An Analysis for Reasoning Bias of Language Models with Small Initialization

Junjie Yao; Zhi-Qin John Xu; Zhongwang Zhang

REVIEW 4 major objections 8 minor 3 cited by

An Analysis for Reasoning Bias of Language Models with Small Initialization

T0 review · 4 major / 8 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The scale of a transformer's weight initialization decides, in one training run, whether it learns compositional rules and generalizes or memorizes its training pairs; early embedding updates are shaped by token label distributions.

desk verdict Useful mechanism story for a known empirical result; the theory is suggestive rather than proven in the tested regime, but the paper deserves a careful referee. read the letter →

arxiv 2502.04375 v2 pith:KLUVLCAP submitted 2025-02-05 cs.CL cs.LG

Junjie Yao , Zhongwang Zhang , Zhi-Qin John Xu This is my paper · ORCID

classification cs.CLcs.LG MSC 68T0768T50

keywords initializationscalereasoningvsmemorizationembeddingdynamicstraininganchorfunctionsself-attentionlanguagemodelscompositionalgeneralization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that the scale of a transformer's parameter initialization is not a neutral training detail: smaller scales nudge the model toward learning compositional reasoning rules, while larger scales nudge it toward memorizing input–output pairs. In a single next-token prediction run over a mixed dataset, a small-initialized model fits reasoning sequences faster and generalizes to unseen reasoning combinations, whereas a large-initialized model fits the memory part but shows no generalization. The reason, the authors argue, is early embedding dynamics: each token's embedding moves along a gradient determined by the label distribution of the samples containing that token. Reasoning anchors have anchor-dependent label distributions, so their embeddings differentiate quickly and form an ordered geometry; memory anchors share one uniform distribution, so their embeddings stay similar and the memory task stays hard. The claim is supported by synthetic composition tasks, a simplified embedding-plus-MLP model, a gradient-flow theory, and GPT-2 runs on real language data.

What carries the argument

The load-bearing object is the early-time gradient flow of the embedding row $w_{\mathrm{emb},s}$, which under small initialization becomes $d w_{\mathrm{emb},s}/dt = r_s (P_s - \frac{1}{d}\mathbf{1})\tilde W$ (Propositions 1–3), where $P_s$ is the label distribution of the samples containing token $s$ and $\tilde W$ collects the effective value/output and feedforward paths. The argument hangs on the contrast between (7) and (8): every memory anchor has the same uniform label distribution, while each reasoning anchor has a shifted distribution, so the same gradient law differentiates reasoning embeddings and leaves memory embeddings similar. Theorem 1 then approximates the reasoning-anchor embedding by a Gaussian bump $e^{-(i-s)^2/(2\sigma_P)}$, and the first attention module is shown to act as a fixed average operator whose value projection is dominated by reasoning anchors, carrying the differentiated signal forward.

What would settle it

Track the norm of the attention-matrix derivative $\|dA/dt\|$ against the embedding-gradient norm $\|dw_{\mathrm{emb}}/dt\|$ during the first 200 epochs of the synthetic task at $\gamma = 0.8$; if attention gradients are not comparatively small, the frozen-attention premise fails. Alternatively, relabel the reasoning task so that every reasoning anchor has the same label distribution as every other anchor; if the reasoning bias persists under small initialization, the label-distribution mechanism is not the whole story.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that smaller initialization scales bias language models toward reasoning tasks and larger scales bias them toward memorization, and that the bias is governed by how label distributions shape the embedding space during the earliest training steps. Formally, with small initialization ($\gamma > 0.5$ in the paper's scheme, where weights are drawn as $N(0, d^{-2\gamma})$), the attention matrix of the first layer acts like an averaging operator, and the gradient flow of a token's embedding $w_{\mathrm{emb},s}$ reduces to $d w_{\mathrm{emb},s}/dt = r_s (P_s - \frac{1}{d}\mathbf{1})\tilde{W}$, with $P_s$ the label distribution over samples containing $s$. Because $P_s$ is the same uniform distribution for every memory anchor but shifts with the anchor value for reasoning anchors, reasoning embeddings separate early and acquire a hierarchical, number-ordered structure, while memory embeddings remain undifferentiated. Theorem 1 approximates the resulting reasoning-anchor embeddings by $e^{-(i-s)^2/(2\sigma_P)}$ and bounds the inner-product error. The value projection $W^V$ of the first attention block develops a dominant singular direction aligned with reasoning anchors, so the reasoning signal propagates to all subsequent tokens, whereas memory anchors are not distinctly captured.

Load-bearing premise

Everything in the theoretical explanation depends on treating the attention matrix as a fixed averaging operator during early training and ignoring the gradients of the attention weights; if those gradients are not actually small in the window where the bias forms, the mechanism is incomplete.

