Inverse Depth Scaling From Most Layers Being Similar Yizhou Liu, Sara Kangaslahti, Ziming Liu, Jeff Gore Abstract Neural scaling laws relate loss to model size in large language models (LLMs), yet depth and width may contribute to performance differently, requiring more detailed studies. Here, we quantify how depth affects loss via analysis of LLMs and toy residual networks. We find loss scales inversely proportional to depth in LLMs, probably due to functionally similar layers reducing error through ensemble averaging rather than compositional learning or discretizing smooth dynamics. This regime is inefficient yet robust and may arise from the architectural bias of residual networks and target functions incompatible with smooth dynamics. The findings suggest that improving LLM efficiency may require architectural innovations to encourage compositional use of depth. 1 1 1 Code available at github.com/liuyz0/DepthScaling 1 Introduction Figure 1 : Three regimes of how LLMs may utilize their layers. Neural scaling laws played an important role in the success of today’s large language models (LLMs), revealing how model performance predictably improves with increases in model size and dataset size. Quantitatively, loss is found to have power-law relationships with these factors ( 17 ; 22 ; 18 ) . Yet, the origin of these power laws remains largely unclear. Studies on neural scaling laws have primarily focused on total parameter count, treating model size as a monolithic quantity ( 40 ; 7 ; 20 ; 31 ; 38 ; 33 ; 29 ; 6 ) . In practice, there is another line of research studying the optimal relationship between model width and depth ( 35 ; 25 ) , indicating their different contributions to loss. Width heuristically limits the representation capacity, while depth restricts the quality of transformation. Recently, theoretical works on neural scaling laws also suggested separating the scaling with model size into one with width and another with depth ( 27 ; 8 ) . In this work, we try to dive into the scaling with depth, asking: Q: Neural depth scaling? How do LLMs use their depth, and what is the quantitative law between depth and performance? There have been many empirical investigations on how LLMs utilize their layers ( 23 ; 26 ; 15 ; 37 ; 32 ; 41 ; 13 ; 16 ; 19 ) . Despite the differences in approaches and analysis, these works conclude that LLMs use their depth inefficiently: a lot of layers are redundant or make incremental updates on hidden states. However, while these studies provide compelling evidence of depth inefficiency, they lack a quantitative framework for understanding why this inefficiency exists and what scaling relationship depth follows. Different theories about how depth can contribute to neural network performance have been proposed. One of the fundamental intuitions for multiple layers in neural networks is that depth allows learning hierarchical abstractions through composition ( 3 ; 2 ; 24 ) . For LLMs, this would correspond to layers building increasingly abstract representations, with early layers handling syntax, middle layers capturing semantics, and deeper layers performing complex reasoning ( 25 ; 10 ; 13 ) . The loss scaling in this regime may then heavily depend on the data properties. We will call this perspective on depth usage as compositional assembly ( Figure 1 ). A second perspective views deep residual networks as discrete approximations of smooth dynamical systems, i.e., neural ODEs ( 11 ; 36 ; 12 ) , which we refer to as procedural assembly . In this regime, layers perform incremental refinement along a smooth path, and the loss arising from discrete approximations is often a power law with depth or number of integration “time” grid points. Finally, a third possibility is that layers function as an ensemble of similar shallow subnetworks ( 42 ; 30 ) , each contributing to the outcome with the same mean but a different error that can be partly canceled through averaging. In this regime of depth utilization, which we call ensemble averaging , the loss will be a power law with depth due to the central limit theorem ( 1 ; 39 ) . Although there has been empirical understanding of how LLMs utilize their depth and theoretical expectations on loss scaling with depth, the connection between LLM reality and theory has not yet been established. To have a quantitative framework of the actual depth scaling of LLMs, we evaluate the behaviors of hidden states across layers in LLMs and compare with theory ( Section 2 ). We find from hidden states that most layers in LLMs are not in the compositional assembly regime. Combining such insights with previous works, we then propose a power-law loss decomposition and find that the depth-limited part is roughly inversely proportional to depth. To understand the inverse depth scaling, we design toy model training experiments ( Section 3 ) to study both the loss and hidden states in the procedural assembly and the ensemble averaging regimes ( Section 4 ). After comparing our controlled experiments with LLM evaluations, we conclude that LLMs utilize most layers similarly to the ensemble averaging regime. We compare our findings to previous works in Section 5 , and discuss limitations and implications in Section 6 . Our contributions are the following: \bullet We find the inverse depth scaling of loss in LLMs. \bullet We find a mechanistic reason for this scaling behavior: similar layers form an ensemble, reducing error by averaging. To summarize, our work provides a more quantitative explanation for how LLMs utilize their layers and pushes neural scaling laws to a more detailed level. Our findings have direct implications for the optimal depth-width relationship and future directions of architectural innovations. 2 LLM Experiments Figure 2 : Most data in most layers are processed in an even and incremental way. (a) Updates of hidden states measured by the angle between neighboring hidden states \theta(h_{l},h_{l+1}) show incremental changes in most middle layers. (b) By PCA of \theta(h_{l},h_{l+1}) , most tokens are updated evenly in the middle, a small fraction stops updating early, which usually corresponds to the first tokens in documents. (c) Mean update decreases with depth. (d) The mean updates scales approximately inversely proportional to depth, suggesting more fine-grained updates rather than decomposing higher-level information. (e) Correlation between neighboring updates is small, suggesting non-smooth dynamics. (f) LLM loss roughly follows an inverse depth scaling. More details about hidden state analysis are in Appendix A , and those of loss fitting are in Appendix B . To connect theoretical expectations with the behavior of real LLMs, we design experiments to probe how hidden states evolve across layers. Among the three regimes, compositional assembly should exhibit qualitatively distinct signatures. If different layers encode progressively higher levels of abstraction, simpler inputs should require fewer layers: some hidden states may converge or stop changing early, while others continue to evolve deeper into the network, depending on the input complexity. In contrast, procedural assembly and ensemble averaging are unlikely to show sharp depth-wise stopping. Procedural assembly approximates smooth dynamics, while ensemble averaging implies many layers perform similar transformations. Under compositional assembly, networks of different depths should share similar early-layer representations at matched layer indices, since these layers encode the same lower-level features; deeper networks would then add new layers that capture higher-level structures inaccessible to shallower models. By contrast, in procedural assembly or ensemble averaging, the per-layer update magnitude should systematically decrease as depth increases, making more fine-grained and accurate estimations, rather than introducing qualitatively new representations. Our experiments then focus on the hidden state updates across layers for each input. We start by studying the angular change of hidden states across layers for individual data points. Due to the wide use of layer norms in LLMs ( 9 ; 5 ; 44 ; 43 ; 21 ) , the angular updates of hidden states capture the essential dynamics. Given an input token x at a certain position in the document, the embedding matrix first turns the token into a vector h_{0} , Transformer layer l then update h_{l-1} (at the same position along context length direction) into h_{l} for l=1,2,...