned structure is in the central region and in how it changes during training (Figures 2 and 3). The boundary regions stay near the capacity from the first step, while the entanglement valley deepens under LoRA and is lifted under FFT.

Re: W3 & Q1

  • "The attention matrix ... Can you test Q, k?"

Yes. We repeated the no-hair test with two updates trained on WQW_Q and WKW_K instead of WQW_Q and WVW_V: LoRA with rank 256, and a dense update without a rank constraint (Dense-add). Both were trained for 1000 steps with the learning rates of our Q/V runs (Appendix N.2). The conclusion does not change: at the central cut, the EE of ΔW\Delta W is 0.99 bits for LoRA and 6.74 bits for Dense-add, while the EE of the attention matrix is 1.736 and 1.719 bits, and the EE of the attention output operator XX⊤XX^\top is 0.565 and 0.553 bits.

  • "Can you repeat the no-hair test for attention outputs X while changing only ΔWV\Delta W_V?"

Yes. We kept the trained ΔWQ\Delta W_Q of LoRA with rank 256 and used either the trained ΔWV\Delta W_V of this LoRA or that of Dense-add, without retraining (all 16 layers, step 1000, Appendix N.2, Table 7). The EE of the attention output operator XX⊤XX^\top at the central cut is 0.559 and 0.558 bits, respectively.

We also checked the output XX itself. At layer 0, XX changes by 7.8% (relative Frobenius norm) when we swap ΔWV\Delta W_V, although the attention matrix is the same in both cases, and the predictions also change slightly (Table 7). So swapping ΔWV\Delta W_V does change the output, while the EE of XX⊤XX^\top stays the same. Our claim in Section 5 concerns only this EE, and Section 5 now notes that updates with nearly equal EE can still differ in their predictions (blue text).

Re: W4

  • "In Figs. 11 and 20 ... to W0W_0."

Thank you for this point. You are right that the updates are small, and we now quantify this (Appendix N.3, Table 9). In Frobenius norm, ΔWQ\Delta W_Q and ΔWV\Delta W_V are 0.6% to 2.6% of the pretrained weights. Relative to the base model, the attention matrix moves by a row-wise total variation of about 0.08, and its entanglement entropy (EE) at the central cut rises by about 0.12 bits, which is consistent with what you observe in Figs. 11 and 20. The same updates lower the held-out NLL from 2.71 to 1.57. Our claim is therefore not that AA stays at its pretrained value. Our claim concerns the contrast with ΔW\Delta W: updates whose EE differs by several bits all move AA by the same small amount. The three methods agree to within 0.007 bits, well below the 0.12-bit shift from the base model.

We agree that a lazy-training or NTK-like regime may well be the reason. The updates are small, and the attention EE does change when a very large scaling factor makes training fail (Re: Q2). We have not tested this explanation directly, so we do not claim a mechanism and present the result as a descriptive one.

Re: Q2

  • "Is there any setting ... metric is useful."

Yes. Appendix N.3 (Table 10) retrains LoRA with large scaling factors α\alpha and reports the attention EE at the central cut together with the held-out NLL. When training is normal, the attention EE is 1.70 to 1.73 bits and the NLL is 1.56 to 1.57. When training degrades, both change:

α\alphaLearning rateTrainingAttention EE (bits)Held-out NLL
1610−410^{-4}normal1.7311.570
409610−510^{-5}normal1.7241.565
5123×10−43\times10^{-4}degraded1.6191.659
409610−410^{-4}degraded1.4791.762
40963×10−43\times10^{-4}degraded1.0372.900

Re: W5

  • "Assumption 1 comes from ... is very different."

You are right, and we now test this directly (Appendix N.4). Assumption 1 follows the setting of Saada et al. [1], and it does not hold for trained Llama attention. It requires the top singular value s1(A)→1s_1(A)\to 1 and ∥A⊥∥F2=O(1)\|A^\perp\|_F^2=O(1). For the base model at layer 8 with the causal mask, however, s1(A)s_1(A) grows from 6.5 to 14.2 and ∥A⊥∥F2\|A^\perp\|_F^2 from 49 to 267 as TT goes from 128 to 1024. The assumption still fails when we remove the causal mask, and it also fails for the fine-tuned models.