Editorial extensions

If this is right

Initialization scale becomes a practical control knob: with $\gamma > 0.5$ a small transformer on a mixed reasoning/memory dataset will preferentially learn and generalize the reasoning part, while $\gamma < 0.5$ makes it memorize the training pairs instead.
The reasoning bias is not about the model's capacity or the data's complexity but about early embedding differentiation, so tasks whose per-token label distributions are structured inherit the bias automatically.
Memory-style tasks can be made learnable even as arbitrary mappings if their label distributions vary with the anchor: the paper shows that an anchor-dependent memory task is learned faster than a fully uniform one.
Real pretraining may inherit the same effect: in GPT-2 runs on PrOntoQA plus TinyStories, the loss advantage of the reasoning dataset grows as the initialization scale shrinks, and the reasoning-token embeddings become more distinguishable.
Under small initialization, the first attention block's value projection aligns with reasoning anchors while the attention matrix stays near an average operator, giving a concrete internal signature that distinguishes a reasoning-biased run from a memorizing one.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An untested implication is that the bias generalizes beyond arithmetic: any per-token label structure that encodes an ordering or a rule, such as dates, coordinates, or hierarchical codes, should benefit from small initialization, while flat, arbitrary fact associations should be chronically disfavored.
As an extension not explored in the paper, one could deliberately engineer label-distribution diversity for key tokens through auxiliary objectives or embedding regularization to induce reasoning-like generalization without shrinking the initialization scale.
The frozen-attention window may be a narrow early-training phase; at larger widths, deeper stacks, or with aggressive learning-rate schedules, attention gradients could become significant before the embedding geometry sets, which would make the bias a small-model, early-training phenomenon rather than a universal LLM property.
A testable extension: after a small-initialization pretraining phase that has already differentiated reasoning embeddings, introducing a fresh memory-style task should proceed slowly; if correct, task order should interact with the bias.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. The headline phenomenon, small initialization favoring reasoning over memorization, was already in the authors' own NeurIPS 2024 paper and its follow-up, and the paper says so. What is actually new is the mechanism story: label distributions shape embedding differentiation early in training. They test that with synthetic anchor tasks, an Emb-MLP, and a real GPT-2 run on PrOntoQA and TinyStories. The core empirical trend looks consistent, and the F_mem,2 control in Section 3.5, where the only change is making memory labels anchor-dependent, is a genuine controlled test of the label-distribution intuition. I believe the empirical direction.

The soft spots are real but not fatal. The Transformer gradient-flow derivation in Lemma 3, Appendix B.5, differentiates the embedding only through the value and residual paths and freezes the attention matrix at its average-operator limit. The derivative of the attention matrix with respect to the embedding is generically nonzero and is never bounded. The stress-test note adds that Lemma 1's average-operator guarantee only kicks in around gamma ≈ 3 for the stated dimensions, while all experiments use gamma = 0.3, 0.5, 0.8. So the theory does not lock down the regime it tests; it is a suggestive model rather than a proof of the mechanism in the actual experiments. That is the load-bearing gap, and the paper itself, in Appendix B.8, admits that a general explanation for the W_V preference would require more sophisticated analysis. I would want the claims scaled back or the omitted attention-derivative term addressed.

Minor issues: single runs without seeds/error bars; no code or data release; the validation of Theorem 1 in Eq. (14) fixes the Gaussian width and normalization by hand; and identifying TinyStories as a memorization task is asserted rather than justified. The label-distribution difference between memory and reasoning anchors is built into the synthetic task construction, which is fine, but it means the theory is explaining a property of the design rather than an independent discovery.

Who is this for? Researchers who care about initialization, training dynamics, and small-model interpretability. Even with the theory incomplete, the phenomenon itself is worth knowing, and the paper is coherent enough to deserve a serious referee rather than a desk reject. I would send it to peer review with a request for seeds and error bars, code/data release, and either a bounded treatment of the attention derivative or a more modest statement of what the theory establishes.

Referee Report

4 major / 8 minor

Summary. The paper studies how the parameter initialization scale of a Transformer affects whether the model learns compositional reasoning mappings or memorizes input–output pairs. In a synthetic anchor-function task and in a GPT-2 experiment mixing PrOntoQA and TinyStories, the authors report that a smaller initialization scale (γ=0.8) favors reasoning, while a larger scale (γ=0.3) favors memorization. They propose a mechanism in which early embedding gradients are driven by the label distribution of each token: memory anchors have uniform label distributions and thus undifferentiated embeddings, whereas reasoning anchors have anchor-dependent label distributions and thus develop a hierarchical embedding structure. The paper formalizes this with an Embedding-MLP analysis (Proposition 1), a one-layer Transformer gradient-flow analysis (Propositions 2–3, Theorem 1), and an analysis of the first and second attention modules, and it validates the qualitative embedding patterns against experiments.

Significance. If the empirical finding holds, initialization scale is an easily controllable training bias that determines whether a small Transformer learns a compositional rule and generalizes or simply memorizes training data. The synthetic task is carefully designed, the empirical direction is consistent across the Transformer, Emb-MLP, and GPT-2 settings, and the paper offers a concrete, testable mechanism based on label distributions. The theoretical framework, however, is not yet on solid ground: the Transformer gradient-flow derivation neglects derivatives through the attention weights, Lemma 1 holds provably only for initialization scales much larger than those tested, and the validation of Theorem 1 fixes Gaussian width and normalization by hand. The empirical contribution is valuable, but the claimed explanatory mechanism requires additional analysis.

major comments (4)