\ell where we use \ell to represent the depth or the number of layers, and at the end, h_{\ell} will go through the final layer norm and the language model head to generate logits for next-token prediction. In each Transformer layer, e.g., layer l , information about previous tokens can be grabbed into h_{l} . Yet, we still refer to h_{l} as the hidden state of the input token x after layer l . We calculate the angle between neighboring hidden states, denoted as \theta(h_{l},h_{l+1}) for analysis (details in Appendix A ). If the hidden states are following smooth dynamics in the depth direction, \theta(h_{l},h_{l+1}) is related to the first-order derivative. We show the angles between hidden states, \theta(h_{l},h_{l+1}) , measured from Pythia-410m ( 5 ) and evaluated on the FineWeb dataset ( 34 ) in Figure 2 a. Each gray line represents \theta(h_{l},h_{l+1}) as a function of l for a single input token, while the dark line shows the dataset-averaged angle, denoted as \langle\theta(h_{l},h_{l+1})\rangle_{\mathcal{D}} . We observe that only a few inputs exhibit early stopping of hidden-state updates, reminiscent of compositional assembly. The first and last layers rotate hidden states by large angles close to \pi/2 , in contrast to intermediate layers. Most middle layers update hidden states by similarly small angles for the majority of inputs, resembling procedural assembly and ensemble averaging. The norms of hidden states convey similar messages: the first and last layer changes norm a lot, and the middle layers do not update the norms much ( Appendix A ). Overall, these observations suggest that no single theoretical regime fully explains LLM behavior; instead, one regime may dominate while others coexist. We proceed to quantify the ratios of different behaviors in Pythia-410m. A trajectory \theta(h_{l},h_{l+1}) with l from 0 to \ell-1 can be thought of as an \ell -dimensional vector. And there are |\mathcal{D}| (the number of evaluated tokens) such vectors. We perform principal component analysis (PCA) of these vectors ( Appendix A ) and find that there are two clusters ( Figure 2 b). To interpret the two clusters, we construct two ideal trajectories. One trajectory is called “early stop”, whose angles in the middle are zero, and another one is “evenly in the middle” with angles in the middle being 0.45 rad. The “early stop” trajectory is projected into the subspace of the first two components, denoted by the star, and the “evenly in the middle” one is the cross ( Figure 2 b). We can then conclude that the small cluster, which accounts for 0.4 % of the tokens, roughly contains data not processed in the middle, and the big cluster, which occupies the remaining 99.6 %, contains data that is approximately evenly updated in the middle layers. We check some of the tokens whose trajectories are in the small cluster and find that they are the first tokens of the documents. These behaviors are generally true for other model sizes and model families ( Appendix A ). Since we study the loss averaged over all data, and how it scales with depth, we can focus on the majority of tokens – those being “evenly updated in the middle”. We average the angle updates across tokens to identify the dominant depth-wise behavior. The dataset-averaged angles \langle\theta(h_{l-1},h_{l})\rangle_{\mathcal{D}} , for l=1,2,\dots,\ell , are plotted against the normalized layer index l/\ell in Figure 2 c for different Pythia models. We find that the first and last layers exhibit similar large update angles across model depths, whereas the intermediate layers show systematically decreasing angles as depth increases. This suggests that the first and last layers may implement depth-independent functions, reminiscent of compositional assembly. However, since our focus is on depth scaling, the dominant contribution comes from the bulk of the layers. For these middle layers, increasing depth corresponds to finer-grained or more accurate incremental updates, consistent with procedural assembly or ensemble averaging. We average \langle\theta(h_{l-1},h_{l})\rangle_{\mathcal{D}} over the middle layers l=2,3,\dots,\ell-1 , denoted as \langle\theta(h_{l},h_{l+1})\rangle_{\mathcal{D},l} , and find that this quantity scales approximately inversely with depth ( Figure 2 d). This scaling further supports the procedural assembly or ensemble averaging interpretation. We therefore conclude the following Result 1: Not compositional assembly For the majority of data and the majority of layers, LLM depth behavior may be best described by procedural assembly or ensemble averaging. To distinguish procedural assembly from ensemble averaging in the middle layers, we evaluate the correlation of incremental hidden-state updates across layers. We denote the update produced by Transformer layer l as \Delta h_{l}=h_{l}-h_{l-1} , and quantify inter-layer correlation by the angle between neighboring updates, \theta(\Delta h_{l},\Delta h_{l+1}) . We find that correlations are generally weak: \theta(\Delta h_{l},\Delta h_{l+1}) is typically large, which is inconsistent with smooth dynamics ( Figure 2 e). However, this conclusion is not definitive, as we lack a calibrated reference for comparison. Controlled training experiments with known data-generating processes are required to disentangle procedural assembly and ensemble averaging (later Sections), after which detailed signatures can be compared to LLM measurements to draw a definitive conclusion. Figure 3 : The toy model can exhibit inverse depth scaling when the underlying transformation dynamics to learn are smooth, and the target distribution is sharp, or when the underlying dynamics to learn are noisy. (a) Architecture of the toy model. (b) Tied teacher weights ( \rho=1 ) produce smooth dynamics and yield \alpha_{\ell}=1 at low teacher temperature (peaked output distributions). Independent teacher weights ( \rho=0 ) generate non-smooth dynamics and have \alpha_{\ell}=1 across different temperatures. Error bars are standard errors. Details in Appendix C . In addition to hidden-state dynamics, we can directly examine loss scaling in LLMs. Neural scaling laws are largely empirical and are typically described by the following form in previous works ( 18 ) : L=\frac{c_{N}}{N^{\alpha_{N}}}+\frac{c_{D}}{D^{\alpha_{D}}}+L_{0}, (1) where N is the number of parameters, D is the training dataset size, and L_{0} is the irreducible loss associated with the entropy of the target