The logarithmic form of Theorem E.1 still carries over to trained attention, but its coefficient does not. In trained attention, the EE at the central cut grows by 0.22 to 0.50 bits per doubling of TT (Table 11), about 2 to 4 times smaller than the coefficient of Theorem E.1. The revised Appendices E and N.4 therefore present Theorems E.1 to E.3 as statements about the idealized model only (blue text).

Re: Typos

Thank you very much for pointing these out. We have corrected all of them in the revised PDF. Two of them need a note:

  • The learning rate is 9×10−49\times10^{-4} (it was misprinted as 2×1022\times10^{2}).
  • The caption of Figure 10 and the text describe two different comparisons. At the same α\alpha, MPS adaptation has a larger optimal learning rate than LoRA, while for both methods the optimal learning rate decreases as α\alpha increases. The revised caption now says "at the same α\alpha" to make this clear.

We hope we have addressed your concerns. Thank you very much.

[1] T. N. Saada, A. Naderi, J. Tanner. Mind the Gap: a Spectral Analysis of Rank Collapse and Signal Propagation in Attention Layers. ICML, 2025.

Dear Reviewer pmnf,

Thank you again for your time and detailed comments on our paper. We have uploaded the revised PDF, in which all changes are marked in blue. Following your comments, the revision credits Shuttleworth et al. (2024) for the spectral concentration of LoRA, states the relation between the entanglement entropy, the spectral entropy and the effective rank, and adds the experiments you asked for in Appendix N (Q/K adaptation, changing only ΔWV\Delta W_V, large scaling factors and longer sequences).

We would be grateful to know whether our responses have addressed your concerns. Please feel free to reach out to us if you have any further questions or suggestions. We would be happy to discuss them, and we can still include further changes in the PDF before the revision deadline.

Thank you very much!

Best regards,

Authors

Reviewer_Lhg7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces “Artificial Entanglement” entropy as a mathematical tool to better understand fine tuning of large language models. The authors take the updates to the query and value projection matrices, as well as the attention matrices, and represent them as Matrix Product States from which they compute the entanglement entropy. They find that entanglement entropy of the query and value matrices follow volume law scaling whereas the attention matrices follow area law scaling. Furthermore, the entropy profiles of the query and value matrices depend on the fine tuning hyperparameters whereas the attention matrices do not. They draw comparisons to the “No-Hair Theorem” for black holes and propose that attention matrices coarse-grain the differences between FFT and LoRA. They also derive an “Attention Cardy Formula” using random matrix theory and show that an MPS adaptation behaves similarly.

Critical strengths

This paper proposes an interesting and potentially useful method for understanding the complex parameter updates that occur during fine tuning.

Critical weaknesses

The authors propose that the attention mechanism coarse-grains the differences between FFT and LoRA, which explains why LoRA performs comparable to FFT. But it’s unclear to me why similar entropy profiles for the attention matrices should imply similar performance. How do we know the entropy is actually capturing something relevant to model performance?

I think the clarity and organization of the paper should be improved. In particular, the paper presents the volume and area scaling laws as important findings, but it is not clearly described how the profiles in Figure 2, 3, and 5 show this. Some figure details could be clearer such as that Figure 4 plots the entropy of the middle cut. Additionally, many important results such as the “Attention Cardy Formula” and MPS adaptation are put in the appendices instead of the main text. The “No-Hair Theorem” could also be better motivated.

Questions for authors

Is there any evidence that the similarity in the entropy profiles of the attention matrices for FFT and LoRA are actually related to model performance?

For the analogy of the “No-Hair Theorem”, is there a deeper mathematical connection here or is it just an example where internal complexities don’t matter for external behavior? If there’s no deeper connection, it seems that discussing black holes adds unneeded complexity that won’t help an ML audience.

2 comments

Thank you very much for your constructive comments. We provide a point-by-point response as follows.

Re: W1 & Q1

  • "Is there any evidence that ... related to model performance?" & "How do we know the entropy ... performance?"

Thank you for raising this important point. We agree that similar entanglement entropy (EE) profiles do not imply similar performance, and we have withdrawn that claim. At the request of Reviewers 5XeT and pmnf, we tested the relation directly (Appendices N.2 and N.3).