[§3.3 / Appendix B.5, Lemma 3] The gradient-flow derivation for the one-layer Transformer differentiates fθ with respect to W_emb only through the value and residual paths, treating the attention matrix A as a fixed average operator. Since A = softmax(mask(QK^T)/√dk) depends on W_emb through Q and K, the omitted dA term is generically nonzero, and the paper does not bound it. Propositions 2 and 3 and Theorem 1 therefore do not yet establish that label distributions alone drive the early embedding differentiation in an actual Transformer. I recommend either deriving a bound on the dA contribution during the early-training window or adding a numerical comparison of the retained versus omitted gradient terms.
[Lemma 1 / §3.3] The regime in which A is provably near the average operator is not the regime tested in the experiments. With d_m=200, d_k=64, ε=δ=0.1, the proof of Lemma 1 requires γ ≳ 2.8, while all experiments use γ=0.3, 0.5, and 0.8. Figure 6 shows the average-operator behavior for γ=0.8 only at epoch 200, which is after the early window in which the embedding differentiation is claimed to occur. The theoretical mechanism is therefore not verified for the experimental setting; please extend the lemma to the tested range, justify the extrapolation, or present experiments in the provable regime.
[Theorem 1 / Eq. (14) / Appendix B.7] The validation of Theorem 1 fixes the Gaussian width and normalization in Eq. (14) by hand, and the stated error bound (13) yields O(d_m^{1-γ}(q^{-1/2}+d_m^{-γ})) ≈ O(2.4) for the experimental values (d_m=200, γ=0.8, q=2), which is too loose to certify the cosine-similarity match shown in Figure 5C. The visual agreement could therefore be sensitive to the hand-set constants; a principled estimate of C1, C2, and σP or a quantitative error metric is needed to substantiate the theoretical validation.
[§3.1–§3.4, all figures] All reported experiments appear to be single runs without error bars or multiple seeds. Since the central empirical claim is a comparison of learning speeds and generalization behaviors across γ values, the observed differences could in principle be within run-to-run variability. Please report the number of seeds and the variance of the loss, accuracy, and cosine-similarity measures, or otherwise justify that the effects are stable.

minor comments (8)

[§2.3, Eq. (2)] The masking operation mask(·) is used without definition; please specify the causal mask explicitly.
[§3.1 / §2.2] The relationship between the 'last token prediction' described in §3.1 and the label construction defined in §2.2 should be stated explicitly, since the dataset definition refers to the label of the full sequence.
[§3.5] The notation '{30, ..., 29 + 20 × i}' for the four label-range groups is confusing; the intended ranges (e.g., 30–49, 30–69, 30–89, 30–109) should be written out.
[Appendix E, Assumption 4] The sign convention in cos(w_pos,i, w_pos,j) = cos(|i−j|π/L) should be checked against the increasing attention pattern in Figure 15C; if the cosine is negative for large |i−j|, the stated increasing behavior requires clarification.
[Theorem 2] Theorem 2 assumes N_Z = d_m, which is not satisfied in the experiments (N_Z=100, d_m=200); the paper should state clearly that this result is not directly validated by the reported experiments.
[§2.2 / §3.2] Equations (7) and (8) are immediate consequences of the dataset construction: memory labels are uniform by definition and reasoning labels are sums of keys and anchors by definition. The authors should state this explicitly so that the empirical content is understood to lie in the dynamical response of the embeddings to these distributions.
[Throughout] The paper does not mention whether code and data will be released; an availability statement would improve reproducibility.
[Figure 4 and Appendix B.5] There are several typesetting issues, such as 'P s − 1 dvob 1' in Figure 4B and the phrase 'as the initialization scales decrease to zero' in Appendix B.5, which conflicts with the γ→∞ limit used in Propositions 2–3; please proofread these passages.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the synthetic label distributions are definitions, the empirical loss and embedding results are measured, and the self-citations are background rather than load-bearing.

full rationale

The claimed derivation chain is: (i) the synthetic benchmark defines memory labels as uniform draws from Z and reasoning labels as sums of the key and anchors, which yields the label distributions in Eqs. (7)-(8); (ii) Proposition 1 and the small-initialization limit express the embedding gradient as proportional to P_s minus a uniform vector; (iii) Propositions 2-3 specialize this gradient flow to memory and reasoning anchors; and (iv) the predicted embedding differentiation is compared with measured cosine-similarity matrices, while the reasoning bias itself is read from independently measured loss curves on Dmem, Drsn,train, Drsn,test and on the external PrOntoQA and TinyStories corpora. Equations (7)-(8) are statements of the data-generation process, not fitted predictions, so the fact that reasoning anchors have s-dependent label distributions is a definitional input rather than a circular output. The empirical claims could have failed (for example, a larger initialization could have ignored the label structure), and Figures 3, 5, and 7 report measured quantities. The main theoretical weakness is the frozen-attention approximation in Lemma 3 and Appendix B.5, which omits derivatives of the attention matrix, together with the regime mismatch between Lemma 1's asymptotic guarantee and the tested gamma=0.8; these are rigor and correctness concerns, not circularity. The self-citations (Zhang et al. 2024a,b; Xu et al. 2025; Chen et al. 2024) motivate the anchor-function framework and condensation intuition, but they are not the load-bearing proof because the paper re-derives its gradient-flow results and validates them on external benchmarks. Therefore no load-bearing step reduces by construction to its own input.