distribution. Here, following our findings and prior theoretical arguments ( 25 ; 27 ; 8 ) , we suggest that the model-size term should be decomposed into width- and depth-dependent contributions: \frac{c_{N}}{N^{\alpha_{N}}}=\frac{c_{m}}{m^{\alpha_{m}}}+L_{\ell}+\cdots, (2) where m denotes the model width, which leads to a power-law loss term under multiple theories ( 1 ; 27 ; 8 ) , L_{\ell} is a term that depends only on depth, and the ellipsis denotes additional cross terms involving both width and depth. Conceptually, the width-dependent power law captures representation-limited error that cannot be eliminated even with perfect learning of transformations, while L_{\ell} captures transformation-limited error that cannot be eliminated by perfect feature representation. Cross terms arise from the interplay between imperfect representation and imperfect transformation, and are expected to be higher order and subdominant when both m and \ell are large. We therefore focus on the dominant contributions. Some theories predict that L_{\ell} decays exponentially with depth, and that the optimal width-depth trade-off under a fixed parameter budget N\propto m^{2}\ell implies depth scaling logarithmically with width ( 25 ) . Other theories instead predict a power-law dependence on depth. We argue that if, near the optimal width-depth relation, the loss exhibits a power-law dependence on N without logarithmic corrections, then L_{\ell} must itself scale as a power law in \ell . Motivated by prior experiments and theory, we propose the following decomposed scaling form: L=\frac{c_{m}}{m^{\alpha_{m}}}+\frac{c_{\ell}}{\ell^{\alpha_{\ell}}}+\frac{c_{D}}{D^{\alpha_{D}}}+L_{0}. (3) We next fit real LLM data using this decomposed scaling form. The model contains seven free parameters, but most open-source model families do not provide sufficient coverage in both width and depth to reliably constrain all terms. Moreover, differences across model families introduce confounding architectural and training variations, complicating joint fitting. We therefore use around 200 data points reconstructed from the Chinchilla models ( 18 ; 4 ) for our fitting. We fit the model by minimizing the mean squared error of the logarithm of the loss and obtain \alpha_{m}=0.98\pm 0.08 , \mathbf{\alpha_{\ell}=1.2\pm 0.3} , and \alpha_{D}=0.30\pm 0.01 (see Appendix B ). The relative error between the empirical loss and the fitted model prediction is 0.4\% on average, indicating a good fit and providing partial support for the proposed decomposition. The depth-dependent loss component, defined as the residual after subtracting the fitted irreducible, width-dependent, and dataset-size-dependent terms, is shown in Figure 2 f and exhibits approximately power-law dependence on depth. The fitted width exponent \alpha_{m}\approx 1 is consistent with a recent observation ( 27 ) , and \alpha_{D}=0.30 closely matches the original Chinchilla scaling exponent ( 18 ) . We therefore conclude that Result 2: Empirical inverse depth scaling Equation 3 provides a reasonable description of LLM loss, with the depth-dominant contribution scaling approximately inversely with depth. Both the procedural assembly and the ensemble averaging regimes can lead to power-law scaling. A more detailed understanding is needed to distinguish them. 3 Toy Model Figure 4 : Matching smooth dynamics leads to the inverse depth scaling when not trained well. (a) The depth scaling exponent \alpha_{\ell} is near 1 in the early stage of training but increases to 3 after training. Error bars are standard errors. (b) Low temperature makes training slow, yet all \alpha_{\ell} tend to increase with more training. Maximum number of training steps t_{\rm max}=80000 . (c) Loss versus training steps for \ell=6 from experiments in Figure 3 suggest that low teacher temperature requires longer training, and the corresponding student has not yet converged. Details in Appendix C . To study the procedural assembly and ensemble averaging regimes in isolation, we design controlled experiments on toy models and complement them with theoretical analysis. Theory alone cannot determine which mechanisms are relevant in realistic training, while experiments on full-scale LLMs are prohibitively expensive and confounded by many interacting factors. Toy models then provide a controlled setting in which specific hypotheses can be tested directly. Our toy model captures a key architectural ingredient of Transformers: residual connections. We adopt a teacher-student setup, in which the teacher generates training data for the student. The teacher and student share the same architecture, except that the teacher has greater depth. Inputs x\in\mathbb{R}^{m} are sampled i.i.d. from the standard normal distribution. For concreteness, we describe the student architecture. The input x is first normalized using root mean square norm to obtain the initial hidden state h_{0} ( Figure 3 a), h_{0}=\mathrm{RMSNorm}(x). (4) The hidden state then propagates through \ell residual blocks, h_{l}=h_{l-1}+\mathrm{MLP}_{l}(h_{l-1}),\quad l=1,\dots,\ell, (5) where \ell is the student depth. Each residual block consists of a two-layer MLP, \mathrm{MLP}_{l}(v)=B_{l}\mathrm{ReLU}^{2}(A_{l}\mathrm{RMSNorm}(v)+b_{l}), (6) with parameters B_{l}\in\mathbb{R}^{m\times 4m} , A_{l}\in\mathbb{R}^{4m\times m} , and b_{l}\in\mathbb{R}^{4m} . The logits are obtained by projecting the final hidden state, y=W\,\mathrm{RMSNorm}(h_{\ell})\in\mathbb{R}^{n}, (7) where W\in\mathbb{R}^{n\times m} . The teacher follows the same architecture but has depth \ell^{*}>\ell , producing logits y^{*} . We use an extra * to denote the variables in the teacher. From these logits, we define output distributions and train the student by minimizing the KL-divergence 2 2 2 KL-divergence is cross-entropy subtracting the entropy of target distributions, which is constant, and has no difference in scaling behaviors. between the student and teacher distributions. We adopt these architectural designs from recent practices of LLMs ( 21 ) . The toy model is a highly simplified abstraction of LLMs, but it is sufficient for isolating depth-scaling phenomena. We ignore embedding training and focus on transformation error, interpreting the toy-model input x as the output of an embedding layer in LLMs. By using the same width m for the teacher and student, we eliminate representation-limited error by construction, so that any residual error is transformation-limited. We also omit attention, which in real LLMs induces spatially coupled dynamics that are better described by PDEs rather than ODEs. Nevertheless, depth scaling alone should remain the same for “PDEs” and “ODEs”. Other architectural components are kept as close as possible to standard LLM designs. We tune teacher and data properties to induce different depth-scaling regimes. We fix m=32 , n=128 , teacher depth \ell^{*}=128 , and vary the student depth \ell from 6 to 48 . Teacher MLP weights are initialized with standard schemes and rescaled by 1/\sqrt{\ell} to ensure that the cumulative transformation from h_{0}^{*} to h_{\ell^{*}}^{*} remains O(1) . Among the controllable data properties, we find that the teacher “temperature” strongly affects training dynamics. After obtaining teacher logits y^{*} , we divide them by a temperature parameter before applying softmax to obtain the target distribution. Lower temperature increases the logit norm and produces sharper, more peaked target