The EE of the attention matrices is related to performance only at a coarse level. When training fails at a very large scaling factor α\alpha, the attention EE and the held-out loss both change, and both return to normal when the same α\alpha is trained with a smaller learning rate (Table 10). Among models that train normally, the attention EE agrees to within 0.008 bits, while their predictions still differ slightly (about 1.8 times the seed-to-seed KL). This means that the attention EE cannot resolve such small differences.

The revised PDF therefore no longer claims that attention "coarse-grains" these differences and thereby explains the effectiveness of LoRA (Abstract, Sections 1, 5 and 6, blue text). Our claim is now descriptive: LoRA updates have a much more concentrated spectrum than dense updates trained on the same projection matrices and truncated to the same rank, and the attention EE is insensitive to this difference.

Re: W2

  • "The paper presents the volume and area scaling laws as important findings, but it is not clearly described how the profiles in Figure 2, 3, and 5 show this."

We agree that this part was not clear. Figures 2, 3 and 5 plot the EE at each cut kk of the MPS chain. As a reference, the entanglement capacity min⁡(k,N−k)\min(k,N-k) is the largest EE possible at cut kk. A volume law means that the EE follows this capacity, as it does for a random Gaussian matrix (Figure 12). In Figures 2 and 3, the EE follows the capacity in the boundary regions, and the updates differ from each other mainly in the central region, where the entanglement valley forms. In Figure 5, the EE of the attention matrices stays far below the capacity at every cut. The revised Section 5 now states this (blue text).

We also no longer describe the attention profile as an area law (Section 5 and the caption of Figure 5, blue text), because its EE at the central cut still grows with the sequence length (Appendix N.4).

  • "Some figure details could be clearer such as that Figure 4 plots the entropy of the middle cut."

Thank you. The revised caption of Figure 4 now states that it plots the EE at the middle cut of the MPS chain.

  • "Many important results such as the "Attention Cardy Formula" and MPS adaptation are put in the appendices instead of the main text."

Many thanks for this suggestion. We kept the full statements and proofs in Appendices E to G for two reasons: the page limit, and the fact that the Attention Cardy Formula holds only for an idealized random model. In trained attention, the EE grows with the logarithmic form of the formula but with a smaller coefficient (Appendix N.4). Section 1 now states this scope where the formula is introduced (blue text), and the last paragraph of Section 5 summarizes the MPS adaptation results.

  • "The "No-Hair Theorem" could also be better motivated."

Please see Re: Q2 below.

Re: Q2

  • "For the analogy of the "No-Hair Theorem", is there a deeper mathematical connection here or is it just an example where internal complexities don't matter for external behavior? If there's no deeper connection, it seems that discussing black holes adds unneeded complexity that won't help an ML audience."

It is an analogy, and we do not have a mathematical result that connects the two settings. The no-hair theorem states that a stationary black hole is described from outside by only its mass, charge and angular momentum. Our observation has a similar many-to-one pattern: updates with very different EE of ΔW\Delta W produce nearly the same attention EE. However, our setting has no uniqueness theorem and no conserved quantities. In the idealized model of Theorem E.1, the leading-order EE depends only on one spectral parameter, which is a no-hair-like statement, but we do not claim this connection for trained LLMs.

We agree that the black-hole discussion should not add complexity for an ML audience. The revised Section 1 shortens it to two sentences, which state the analogy and its role as motivation only (blue text).

We hope we have addressed your concerns. Thank you very much.

Dear Reviewer Lhg7,

Thank you again for your time and constructive comments on our paper. We have uploaded the revised PDF, in which all changes are marked in blue. Following your comments, the revision explains how the profiles in Figures 2, 3 and 5 relate to the entanglement capacity, reports how the entanglement entropy relates to the held-out loss (Appendices N.2 and N.3), and shortens the No-Hair discussion in Section 1.

We would be grateful to know whether our response has addressed your concerns. Please feel free to reach out to us if you have any further questions or suggestions. We would be happy to discuss them, and we can still include further changes in the PDF before the revision deadline.

Thank you very much!

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Our Responses and Revisions

Dear ACs and Reviewers,

Thank you very much for your time and constructive comments. We have replied to each review and uploaded a revised PDF, in which all changes are marked in blue.