Assumptions & free parameters 1 free parameters · 7 assumptions · 0 invented entities

The central derivation uses a linearized gradient-flow model of the embedding layer, valid only in the limit of infinitesimal initialization and infinite data. It relies on the specific construction of the anchor-function tasks, and the main quantitative check fixes a Gaussian width by hand. The frozen-attention assumption is the largest unstated simplification.

free parameters (1)

Gaussian width and normalization in Theorem 1 = sigma_P^2 = 6 (exponent (i-s)^2/12), C1*C2 = 1 in Eq. (14)
Theorem 1's ~w^s uses unspecified C1, C2, sigma_P. The validation in Eq. (14) sets them by hand to e^{-(i-s)^2/12}, so the later cosine-similarity match is a fit, not a prediction.

assumptions (7)

standard math Assumption 1: activation sigma is C^2 with bounded first and second derivatives, sigma(0)=0, sigma'(0)=1.
Used to linearize the network at small initialization so that sigma' = 1 and softmax = uniform. Standard smoothness assumption, but real transformers use activations that are not exactly linear at initialization.
ad hoc to paper The attention matrix A is treated as constant when deriving the embedding gradient in the one-layer Transformer (Lemma 3).
Lemma 3 in Appendix B.5 differentiates only the direct W_emb terms and treats the attention matrix as fixed; no bound is given for the neglected derivative of the attention weights.
domain assumption The limits n -> infinity and gamma -> infinity are taken simultaneously with r_s constant.
Propositions 1, 2, 3 replace empirical label frequencies by their expectations and activations by their linear limit. This is an infinite-data, infinitesimal-init idealization that may not hold over the full training window.
standard math The discrete sum of uniform random variables is approximated by a Gaussian via Berry-Esseen (Theorem 1).
Used to justify the Gaussian bump form of ~w^s; the stated error O(q^{-1/2}) follows from CLT but the constants in the approximation remain unset.
domain assumption Task construction: memory labels are uniform over Z independent of the anchor, reasoning labels are the sum of the key and anchors (Eqs. 7 and 8).
These definitions directly imply that memory anchors share a label distribution while reasoning anchors have anchor-dependent distributions. The differentiated label structure is enforced by the benchmark design, not learned.
ad hoc to paper Word embeddings and position embeddings are orthogonal and position-position cosine follows cos(|i-j| pi / L) (Assumptions 2-4, Appendix E).
Used to build the cliff-sequence reconstruction of the second attention module; these are idealized assumptions not verified in the trained models.
ad hoc to paper Theorem 2 assumes NZ = dm and a 1:1 ratio of memory to reasoning mappings.
This special constrained condition is explicitly conceded in Appendix B.8 to be a limitation of the general explanation for the W^V preference.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Analysis for Reasoning Bias of Language Models with Small Initialization." pith.science (2026). https://pith.science/paper/KLUVLCAP

@misc{pith2026250204375,
  author       = {Pith},
  title        = {Pith review of: An Analysis for Reasoning Bias of Language Models with Small Initialization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KLUVLCAP}},
  note         = {Machine review of arXiv:2502.04375}
}

read the original abstract

Transformer-based Large Language Models (LLMs) have revolutionized Natural Language Processing by demonstrating exceptional performance across diverse tasks. This study investigates the impact of the parameter initialization scale on the training behavior and task preferences of LLMs. We discover that smaller initialization scales encourage models to favor reasoning tasks, whereas larger initialization scales lead to a preference for memorization tasks. We validate this reasoning bias via real datasets and meticulously designed anchor functions. Further analysis of initial training dynamics suggests that specific model components, particularly the embedding space and self-attention mechanisms, play pivotal roles in shaping these learning biases. We provide a theoretical framework from the perspective of model training dynamics to explain these phenomena. Additionally, experiments on real-world language tasks corroborate our theoretical insights. This work enhances our understanding of how initialization strategies influence LLM performance on reasoning tasks and offers valuable guidelines for training models.

Figures

Figures reproduced from arXiv: 2502.04375 by the authors.

**Figure 1.** Comparison of training loss between PrOntoQA and TinyStories in one next-token prediction training for this mix dataset. The red line represents the training loss on the PrOntoQA dataset, while the blue line depicts the training loss on the TinyStories dataset. wide range of tasks, from text generation to complex reasoning (Wei et al., 2022a; Achiam et al., 2023; Liu et al., 2024). Reasoning, in particular, is a c… view at source ↗

**Figure 2.** Schematic diagram of the synthetic composition task. The gray-shaded area illustrates the specific setup used in this example. Each block represents a token within the input sequence, with different face colors indicating distinct token types (blue: noise, orange: key, green: anchor). Each row corresponds to an input sequence paired with its respective label. The left section depicts four examples of memory mapping,… view at source ↗

**Figure 3.** A: Loss and prediction accuracy of the models on different datasets under varying initialization scales (γ = 0.3, 0.5, 0.8). The top row depicts the evolution of the loss during training for three datasets: Dmem (blue lines), Drsn,train (purple lines), and Drsn,test (orange lines). The bottom row presents the corresponding prediction accuracies for these datasets. Each column represents results obtained with differe… view at source ↗

Figures from the paper (14 more)

**Figure 4.** Figure 4: A: Cosine similarity matrices for memory (top row) and reasoning (bottom row) anchors at epoch 50 (left) and epoch 900 (right) of a model initialized with γ = 0.8. B: Distribution of P s − 1 dvob 1 for different reasoning anchor s. C: Cosine similarity between P si − 1…

**Figure 5.** Figure 5: Embedding structure of a Transformer model with small initialization scale. A: Cosine similarity matrices for memory (top) and reasoning (bottom) anchors at epoch 200 (left) and epoch 900 (right). B: Visualization of the embedding space projected onto the first two pri…

**Figure 6.** Figure 6: Characteristics of the first attention module under small initialization (γ = 0.8) in the early training stage (epoch 200). A: Heatmap of the attention matrix for a random sample. B: Distribution of the relative error between attention Ajk and 1 j across all training s…