distributions. Another key control is whether teacher MLP weights are tied across layers, i.e., A_{l}^{*} are identical, and B_{l}^{*} are identical for all l . Weight tying induces smooth hidden-state dynamics in the limit \ell^{*}\to\infty , with derivatives of arbitrary order, and is expected to produce procedural assembly. In contrast, sampling teacher MLP weights i.i.d. across layers yields hidden-state dynamics resembling a random walk, which is non-smooth and likely leads to ensemble averaging for students. Students are trained for 40000 steps using Adam ( Appendix A ). We fit the loss using L=\frac{c_{\ell}}{\ell^{\alpha_{\ell}}}+L_{\backslash\ell}, (8) where L_{\backslash\ell} denotes depth-independent contributions. For independent teacher weights (weight correlation \rho=0 in Figure 3 b), we observe inverse-depth scaling with \alpha_{\ell}\approx 1 . For tied teacher weights ( \rho=1 in Figure 3 b), the fitted depth exponent \alpha_{\ell} increases from 1 to 3 as teacher temperature increases. These results indicate that both procedural assembly and ensemble averaging may yield inverse-depth scaling in certain regimes. We next analyze the mathematical origins of these scaling behaviors. 4 Toy Model Analysis Figure 5 : Matching the full transformation as an ensemble may explain the inverse depth scaling in LLMs. (a) Hidden state updates of a student whose teacher has independent MLP weights are even across layers. Each gray line is from one input, and the dark line is averaged on dataset. (b) Mean update in one layer scales inversely proportional to depth. (c) Correlation between neighboring updates is small, suggesting no smooth dynamics. These hidden state features agree with ensemble averaging and are similar to those in LLMs. Details in Appendix C . We first analyze the case where teacher weights are tied. In this setting, the underlying transformation can be modeled as a smooth continuous dynamics, and the student may implement procedural assembly, which we aim to test explicitly. The normalized layer index s=l/\ell plays the role of “time” in a neural ODE, with step size \Delta s=1/\ell . The residual dynamics for a given input can be written as h(s+\Delta s)=h(s)+f(s)\Delta s. (9) Assuming the teacher hidden states are smooth in s (with \ell^{*}\gg\ell ), we study the deviation between student and teacher trajectories: \displaystyle h\Big(\frac{l+1}{\ell}\Big) \displaystyle-h^{*}\Big(\frac{l+1}{\ell}\Big)=h\Big(\frac{l}{\ell}\Big)-h^{*}\Big(\frac{l}{\ell}\Big) \displaystyle\quad+\frac{1}{\ell}\left(f\Big(\frac{l}{\ell}\Big)-f^{\circ}\Big(\frac{l}{\ell}\Big)\right) \displaystyle\quad+\int_{l/\ell}^{(l+1)/\ell}\left(f^{\circ}\Big(\frac{l}{\ell}\Big)-f^{*}(s)\right)\mathrm{d}s. (10) Here, f denotes the student transformation, f^{\circ} is the optimal student transformation attainable by adjusting weights, and f^{*} is the true teacher transformation. The second term in Equation 10 captures the error due to imperfect training to match the local transformation, while the third term corresponds to the discretization error. We next analyze the scaling behavior of these contributions. If the error in matching the local transformation dominates, each student layer contributes an O(1/\ell) error as indicated in Equation 10 . With \ell layers, naive accumulation yields \|h(1)-h^{*}(1)\|=O(1) . However, if errors across layers are independent, the central limit theorem implies diffusive accumulation, giving a typical behavior \|h(1)-h^{*}(1)\|\stackrel{{\scriptstyle\rm t}}{{=}}O(1/\sqrt{\ell}). (11) When h(1) is close to the optimal trajectory h^{*}(1) , we can Taylor expand the analytic loss around h^{*}(1) where the gradient is zero, yielding that the leading loss contribution is quadratic in \|h(1)-h^{*}(1)\| . Consequently, the loss can range from an O(1) upper bound in the worst case to a typical scaling L\sim 1/\ell . Once training converges, the loss is dominated by discretization error. The residual stream corresponds to a first-order integration scheme, and for smooth f^{*} we have f^{\circ}(l/\ell)-f^{*}(s)=O(\Delta s) for s\in[l/\ell,(l+1)/\ell] . Consequently, the error contributed by a single student layer is O(\Delta s^{2}) , i.e., O(1/\ell^{2}) . In the worst case, errors accumulate coherently, yielding \|h(1)-h^{*}(1)\|=O(1/\ell) and hence L\sim 1/\ell^{2} . Under independent layer-wise discretization errors, the typical behavior is \|h(1)-h^{*}(1)\|\stackrel{{\scriptstyle\rm t}}{{=}}O(1/\ell^{3/2}), which implies L\sim 1/\ell^{3} . We test these theoretical predictions using experiments at a temperature 1 , where training converges reliably ( Appendix C ). At early training times t , the fitted depth-scaling exponent is close to 1 ( Figure 4 a), consistent with the training-dominated behavior. During intermediate training, additional dynamics obscure a clean power-law fit in depth. At late training, however, the exponent approaches 3 , in agreement with the predicted typical discretization-dominated scaling. We modify the student architecture to mimic a higher-order integral scheme, which increases \alpha_{\ell} as expected under tied teacher weights ( Appendix C ). These results confirm that tied teacher weights induce procedural assembly (discretization of smooth dynamics) in the student. To understand why the scaling exponent reduces to 1 at low teacher temperature, we analyze \alpha_{\ell} across temperatures and training steps ( Figure 4 b). We find that \alpha_{\ell} increases with t for all temperatures, but grows significantly more slowly at low temperatures. At high temperatures, the loss saturates by the end of training, whereas at low temperatures it continues to decrease, indicating incomplete optimization ( Figure 4 c). We thus attribute the observed \alpha_{\ell}\approx 1 regime to imperfect training: low temperature can map a wide range of hidden states to similar peaked output distributions, filtering information and reducing gradient signal magnitudes ( 28 ) , thereby slowing convergence. Since our primary focus is on depth scaling after training convergence, rather than cross terms involving training dynamics and depth ( Equation 3 ), we argue that procedural assembly is unlikely to be the dominant mechanism in LLM training, as it predicts a converged depth exponent \alpha_{\ell}=3 . We next study the case of independent teacher weights. In this setting, f^{*}(s) is non-smooth and its derivative does not exist. As a result, the discretization term in Equation 10 (third row) is O(\sqrt{\Delta s}) if f^{\circ} attempts to match the local value f^{*}(l/\ell) . Even under typical error accumulation, this yields an O(1) deviation in the final hidden state and a non-vanishing loss. Independent teacher weights therefore cannot yield procedural assembly. For general transformations that cannot be modeled as smooth dynamics, the network cannot decompose the target function into local incremental updates and may instead follow ensemble averaging: \displaystyle h(1)-h^{*}(1)=\frac{1}{\ell}\sum_{l=0}^{\ell-1}\left(f(l/\ell)-\int_{0}^{1}f^{*}(s)\mathrm{d}s\right). (12) In this picture, each layer (or subnetwork) attempts to approximate the full transformation with some error. Each layer contributes an O(1/\ell) error, and summing over \ell layers yields an upper bound \|h(1)-h^{*}(1)\|=O(1) , with a typical behavior \|h(1)-h^{*}(1)\|\stackrel{{\scriptstyle\rm t}}{{=}}O(1/\sqrt{\ell}) under independent layer-wise errors. Consequently, the typical