Strengths noted by the reviewers

  • Reviewer pmnf found the use of Matrix Product States and entanglement entropy (EE) to compare fine-tuning methods new, and Reviewer Lhg7 found it an interesting and potentially useful method for understanding the parameter updates during fine-tuning.
  • The control experiments in Appendices J to M are very good (Reviewer pmnf), and those in Appendices J and K separate the spectral concentration of LoRA from its rank constraint (Reviewer 5XeT).
  • The robustness checks in Appendices L and M strengthen the descriptive findings (Reviewer 5XeT).
  • Lemma I.1 gives a clean and correct reason for the position of the entanglement valley (Reviewer pmnf).

Main concerns and our responses

1. Relation between the EE and model performance (Reviewers Lhg7, 5XeT and pmnf). We agree that similar EE profiles do not imply similar performance. We have withdrawn the claim that the attention-level result explains why LoRA performs comparably to full fine-tuning, and the revised Abstract and Sections 1, 5 and 6 state it descriptively. Our new functional measurements (Appendices N.2 and N.3) show the following.

  • The EE of ΔW\Delta W does not predict function. A transplanted update with the singular values of LoRA and the singular vectors of the dense update has the LoRA profile, but its held-out negative log-likelihood (NLL) is the worst, at 1.92 compared with 1.57 for LoRA.
  • Updates whose attention EE is within 0.008 bits of LoRA still differ in their predictions, so the attention EE does not capture such small differences in function.
  • When training degrades at a very large scaling factor α\alpha, the attention EE and the held-out NLL change together, and both return to normal at the same α\alpha with a smaller learning rate. The attention EE is therefore related to performance only at a coarse level.

2. Relation to spectral concentration and effective rank (Reviewer pmnf). The revised Section 2 credits Shuttleworth et al. (2024) for the spectral concentration of LoRA. It also states that the EE at the central cut equals the spectral entropy of ΔW\Delta W, and that we do not claim the EE is a better measure than the spectral entropy or the effective rank. A new control (Appendix N.1) keeps the singular values of ΔW\Delta W and replaces its singular vectors with random ones. The profile at the other cuts then changes by at most 0.12 bits, so the entanglement valley is mainly a property of the spectrum. Our main finding is the contrast between weight updates and attention matrices: the entanglement valley of LoRA is 3.5 to 4.2 bits deeper than that of a dense update trained on the same Q/V projections, while their attention EE differs by about 0.01 bits or less.

3. Theory (Reviewers 5XeT and pmnf).

  • We corrected Lemma E.2/F.2 with the exact tail mass. The corrected bound is tight for M=I2M=I_2, and the conclusion of Theorem E.3/F.3 still holds.
  • We added Corollary F.4 for the measured object XX⊤XX^\top.
  • We withdrew the area-law interpretation and now state the scope of Theorem E.1: the central cut, the limit T→∞T\to\infty, and the idealized model of Assumption 1.
  • In trained causal attention, Assumption 1 does not hold (Appendix N.4). The EE at the central cut still grows logarithmically in TT, but with a coefficient 2 to 4 times smaller than in Theorem E.1.

4. Additional experiments (Reviewer pmnf, Appendix N).

  • Adapting WQW_Q and WKW_K does not change the conclusion. At the central cut, the EE of ΔW\Delta W is 0.99 bits for LoRA with rank 256 and 6.74 bits for a dense update, while the attention EE is 1.736 and 1.719 bits.
  • When only ΔWV\Delta W_V is changed, the EE of XX⊤XX^\top stays at 0.559 and 0.558 bits, although the output XX at layer 0 changes by 7.8%.
  • ΔWQ\Delta W_Q and ΔWV\Delta W_V are only 0.6% to 2.6% of the pretrained weights, yet they lower the held-out NLL from 2.71 to 1.57 and raise the attention EE by about 0.12 bits. The three methods agree to within 0.007 bits.

5. Clarity (Reviewers Lhg7 and pmnf). Section 5 now explains Figures 2, 3 and 5 with the entanglement capacity min⁡(k,N−k)\min(k,N-k), the largest EE possible at cut kk. The EE of ΔW\Delta W follows this capacity in the boundary regions, as a random Gaussian matrix does, and the updates differ mainly in the central region, where the entanglement valley forms. The attention EE stays far below the capacity at every cut. We also clarified the caption of Figure 4, stated the scope of the Attention Cardy Formula in Section 1, shortened the No-Hair discussion to two sentences, and corrected the typos.

We hope we have addressed your concerns, and we would be happy to discuss any further questions or suggestions. Thank you very much!

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.