**Figure 7.** Figure 7: Reasoning bias of GPT-2 in real language tasks. A: dynamics of ∆L during early training stage with initializations scales γ = 0.3, 0.5 and 0.8. B: Cosine similarity among embedding space of PrOntoQA dataset and TinyStories dataset at step 5000 with γ = 0.8. detailed fo…

**Figure 8.** Figure 8: Left: Distribution of targets and predictions in 4 groups of memory tasks. Red represents the target distribution in each group and blue represents the prediction distribution. Right: Learning speed comparison for Fmem,1 (red) and Fmem,2 (blue). Previous sections reve…

**Figure 9.** Figure 9: Cosine similarity comparison between experimental results cos wemb,si , wemb,sj with theoretical approximations cos w˜ emb,si , w˜ emb,sj (see (12)), for any si, sj ∈ Arsn. Theorem 2. Let n, γ → ∞, NZ = dm, Define that A ∼ U (Arsn) and Y as a random variable which …

**Figure 10.** Figure 10: exhibits the cosine similarity within the embedding space of Emb-MLP models with initialization rates γ = 0.3 and γ = 0.5. The results indicate that under a large initialization scale, the embedding space of the model becomes less influenced by the label distributions…

**Figure 11.** Figure 11: Characteristic of embedding space of Transformer with initialization rates γ = 0.3, 0.5. The left and middle panels depict the cosine similarity among embedding vectors of memory anchors and reasoning anchors at epochs 200 and 900. The right panel shows a PCA projecti…

**Figure 12.** Figure 12: Characteristics of the first attention module of Transformers (step 200) with initialization rates γ = 0.3 (top row) and γ = 0.5 (bottom row). A: Heatmap of the attention matrix for a random sample. B: Distribution of the relative error between attention Ajk and 1 j a…

**Figure 13.** Figure 13: The distribution of singular value in different parameter matrices under different initialization scales (step 200). We denote the singular value vector by S and the i-th largest singular value by si. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_13.png]

**Figure 14.** Figure 14: exhibits the cosine similarity within the embedding space of PrOntoQA and TinyStories tasks trained by GPT-2 models with initialization rates γ = 0.3 and γ = 0.5. It’s noted that under a large initialization scale, the embedding vectors are mutually orthogonal, indica…

**Figure 15.** Figure 15: Characteristic of the second attention module. A: The last row of the second attention matrix (without applying softmax) for a randomly selected sequence. B: A heatmap of the second attention matrix for the same sequence. C: The last row of the matrix QWposWK/ √ dk im…

**Figure 16.** Figure 16: Training dynamics of Transformers under γ = 0.3, 0.5, 0.8 without Layer Normalization. G. Learning Rate We conduct experiments with the learning rate belonging to 10−5 , 5 × 10−4 [PITH_FULL_IMAGE:figures/full_fig_p030_16.png]

**Figure 17.** Figure 17: Training dynamics of Transformers under γ = 0.3, 0.5, 0.8 and varying learning rates. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_17.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Understanding LoRA as Knowledge Memory: An Empirical Analysis
cs.LG 2026-03 conditional novelty 7.0 of 10

LoRA modules function as composable knowledge memories for LLMs with measurable storage capacity, internalization efficiency, and advantages in multi-module long-context reasoning.
Unveiling the Mechanisms of Multi-Hop Reasoning in Transformers via Identity Bridge
cs.LG 2025-09 conditional novelty 6.0 of 10

Adding identity supervision on bridge tokens enables out-of-distribution two-hop reasoning in simple transformers, with a nuclear-norm theory explaining the benefit.
Scalable Complexity Control Facilitates Reasoning Ability of LLMs
cs.LG 2025-05 conditional novelty 6.0 of 10

Controlling model complexity through smaller initialization rates and stronger weight decay improved LLM benchmark scores and made loss-versus-scale curves descend faster.

Reference graph

Works this paper leans on

60 extracted references · 22 canonical work pages · cited by 3 Pith papers

[1]

J., Javaheripi, M., Kauffmann, P., Lee, J

Abdin, M., Aneja, J., Behl, H., Bubeck, S., Eldan, R., Gunasekar, S., Harrison, M., Hewett, R. J., Javaheripi, M., Kauffmann, P., Lee, J. R., Lee, Y. T., Li, Y., Liu, W., Mendes, C. C. T., Nguyen, A., Price, E., de Rosa, G., Saarikivi, O., Salim, A., Shah, S., Wang, X., Ward, R., Wu, Y., Yu, D., Zhang, C., and Zhang, Y. Phi-4 technical report, 2024. URL h...

arXiv 2024
[2]

L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

arXiv 2023
[3]

S., Hu, W., Li, Z., Salakhutdinov, R

Arora, S., Du, S. S., Hu, W., Li, Z., Salakhutdinov, R. R., and Wang, R. On exact computation with an infinitely wide neural net. In Advances in Neural Information Processing Systems, pp.\ 8141--8150, 2019

work page 2019
[4]

Reflections after refereeing papers for nips

Breiman, L. Reflections after refereeing papers for nips. The Mathematics of Generalization, XX: 0 11--15, 1995

work page 1995
[5]

and Rathie, P

Caiado, C. and Rathie, P. Polynomial coefficients and distribution of the sum of discrete uniform variables. 01 2007

work page 2007
[6]