loss scaling is L\sim 1/\ell , i.e., \alpha_{\ell}\geq 0 with \alpha_{\ell}\stackrel{{\scriptstyle\rm t}}{{=}}1 . Consistent with this prediction, we find that with independent teacher weights the measured \alpha_{\ell} is robustly close to 1 ( Figure 3 b). Examining hidden states of trained students, we observe that layer-wise updates, quantified by \theta(h_{l},h_{l+1}) , are approximately uniform across depth ( Figure 5 a). The averaged update magnitude \langle\theta(h_{l},h_{l+1})\rangle_{\mathcal{D},l} scales inversely with the number of layers ( Figure 5 b). Moreover, correlations between neighboring updates, measured by \theta(\Delta h_{l},\Delta h_{l+1}) , are large and close to \pi/2 , rejecting the hypothesis of smooth dynamics ( Figure 5 c). Together, these hidden-state signatures and loss-scaling results indicate that independent teacher weights induce ensemble averaging. The hidden state behaviors in Figure 5 are qualitatively similar to those of most data in most layers in LLMs ( Figure 2 ). The quantitative prediction of inverse depth scaling from ensemble averaging also agrees with the LLM scaling result. Combining all the evidence, we conclude that Result 3: Depth scaling from ensemble averaging Inverse depth scaling in LLMs is likely due to ensemble averaging. 5 Related Works Building on empirical studies of neural scaling laws ( 17 ; 22 ; 18 ) , our work suggests a finer decomposition of scaling behavior into separate depth- and width-dependent contributions. Compared to recent theoretical works that propose such separations ( 27 ; 8 ) , we provide a direct measurement on LLMs. Our conclusion that LLMs predominantly operate in the ensemble averaging regime is consistent with prior findings that Transformer layers are highly redundant or inefficient ( 15 ; 37 ; 41 ; 32 ) , fail to exploit compositional structures in data ( 13 ) , and are robust to layer deletion or permutation ( 23 ) . We extend these observations by providing a quantitative description, i.e., the inverse depth loss scaling. Previous theoretical analyses of residual networks ( 12 ; 36 ) primarily focus on worst-case error bounds. In contrast, our theory–experiment framework shows that the observed loss follows the typical regime, where layer-wise errors accumulate diffusively and are effectively independent. 6 Discussion We find that loss scales roughly inversely proportional to depth in LLMs, and provide a mechanistic reason by a combination of experiments and theory: the majority of layers in LLMs make similarly incremental updates when facing most of the data and reduce error by averaging. Our work is limited in several aspects. Loss decomposition Equation 3 was not derived from first principles but proposed based on a combination of previous empirical findings and theoretical insights, but is supported by the quality of fitting. The study of the interplay between width, depth, and training dataset size can be important, especially for small models. We evaluated the three popular theories of layer utilization and concluded that ensemble averaging is more relevant via the hidden state behaviors and quantitative depth scaling. Due to the lack of first-principle derivations, we cannot rigorously exclude the possibility that there are other mechanisms that can produce the same inverse depth scaling and similar hidden state behaviors. Within ensemble averaging, there can be more general situations that our study cannot distinguish. The middle layers may not be in one ensemble for one function group, but a fixed number of ensembles or function groups. Increasing the number of total layers may make each ensemble grow proportionally and still yield an inverse depth scaling. We find the inverse depth scaling in LLMs and explain it by ensemble averaging, yet a more mechanistic explanation for the emergence of ensemble averaging is needed. The architectural bias of the residual connection may play an important role in not utilizing data compositionality ( 42 ) . Next-token prediction may not be modeled by a smooth dynamical system, which forces the LLMs to use ensemble averaging. The findings that fitted \alpha_{m}\approx 1 and \alpha_{\ell}\approx 1 imply that, under a fixed number of parameters N\propto m^{2}\ell , the optimal width-depth relationship is m\propto\ell . The result suggests that we should scale width and depth in a proportional manner, which is not rigorously true in practice ( Appendix B ), yet not far away. Beyond current architecture, if we want to make LLMs more efficient in depth scaling, one direction may be encouraging compositional use of depth. The use of recurrent depth seems to better explore data hierarchy ( 14 ) . We hope our insights can help the continued development of LLMs, from hyperparameter choosing to architecture design. Appendix A LLM Hidden States A.1 Methods Our script is designed to probe how internal representations evolve across depth in a pretrained causal language model. The high-level goal is to take real text, run it through the model, and then compute layer-by-layer geometric and predictive diagnostics of the hidden states, such as norms, rotation angles, and the cross-entropy that would be obtained if we “stopped” at an intermediate layer and applied the model’s final normalization and output head. The code runs entirely in inference mode (no training), and it accumulates per-token measurements over a large sample of documents before saving the resulting tensors for downstream analysis. The experiment uses a streaming text source to avoid storing the full dataset locally. Specifically, it loads the HuggingFaceFW/fineweb dataset with streaming=True and repeatedly draws batches of raw text strings from the training split. The key data hyperparameters are batch_size = 48 documents per step and max_length = 1024 tokens per document (with padding and truncation enabled). The total number of processed documents is set as num_docs = batch_size * 100 = 4800 , so the main loop runs for exactly num_docs // batch_size = 100 iterations. We can scan different models by changing the model name. The tokenizer is loaded from the same model name, and if no pad token is defined, the script assigns tokenizer.pad_token = tokenizer.eos_token to ensure that padded batches can be formed. A central implementation detail is that different Hugging Face model families store their final normalization layer and output head under different attribute names. To make the analysis code reusable, the script defines get_final_norm_and_head(model) , which checks common architectures (GPT-NeoX/Pythia, GPT-2, Llama, and MPT) and returns a pair (final_norm, lm_head) . For Pythia specifically, this resolves to model.gpt_neox.final_layer_norm and model.embed_out . In addition, the script stores a reference to the last transformer block last_layer = model.gpt_neox.layers[-1].eval() , because later it will apply this final block to intermediate hidden states in order to compute intermediate-layer logits in a way that matches the model’s own last-step computation. Specific operations for other model families are similar and can be found in the corresponding code. Each batch of text is tokenized using padding="max_length" , truncation=True , and max_length=1024 , which produces input_ids and an attention_mask of shape (batch_size, seq_len) . The analysis is done at next-token prediction positions, so the script forms targets as targets = input_ids[:, 1:] and considers prediction