Chen, Z.-A., Li, Y., Luo, T., Zhou, Z., and Xu, Z.-Q. J. Phase diagram of initial condensation for two-layer neural networks. CSIAM Transactions on Applied Mathematics, 5 0 (3): 0 448--514, 2024. ISSN 2708-0579. doi:https://doi.org/10.4208/csiam-am.SO-2023-0016. URL http://global-sci.org/intro/article_detail/csiam-am/23306.html

work page doi:10.4208/csiam-am.so-2023-0016 2024
[7]

and Bach, F

Chizat, L. and Bach, F. On the Global Convergence of Gradient Descent for Over -parameterized Models using Optimal Transport . In Advances in Neural Information Processing Systems 31 , pp.\ 3036--3046. 2018

work page 2018
[8]

and Shanahan, M

Creswell, A. and Shanahan, M. Faithful reasoning using large language models. arXiv preprint arXiv:2208.14271, 2022

arXiv 2022

Show all 60 references

[9]

Selection-inference: Exploiting large language models for interpretable logical reasoning

Creswell, A., Shanahan, M., and Higgins, I. Selection-inference: Exploiting large language models for interpretable logical reasoning. arXiv preprint arXiv:2205.09712, 2022

2022 arXiv
[10]

The neural data router: Adaptive control flow in transformers improves systematic generalization

Csord \'a s, R., Irie, K., and Schmidhuber, J. The neural data router: Adaptive control flow in transformers improves systematic generalization. arXiv preprint arXiv:2110.07732, 2021

2021 arXiv
[11]

Ctl++: Evaluating generalization on never-seen compositional patterns of known functions, and compatibility of neural representations

Csord \'a s, R., Irie, K., and Schmidhuber, J. Ctl++: Evaluating generalization on never-seen compositional patterns of known functions, and compatibility of neural representations. arXiv preprint arXiv:2210.06350, 2022

2022 arXiv
[12]

L., Jiang, L., Lin, B

Dziri, N., Lu, X., Sclar, M., Li, X. L., Jiang, L., Lin, B. Y., Welleck, S., West, P., Bhagavatula, C., Le Bras, R., et al. Faith and fate: Limits of transformers on compositionality. Advances in Neural Information Processing Systems, 36, 2024

2024
[13]

A comparative analysis of optimization and generalization properties of two-layer neural network and random feature models under gradient descent dynamics

E, W., Ma, C., and Wu, L. A comparative analysis of optimization and generalization properties of two-layer neural network and random feature models under gradient descent dynamics. Sci. China Math., 63, 2020

2020
[14]

and Li, Y

Eldan, R. and Li, Y. Tinystories: How small can language models be and still speak coherent english?, 2023. URL https://arxiv.org/abs/2305.07759

2023 arXiv
[15]

How does gpt obtain its ability? tracing emergent abilities of language models to their sources

Fu, Y., Peng, H., and Khot, T. How does gpt obtain its ability? tracing emergent abilities of language models to their sources. Yao Fu’s Notion, 2022

2022
[16]

Delving deep into rectifiers: Surpassing human-level performance on imagenet classification

He, K., Zhang, X., Ren, S., and Sun, J. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pp.\ 1026--1034, 2015

2015
[17]

S., Perez, F., Ba, J., and Volkovs, M

Huang, X. S., Perez, F., Ba, J., and Volkovs, M. Improving transformer optimization through better initialization. In International Conference on Machine Learning, pp.\ 4475--4483. PMLR, 2020

2020
[18]

Learning compositionally through attentive guidance

Hupkes, D., Singh, A., Korrel, K., Kruszewski, G., and Bruni, E. Learning compositionally through attentive guidance. arXiv preprint arXiv:1805.09657, 2018

2018 arXiv
[19]

Neural Tangent Kernel : Convergence and Generalization in Neural Networks

Jacot, A., Gabriel, F., and Hongler, C. Neural Tangent Kernel : Convergence and Generalization in Neural Networks . In Advances in Neural Information Processing Systems 31 , pp.\ 8571--8580. 2018

2018
[20]

B., and M \"u ller, K

LeCun, Y., Bottou, L., Orr, G. B., and M \"u ller, K. R. Efficient BackProp, pp.\ 9--50. Springer Berlin Heidelberg, Berlin, Heidelberg, 1998. ISBN 978-3-540-49430-0. doi:10.1007/3-540-49430-8_2. URL https://doi.org/10.1007/3-540-49430-8_2

1998 doi
[21]

A., Serre, T., and Pavlick, E

Lepori, M. A., Serre, T., and Pavlick, E. Break it down: Evidence for structural compositionality in neural networks. arXiv preprint arXiv:2301.10884, 2023

2023 arXiv
[22]

Not all tokens are what you need for pretraining

Lin, Z., Gou, Z., Gong, Y., Liu, X., yelong shen, Xu, R., Lin, C., Yang, Y., Jiao, J., Duan, N., and Chen, W. Not all tokens are what you need for pretraining. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/for...

2024
[23]

Deepseek-v3 technical report

Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024

2024 arXiv
[24]

T., Goel, S., Krishnamurthy, A., and Zhang, C

Liu, B., Ash, J. T., Goel, S., Krishnamurthy, A., and Zhang, C. Transformers learn shortcuts to automata. arXiv preprint arXiv:2210.10749, 2022

2022 arXiv
[25]

Understanding the difficulty of training transformers

Liu, L., Liu, X., Gao, J., Chen, W., and Han, J. Understanding the difficulty of training transformers. arXiv preprint arXiv:2004.08249, 2020

2004 arXiv
[26]

J., Ma, Z., and Zhang, Y

Luo, T., Xu, Z.-Q. J., Ma, Z., and Zhang, Y. Phase diagram for two-layer relu neural networks at infinite-width limit. Journal of Machine Learning Research, 22 0 (71): 0 1--47, 2021

2021
[27]

Marcus, G. F. The algebraic mind: Integrating connectionism and cognitive science. MIT press, 2003

2003
[28]

A mean field view of the landscape of two-layer neural networks

Mei, S., Montanari, A., and Nguyen, P.-M. A mean field view of the landscape of two-layer neural networks. Proceedings of the National Academy of Sciences, 115 0 (33): 0 E7665--E7671, 2018. doi:10.1073/pnas.1806579115

2018 doi
[29]

S., Dick, R

Okawa, M., Lubana, E. S., Dick, R. P., and Tanaka, H. Compositional abilities emerge multiplicatively: Exploring diffusion models on a synthetic task. arXiv preprint arXiv:2310.09336, 2023

2023 arXiv
[30]

Language models are unsupervised multitask learners

Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. Language models are unsupervised multitask learners. 2019. URL https://api.semanticscholar.org/CorpusID:160025533

2019
[31]

P., Tanaka, H., and Lubana, E

Ramesh, R., Khona, M., Dick, R. P., Tanaka, H., and Lubana, E. S. How capable can a transformer become? a study on synthetic, interpretable tasks. arXiv preprint arXiv:2311.12997, 2023

2023 arXiv
[32]

and Vanden-Eijnden, E

Rotskoff, G. and Vanden-Eijnden, E. Parameters as interacting particles: long time convergence and asymptotic error scaling of neural networks. In Advances in Neural Information Processing Systems 31 , pp.\ 7146--7155. 2018

2018
[33]

and He, H

Saparov, A. and He, H. Language models are greedy reasoners: A systematic formal analysis of chain-of-thought. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=qFVVBzXxR2V

2023
[34]

and Spiliopoulos, K

Sirignano, J. and Spiliopoulos, K. Mean field analysis of neural networks: A central limit theorem. Stochastic Processes and their Applications, 130 0 (3): 0 1820--1852, 2020. doi:10.1016/j.spa.2019.06.003

2020 doi
[35]

Neurocompositional computing: From the central paradox of cognition to a new generation of ai systems

Smolensky, P., McCoy, R., Fernandez, R., Goldrick, M., and Gao, J. Neurocompositional computing: From the central paradox of cognition to a new generation of ai systems. AI Magazine, 43 0 (3): 0 308--322, 2022

2022
[36]

Srivastava, A., Rastogi, A., Rao, A., Shoeb, A. A. M., Abid, A., Fisch, A., Brown, A. R., Santoro, A., Gupta, A., Garriga-Alonso, A., et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615, 2022

2022 arXiv
[37]

and Kolter, J

Trockman, A. and Kolter, J. Z. Mimetic initialization of self-attention layers. In International Conference on Machine Learning, pp.\ 34456--34468. PMLR, 2023

2023
[38]

Deepnet: Scaling transformers to 1,000 layers

Wang, H., Ma, S., Dong, L., Huang, S., Zhang, D., and Wei, F. Deepnet: Scaling transformers to 1,000 layers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024 a

2024
[39]

Label words are anchors: An information flow perspective for understanding in-context learning

Wang, L., Li, L., Dai, D., Chen, D., Zhou, H., Meng, F., Zhou, J., and Sun, X. Label words are anchors: An information flow perspective for understanding in-context learning. arXiv preprint arXiv:2305.14160, 2023

2023 arXiv
[40]

Improving generalization and convergence by enhancing implicit regularization

Wang, M., He, H., Wang, J., Wang, Z., Huang, G., Xiong, F., Li, Z., Wu, L., et al. Improving generalization and convergence by enhancing implicit regularization. arXiv preprint arXiv:2405.20763, 2024 b

2024 arXiv
[41]

Wang, M. et al. Understanding the expressive power and mechanisms of transformer for sequence modeling. arXiv preprint arXiv:2402.00522, 2024 c

2024 arXiv
[42]

Wang, Z., Wang, Y., Zhang, Z., Zhou, Z., Jin, H., Hu, T., Sun, J., Li, Z., Zhang, Y., and Xu, Z.-Q. J. Towards understanding how transformer perform multi-step reasoning with matching operation. arXiv preprint arXiv:2405.15302, 2024 d

2024 arXiv
[43]

H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., and Fedus, W

Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., and Fedus, W. Emergent abilities of large language models. Transactions on Machine Learning Researc...

2022
[44]

Chain of thought prompting elicits reasoning in large language models

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Chi, E., Le, Q., and Zhou, D. Chain of thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2022 b

2022 arXiv
[45]

T., Panozzo, D., Zorin, D., and Bruna, J

Williams, F., Trager, M., Silva, C. T., Panozzo, D., Zorin, D., and Bruna, J. Gradient dynamics of shallow univariate relu networks. CoRR, abs/1906.07842, 2019. URL http://arxiv.org/abs/1906.07842

1906 arXiv
[46]

J., Zhang, Y., and Zhou, Z

Xu, Z.-Q. J., Zhang, Y., and Zhou, Z. An overview of condensation phenomenon in deep learning. arXiv preprint arXiv:2504.09484, 2025

2025 arXiv
[47]

Do vision-language pretrained models learn composable primitive concepts? arXiv preprint arXiv:2203.17271, 2022