sites aligned to input_ids[:, :-1] . To ensure that both the conditioning token and the target token are non-padding, it defines a boolean mask \texttt{valid\_mask}=(\texttt{attention\_mask}[:,1:]>0)\ \&\ (\texttt{attention\_mask}[:,:-1]>0), which has shape (batch_size, seq_len-1) and is used to filter out padded positions from every saved statistic. The forward pass calls the model with output_hidden_states=True , use_cache=False , and return_dict=True . This returns out.hidden_states , a tuple of length num_layers+1 containing tensors of shape (batch_size, seq_len, hidden_size) ; index 0 corresponds to the input embedding output, and index num_layers corresponds to the final hidden state before the output head. From these hidden states, the script computes per-layer increments dh[layer] = hidden_states[layer+1] - hidden_states[layer] for all layers. These quantities are then used to quantify geometric alignment across depth. The script accumulates several layerwise statistics across all valid tokens, storing them on CPU to keep GPU memory usage bounded. For each layer index layer in 0,…,L , it saves the hidden-state norm ||h_layer|| at prediction positions, implemented as hidden_states[layer][:,:-1,:].norm(dim=-1) and filtered by valid_mask . To evaluate predictive quality as a function of depth, the script computes cross-entropy for each layer’s representation using a consistent readout pipeline. For the final layer \ell -1 , it directly uses out.logits from the model forward pass and computes per-token cross-entropy at each position as CE = -log_p(target) after applying log_softmax . For all earlier layers 0,…, \ell -2 , it first applies the model’s final transformer block to that layer’s hidden state, using last_layer(hidden_states[layer], attention_mask=prepared_mask, position_ids=position_ids)[0] . Here prepared_mask is a broadcasted attention mask, constructed so that padded keys contribute a large negative value ( torch.finfo(dtype).min ) in attention logits, and position_ids are derived by cumulatively counting non-pad tokens so that positions begin at 0 over real tokens. The resulting activations are then passed through the model’s final layer norm and output head, logits = lm_head(final_norm(hidden_state_)) , and the script computes per-token cross-entropy values CE_layer that are filtered by valid_mask and concatenated across batches. Although the script also computes tokenwise entropy H_layer , that portion is commented out and is not saved in the current run. In addition to predictive loss, the script records geometric rotation diagnostics across depth. For each layer l , it computes Theta[ l ] , the angle between consecutive hidden states at that token position, defined as \theta(h_{l},h_{l+1})=\arccos\!\Big(\mathrm{clip}(\cos(h_{l},h_{l+1}),-1,1)\Big), implemented via F.cosine_similarity(...).acos() with clipping for numerical stability. It also computes Theta_dh[ l ] , the angle between consecutive depth increments dh[ l ] and dh[ l +1] , which measures how the direction of the representation update changes from layer to layer. Finally, it computes angle_to_end[ l ] , the angle between the layer- l hidden state and the final hidden state h_ \ell , which quantifies how aligned intermediate representations are with the model’s terminal representation. All of these angles are computed at prediction positions (:,:-1,:) and stored only for valid_mask positions. After processing all 4800 documents, the script prints the wall-clock time for the full run and then stacks the accumulated CPU vectors into dense tensors. The resulting saved arrays have shapes (N_tokens, \ell ) for cross_entropy , theta , and angle_to_end ; shape (N_tokens, \ell -1) for theta_dh ; and shape (N_tokens, \ell +1) for norms , where N_tokens is the total number of valid (non-pad) next-token prediction positions observed across all batches. These tensors are saved with torch.save to the output path under the keys cross_entropy , theta , theta_dh , norms , and angle_to_end , providing a compact dataset for subsequent statistical analysis of depth-dependent behavior in the model. A.2 Main Text Figures From Section A.1 , we can obtain \theta(h_{l},h_{l+1}) for each token (i.e., Theta[ l ] ) and \theta(\Delta h_{l},\Delta h_{l+1}) for each token (i.e., Theta_dh[ l ] ). Panels a and e in Figure 2 simply plot the raw data for the first 2000 tokens in the dataset. In total, we evaluated 2M tokens in the FineWeb dataset ( 34 ) . \langle\cdot\rangle_{\mathcal{D}} is obtained by averaging the value over all the tokens evaluated. And \langle\cdot\rangle_{\mathcal{D},l} is obtained by averaging the value over all the tokens and middle layers (first and last layers excluded). Those averaged values yield panels c and d in Figure 2 . \theta(h_{l},h_{l+1}) , l=0,...,\ell-1 , for one token form an \ell -dimensional vector. Across the dataset, we can then have as many \ell -dimensional vectors as the number of evaluated tokens. After performing PCA on these vectors with scipy and choosing the first two principal components to show the data, we obtain panel b in Figure 2 . The early stopping vector is constructed as np.array([np.pi/2]+[0.45]*6+[0.0]*16+[np.pi/2]) ( \ell=24 for Pythia-410m) and the evenly in the middle vector is np.array([np.pi/2]+[0.45]*22+[np.pi/2]) . We project the two \ell -dimensional vectors into the subspace spanned by the first two principal components. A.3 Additional Results Figure 6 : Hidden state properties from more model sizes in Pythia. Hidden state norms are included. The results qualitatively agree with Figure 2 . Figure 7 : Hidden state properties from Qwen-2.5 models ( 43 ) . Hidden state norms are included. The results qualitatively agree with Figure 2 . Figure 8 : Hidden state properties from OPT models ( 44 ) . Hidden state norms are included. The results qualitatively agree with Figure 2 . The OPT models seem to be more different from Pythia and Qwen models. The last layer does not rotate a large angle on average, and the clusters in the PCA plots are very close. We hypothesize that early layers (the first layer excluded) of OPT are in ensemble averaging, and the last few layers, not only the last one, are performing a fixed operation. Although the clusters are less distinguishable, the small one related to the early stopping still occupies a small fraction of tokens, e.g., 1.3 % for OPT-13b (we use PC1 >0.5 as the criterion of being in the small cluster). We tested more model sizes other than 410m in the Pythia suit ( 5 ) . The results agree with Figure 2 , see Figure 6 . We tested more model families, Qwen-2.5 ( 43 ) , and OPT ( 44 ) . The results also agree with Figure 2 , see Figure 7 and Figure 8 , respectively. Appendix B LLM Scaling Laws B.1 Methods We used the extracted data in 4 from the Chinchilla scaling law paper ( 18 ) . The data contains model size N , dataset size D , and loss L . We next used the table at the end of 18 to find the width m and depth \ell for each model size N . We next describe how we fit the scaling exponents and prefactors for depth, width, and dataset size in the empirical scaling law model. The high-level goal is to model the validation loss as a sum of independent power-law contributions from model width, model depth, and dataset size, and to estimate both the exponents and prefactors directly from empirical measurements. The fitting procedure is implemented as a nonlinear regression in log-space using gradient-based optimization, followed by an uncertainty analysis using the Jacobian of the residuals. We load empirical scaling measurements from the file get_fitting_data_chinchilla.csv , which contains the model width m (hidden dimension), number of layers \ell , dataset size D (number of training tokens), and the observed validation loss. We exclude a subset of high-loss models to avoid outliers dominating the fit. Concretely, we exclude nr_of_models_excluded = 40 data points with the largest losses, and retain all other data points for fitting. Excluding fewer data points will make the dataset part loss (as is in Figure 10 ) not like a power law visually after plotting. Excluding more data points will result in similar fitting values but with a larger uncertainty. There are 203 data points left. The remaining variables are converted to PyTorch tensors for automatic differentiation. We assume an additive scaling law of the form L(m,\ell,D)=\frac{c_{m}}{m^{\alpha_{m}}}+\frac{c_{\ell}}{\ell^{\alpha_{\ell}}}+\frac{c_{D}}{D^{\alpha_{D}}}+L_{0}, where c_{m},c_{\ell},c_{D} are positive coefficients, \alpha_{m},\alpha_{\ell},\alpha_{D} are scaling exponents, and L_{0} is an irreducible loss floor. To ensure positivity of the coefficients during optimization, we parametrize them in log-space as c_{x}=\exp(\ln c_{x}) . The fitted parameters are therefore (\ln c_{m},\ln c_{\ell},\ln c_{D},\alpha_{m},\alpha_{\ell},\alpha_{D},L_{0}) . We optimize the parameters using the Adam optimizer with separate learning rates for coefficients and exponents: \text{lr}(\ln c_{x})=0.005,\quad\text{lr}(\alpha_{x})=0.0005,\quad\text{lr}(L_{0})=0.005, for a total of num_epochs = 50000 gradient steps, which ensures convergence. The objective function is the mean squared error in log-loss space, \mathcal{R}=100\cdot\mathbb{E}\big[(\log L_{\text{obs}}-L_{\text{model}})^{2}\big], which stabilizes optimization across several orders of magnitude in dataset size and model scale. After convergence, we compute residuals r_{i}=L_{\text{obs},i}-L_{\text{model},i}, and relative residuals r_{i}/L_{\text{obs},i} for diagnostic plots. To quantify parameter uncertainty, we compute the Jacobian of the residuals with respect to all fitted parameters using PyTorch automatic differentiation. Let J be the Jacobian matrix with entries J_{i\theta}=\partial r_{i}/\partial\theta . Assuming homoscedastic Gaussian noise, the parameter covariance matrix is estimated as \Sigma=\sigma^{2}(J^{\top}J)^{-1}, where \sigma^{2} is the empirical variance of the residuals. Standard errors for each fitted parameter are extracted from the diagonal of \Sigma . Finally, we visualize the fitted scaling law by subtracting the fitted width and depth contributions from the observed losses and plotting the residual dataset-size scaling term on log-log axes. B.2 Main Text Figures The data points in Figure 2 f is obtained from L-\frac{c_{m}}{m^{\alpha_{m}}}-\frac{c_{D}}{D^{\alpha_{D}}}-L_{0}, where L is actual loss data and other terms are calculated with the fitted values. The line is \frac{c_{\ell}}{\ell^{\alpha_{\ell}}} drawn with fitted c_{\ell} and \alpha_{\ell} . B.3 Additional Results We can do the same to evaluate the width part loss by plotting L-\frac{c_{\ell}}{\ell^{\alpha_{\ell}}}-\frac{c_{D}}{D^{\alpha_{D}}}-L_{0}, along with the theory \frac{c_{m}}{m^{\alpha_{m}}} ( Figure 9 ). Figure 9 : Loss due to the width agrees well with a power law. The dashed line is based on \frac{c_{m}}{m^{\alpha_{m}}} drawn with fitted c_{m} and \alpha_{m}=0.98\pm 0.08 . The dataset part loss is estimated by L-\frac{c_{m}}{m^{\alpha_{m}}}-\frac{c_{\ell}}{\ell^{\alpha_{\ell}}}-L_{0}, which is plotted in Figure 10 . Figure 10 : Loss due to the dataset size agrees well with a power law. The dashed line is based on \frac{c_{D}}{m^{\alpha_{D}}} drawn with fitted c_{D} and \alpha_{D}=0.30\pm 0.01 . Based on our results that \alpha_{m}\approx 1 and \alpha_{\ell}\approx 1 , we should have the optimal width-depth relationship given a fixed number of parameters N\approx 12m^{2}\ell as m\propto\ell where the coefficient is only related to c_{m} and c_{\ell} . We next check the relationship between depth and width in actual models ( Figure 11 ). It seems that \ell\propto m^{0.67} can better fit most of the data we have. Yet, when both m and \ell are large enough, it seems the models are following m\propto\ell , agreeing with our expectation. We are not sure if previous works ( 18 ; 9 ) scanned depth and width and used the optimal relationship. If they did, there might be two reasons at first m\propto\ell is not satisfied. One is that when m is small contribution from attention layers and language model head to N is significant, whose number of parameters is proportional to m rather than m^{2} , making the dependence of N on m complicated and obscuring the optimal m - \ell relation. The second is that when m and \ell are small, some crossing terms of m and \ell may still play a role in loss. Figure 11 : Actual width-depth relationship in models. The dark points are from Chinchilla models ( 18 ) and the red points from GPT-3 ( 9 ) . We find that when both m and \ell are large (the last three red points), m\propto\ell (red dashed line) is true, agreeing with our theory. Chinchilla models are fitted by the dark dashed line, suggesting \ell\propto m^{0.67} . In our observation, the first and the last layers may not be in the ensemble averaging regime. The number of layers in the ensemble averaging regime is better described by \ell-2 . And the loss may be a clearer power law with \ell-2 . We next try to fit the loss with the form L=\frac{c_{m}}{m^{\alpha_{m}}}+\frac{c_{\ell}}{(\ell-2)^{\alpha_{\ell}}}+\frac{c_{D}}{D^{\alpha_{D}}}+L_{0}. It turns out that the newly fitted \alpha_{\ell}=1.1\pm 0.2 . Compared to the previous one, \alpha_{\ell}=1.2\pm 0.3 , we indeed obtain a better power law as the standard deviation decreases, and this better power law has an exponent closer to the prediction 1.0 . The other newly fitted exponents, \alpha_{m}=0.96\pm 0.08 and \alpha_{D}=0.30\pm 0.01 , remain similar to the old ones. The fact that the fitted power law gets closer to the theoretical expectation after making a more accurate estimation of the number of layers that may be in the ensemble averaging regime, based on hidden state behaviors, supports the argument that those layers are in the ensemble averaging regime. We plot the loss by parts as before in Figure 12 (width part), Figure 13 (depth part), and Figure 14 (dataset part). Furthermore, if we plot \ell-2 versus m with data from 18 ; 9 , the fitted relationship \ell-2\propto m^{0.74} ( Figure 15 ) is closer to the asymptotic optimal relationship \ell-2\propto m predicted comparing to previous situation where we studied \ell versus m . We should focus on the scaling relationship between the number of layers in the ensemble averaging regime and the width. Though, \ell-2\approx\ell asymptotically for large depth \ell . Figure 12 : Loss due to the width agrees well with a power law (fitted after replacing \ell by \ell-2 ). The dashed line is based on \frac{c_{m}}{m^{\alpha_{m}}} drawn with fitted c_{m} and \alpha_{m}=0.96\pm 0.08 . Figure 13 : Loss due to the depth agrees well with a power law (fitted after replacing \ell by \ell-2 ). The dashed line is based on \frac{c_{\ell}}{(\ell-2)^{\alpha_{\ell}}} drawn with fitted c_{\ell} and \alpha_{\ell}=1.1\pm 0.2 . Figure 14 : Loss due to the dataset size agrees well with a power law (fitted after