Yun, T., Bhalla, U., Pavlick, E., and Sun, C. Do vision-language pretrained models learn composable primitive concepts? arXiv preprint arXiv:2203.17271, 2022

2022 arXiv
[48]

Improving deep transformer with depth-scaled initialization and merged attention

Zhang, B., Titov, I., and Sennrich, R. Improving deep transformer with depth-scaled initialization and merged attention. arXiv preprint arXiv:1908.11365, 2019 a

1908 arXiv
[49]

Understanding deep learning requires rethinking generalization

Zhang, C., Bengio, S., Hardt, M., Recht, B., and Vinyals, O. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016

2016 arXiv
[50]

J., Luo, T., and Ma, Z

Zhang, Y., Xu, Z.-Q. J., Luo, T., and Ma, Z. A type of generalization error induced by initialization in deep neural networks. arXiv:1905.07777 [cs, stat], 2019 b

1905 arXiv
[51]

Zhang, Y., Zhang, Z., Zhang, L., Bai, Z., Luo, T., and Xu, Z.-Q. J. Linear stability hypothesis and rank stratification for nonlinear models. arXiv preprint arXiv:2211.11623, 2022

2022 arXiv
[52]

and Xu, Z.-Q

Zhang, Z. and Xu, Z.-Q. J. Loss spike in training neural networks. arXiv preprint arXiv:2305.12133, 2023

2023 arXiv
[53]

and Xu, Z.-Q

Zhang, Z. and Xu, Z.-Q. J. Implicit regularization of dropout. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

2024
[54]

Zhang, Z., Li, Y., Luo, T., and Xu, Z.-Q. J. Stochastic modified equations and dynamics of dropout algorithm. arXiv preprint arXiv:2305.15850, 2023

2023 arXiv
[55]

Zhang, Z., Lin, P., Wang, Z., Zhang, Y., and Xu, Z.-Q. J. Initialization is critical to whether transformers fit composite functions by reasoning or memorizing. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 a

2024
[56]

Zhang, Z., Wang, Z., Yao, J., Zhou, Z., Li, X., E , W., and Xu, Z.-Q. J. Anchor function: a type of benchmark functions for studying language models. arXiv preprint arXiv:2401.08309, 2024 b

2024 arXiv
[57]

Zhang, Z., Lin, P., Wang, Z., Zhang, Y., and Xu, Z.-Q. J. Complexity control facilitates reasoning-based compositional generalization in transformers. arXiv preprint arXiv:2501.08537, 2025

2025 arXiv
[58]

Zhou, H., Zhou, Q., Jin, Z., Luo, T., Zhang, Y., and Xu, Z.-Q. J. Empirical phase diagram for three-layer neural networks with infinite width. Advances in Neural Information Processing Systems, 2022

2022
[59]

R., and Goldstein, T

Zhu, C., Ni, R., Xu, Z., Kong, K., Huang, W. R., and Goldstein, T. Gradinit: Learning to initialize neural networks for stable and efficient training. Advances in Neural Information Processing Systems, 34: 0 16410--16422, 2021

2021
[60]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

J., Javaheripi, M., Kauffmann, P., Lee, J

Abdin, M., Aneja, J., Behl, H., Bubeck, S., Eldan, R., Gunasekar, S., Harrison, M., Hewett, R. J., Javaheripi, M., Kauffmann, P., Lee, J. R., Lee, Y. T., Li, Y., Liu, W., Mendes, C. C. T., Nguyen, A., Price, E., de Rosa, G., Saarikivi, O., Salim, A., Shah, S., Wang, X., Ward, R., Wu, Y., Yu, D., Zhang, C., and Zhang, Y. Phi-4 technical report, 2024. URL h...

arXiv 2024

[2] [2]

L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

arXiv 2023

[3] [3]

S., Hu, W., Li, Z., Salakhutdinov, R

Arora, S., Du, S. S., Hu, W., Li, Z., Salakhutdinov, R. R., and Wang, R. On exact computation with an infinitely wide neural net. In Advances in Neural Information Processing Systems, pp.\ 8141--8150, 2019

work page 2019

[4] [4]

Reflections after refereeing papers for nips

Breiman, L. Reflections after refereeing papers for nips. The Mathematics of Generalization, XX: 0 11--15, 1995

work page 1995

[5] [5]

and Rathie, P

Caiado, C. and Rathie, P. Polynomial coefficients and distribution of the sum of discrete uniform variables. 01 2007

work page 2007

[6] [6]

Chen, Z.-A., Li, Y., Luo, T., Zhou, Z., and Xu, Z.-Q. J. Phase diagram of initial condensation for two-layer neural networks. CSIAM Transactions on Applied Mathematics, 5 0 (3): 0 448--514, 2024. ISSN 2708-0579. doi:https://doi.org/10.4208/csiam-am.SO-2023-0016. URL http://global-sci.org/intro/article_detail/csiam-am/23306.html

work page doi:10.4208/csiam-am.so-2023-0016 2024

[7] [7]

and Bach, F

Chizat, L. and Bach, F. On the Global Convergence of Gradient Descent for Over -parameterized Models using Optimal Transport . In Advances in Neural Information Processing Systems 31 , pp.\ 3036--3046. 2018

work page 2018

[8] [8]

and Shanahan, M

Creswell, A. and Shanahan, M. Faithful reasoning using large language models. arXiv preprint arXiv:2208.14271, 2022

arXiv 2022