replacing \ell by \ell-2 ). The dashed line is based on \frac{c_{D}}{m^{\alpha_{D}}} drawn with fitted c_{D} and \alpha_{D}=0.30\pm 0.01 . Figure 15 : Actual width-depth relationship in models. The dark points are from Chinchilla models ( 18 ) and the red points from GPT-3 ( 9 ) . We find that when both m and \ell-2 are large (the last three red points), m\propto\ell-2 (red dashed line) is true, agreeing with our theory. Chinchilla models are fitted by the dark dashed line, suggesting \ell-2\propto m^{0.74} . Appendix C Toy Model Experiments C.1 Methods This experiment studies how learning dynamics depend on the entropy of the target distribution by using a synthetic teacher-student setup. The high-level goal is to generate a controllable next-token distribution from a teacher model with adjustable temperature, train a student model to match the teacher using KL divergence, and measure how training and test loss depend on the teacher entropy. This isolates the effect of peaked versus flat target distributions in a minimal setting, independent of natural language structure. We run the experiment as a distributed job array. The teacher distribution is controlled by a temperature parameter T , with \texttt{n\_temperatures}=16 logarithmically spaced values between \texttt{T\_min}=10^{-2} and \texttt{T\_max}=1 . For each temperature, we run \texttt{n\_teachers}=3 replicates of teachers. Evaluation is performed on n_eval_batches fresh batches of Gaussian inputs. All models are trained in full precision (FP32) on CPU, and no weight decay or dropout is used. The model architecture is described in the main text ( Section 3 ). C.2 Main Text Figures Figure 3 b is based on code exp-9.py (independent teacher weights \rho=0 ) and exp-9-1.py (independent teacher weights \rho=1 ). The learning rates are fixed as 6e-4 . Other basic information is described in Section C.1 . After obtaining the data, we fit the final test loss with Equation 8 via scipy . The training dynamics of models \ell=6 with different teacher temperatures and \rho=1 are shown in Figure 4 c. Realizing the difficulty of training in low temperature, we tried to increase the number of training steps to 80000 in Figure 4 , a and b, whose code is exp-9-3.py . To make training easier, we also gave the student the teacher head and only trained the MLP layers. It turns out that lower temperatures require orders of magnitude longer time to train and cannot converge within our compute budget. We then fit the exponent \alpha_{\ell} as Equation 8 at different steps and temperatures. At the end, to ensure convergence and then study hidden states, we use the MSE between the last hidden states to train the models. The fitted exponents agree with theory well ( Figure 20 ). The code is in exp-9-6.py . We then study the angle between hidden states and the angle between updates, the same as LLM experiments. Some of the results form Figure 5 . C.3 Additional Results Figure 16 : Training dynamics of exp-9.py for Figure 3 . Under high temperatures, the training converge. Under low temperatures, the loss is dominated by training part, not model size part or not limited by depth. Figure 17 : Training dynamics of exp-9-1.py for Figure 3 and Figure 4 c. Under high temperatures, the training converge. Under low temperatures, the loss is dominated by training part, not model size part or not limited by depth. Figure 18 : Training dynamics of exp-9-3.py for Figure 3 , a and b. Under high temperatures, the training converge. Under low temperatures, the loss is dominated by training part, not model size part or not limited by depth. Figure 19 : Training dynamics of exp-9-6.py for Figure 5 . We use MSE, which tend to ensure training convergence. Figure 20 : Fitted \alpha_{\ell} from exp-9-6.py agree with theory. The diamonds correspond to \rho=1 and dots correspond to \rho=0 . The training dynamics, loss versus time, behind Figure 3 is in Figure 16 ( \rho=0 ) and Figure 17 ( \rho=1 ). Under high temperatures, it is obvious that the loss has plateau at the end. While for low temperatures, loss continue to drop and difference due to depth is not obvious, suggesting no convergence. The training dynamics, loss versus time, behind Figure 4 , a and b, is in Figure 18 . The training dynamics, loss versus time, behind Figure 5 is in Figure 19 . We use MSE for Figure 5 . The loss scaling of cross-entropy and MSE is the same, i.e., quadratic in the hidden state difference. The measured exponents \alpha_{\ell} agree with our theory ( Figure 20 ). Figure 21 : Evaluation of hidden state angle \theta(h_{l},h_{l+1}) after training from exp-9-6.py . The first row are from \rho=0 and second row from \rho=1 . The columns are from left to right are \ell=6,12,16,24,32,48 and the teacher. Figure 22 : Evaluation of hidden state angle \theta(\Delta h_{l},\Delta h_{l+1}) after training from exp-9-6.py . The first row are from \rho=0 and second row from \rho=1 . The columns are from left to right are \ell=6,12,16,24,32,48 and the teacher. In addition to the hidden states properties shown in Figure 5 , we plot all the observed data here from exp-9-6.py in Figure 21 (about \theta(h_{l},h_{l+1}) ) and Figure 22 (about \theta(\Delta h_{l},\Delta h_{l+1}) ). We find that correlation between updates \theta(\Delta h_{l},\Delta h_{l+1}) is indeed lower under \rho=1 than those under \rho=0 , suggesting smooth dynamics under \rho=1 as expected. However, the angles under \rho=1 is still far away from zero, which may be due to limited training or symmetries in the network. Figure 23 : Training dynamics of exp-9-4.py for Figure 5 . We use MSE, which tend to ensure training convergence. The new architecture is used. Figure 24 : Fitted \alpha_{\ell} from exp-9-4.py agree with theory. The diamonds correspond to \rho=1 and dots correspond to \rho=0 . The final \alpha_{\ell} under \rho=1 is not 5 , which may be due to the fact larger models have not really converge ( Figure 23 ). We also did experiments with modified student architecture. The hidden state then propagates through \ell blocks, h_{l}=h_{l-1}+\mathrm{BLOCK}_{l}(h_{l-1}),\quad l=1,\dots,\ell. (13) Within each block, there are two MLPs, h_{l-1}^{c}\equiv\frac{1}{2}\mathrm{MLP}_{l,1}(h_{l-1})+h_{l-1} (14) and \mathrm{BLOCK}_{l}(h_{l-1})=\mathrm{MLP}_{l,2}(h_{l-1}^{c}). (15) Ideally, h_{l-1}^{c} is to approximate the hidden state at center point s=(l-1)/\ell+1/2\ell between (l-1)/\ell and l/\ell . The two MLPs are approximating the hidden state gradient with s . This is a second-order method. If the underlying dynamics is smooth, the discretization error in one layer is O(\Delta s^{3}) . After error accumulation, the upper bound is O(\Delta s^{2}) and the typical behavior is O(\Delta s^{5/2}) . The loss based on the upper bound of hidden state error is L\sim 1/\ell^{4} , and the typical behavior is L\sim 1/\ell^{5} . However, if under the ensemble averaging, the theory predicts L\sim 1/\ell regardless of the architecture in one layer. We find that procedural assembly (teacher \rho=1 ) with this new architecture does have a higher \alpha_{\ell}\approx 4 than that of residual network (see Figure 20 ), while ensemble averaging (teacher \rho=0 ) remains \alpha_{\ell}\approx 1 ( Figure 24 ). We therefore further verified that teacher \rho=1 leads to procedural assembly, where a better discretization scheme can help depth scaling, while teacher \rho=0 leads to ensemble averaging, where a better discretization scheme cannot help.