acceptodds
Under review as a conference paper at ICLR 2027

LLMs Meet GNNs for Authorship Attribution: Metric Geometry and OT Calibration

Abstract

Authorship attribution has long relied on function words, which capture syntactic style independently of topic and anonymize the text. Word adjacency networks (WANs) model each text as a Markov chain over function words, and distances between them define a syntactic geometry over a corpus. We exploit this geometry with graph neural networks (GNNs), which process large language model (LLM) embeddings of the anonymized texts as signals over a graph built from WAN distances. GNNs are stable and transferable across geometric graphs, but these guarantees hold only if the graph itself is well-behaved. We therefore design the graph so that its distances form a valid metric, remain stable under small edits, and keep distinct styles separated, and so that its edge weights make informative edges stand out. We characterize how candidate metrics respond to small perturbations, trace a token-level edit through every stage of the pipeline to prove stability and discriminability, and introduce an optional calibration step that reshapes the edge weights through a monotone, bi-Lipschitz optimal transport map. On Early Modern English plays, the calibrated GNN reaches 73.2% accuracy, outperforming an LLM classifier at 70.0% and an uncalibrated GNN at 70.5%, and calibration reduces the accuracy drop under WAN perturbations from 7.3% to 0.3%.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
authorship attribution, stylometry, metric geometry, graph calibration, word adjacency networks, graph neural networks, large language models, optimal transport

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1BgTOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper develops a model for authorship attribution that combines the textual representation of LLMs with that of Word Adjacency Networks. Each text sample is patched to retain only function words and then used to train a LLM that produces a feature embedding and to construct a Word Adjacency Network that models transitions between these function words as a Markov Chain. Then, the stationary distributions of these WANs are compared across text samples to obtain pairwise distances between them. The resulting weighted graph over the collection of texts is is optionally calibrated and used as an input to a GNN that performs author attribution, additionally taking the LLM embeddings as text features. The theoretical contributions focus on the properties of this learning algorithm and can be summarised as a perturbation analysis that shows that the learning procedure remains stable to small textual perturbations while preserving useful stylistic information to distinguish the authors. The method is evaluated on a single dataset of Early Modern English playwrights, achieving marginally better accuracy compared to the baselines, and showing substantially increased robustness to small perturbations of the WAN matrix.

Critical strengths

None.

Critical weaknesses
  1. Neither in the Introduction nor in the Related Work section in Appendix A does the paper state the research gap it is seeking to address. The literature review merely collects the relevant context without identifying the strengths or weaknesses of any existing methodology or established baselines in the filed. The Introduction and the methodology (sections 2-6) are instead a listing of the paper's sequence of methodological choices and results and do not provide a clear contextual motivation for them. Similarly, the evaluation does not consider any established baselines in the field and instead focuses on baselines specifically created by the authors.

  2. The evaluation is insufficient to establish the properties of and empirical claims on the learning algorithm proved in the paper. The empirical findings are narrow and primarily focus on the predictive comparison between the model and selected baselines on authorship attribution and location and time period classification, but testing only on one dataset is insufficient establish that the accuracy improvement of the proposed method generalizes to a sufficient number of other datasets. Robustness to perturbations (Theorems 3.1, 5.1) is tested by perturbing the WAN transition matrix rather than proving the robustness of the whole learning pipeline from direct textual perturbation to GNN output. The claim that distinct syntactic styles remain separated (Theorems 3.2, 5.2) is not empirically verified with a dedicated experiment. That OT calibration produces informative edges is replaced by a descriptive analysis of the graph geometry.

Overall, this paper looks a lot more like a listing of nice theoretical properties of a new learning algorithm than a method whose necessity and usefulness is truly motivated by the existing literature and the experiments. The paper does not engage with the existing literature, does not identify a research gap, and does not provide a comprehensive and logically consistent evaluation procedure. Therefore, in my opinion, it does not meet the minimum standards for acceptance.

Questions for authors
  1. What is the research gap? This should be clarified in the Introduction.
  2. What are other established baselines in the field and why should I choose your method instead of them? This should be established both in the Introduction and in the methodology, and evaluated using additional baselines and benchmarks in the Experiments.
  3. Can you close the logical gaps between the theoretical results and their experimental validation? This should be fixed in the Experiments.
6 comments

We thank the reviewer for the review, for finding the presentation clear, and for recognizing that the paper establishes a listing of nice theoretical properties. We take the concerns about motivation and evaluation seriously and answer each point below, with additional experiments that validate our theoretical findings and a clearer account of the settings in which our method is a competitive choice.

First, a factual clarification: the LLM is not trained in our pipeline. We use a frozen pretrained encoder (all-MiniLM-L6-v2, Appendix D.1.6) to embed the masked texts, and only the GNN/MLP layers are trained. We will make this more visible in §2.

Experimental setting. Unless otherwise stated, all new experiments in this response use the default setting of the paper (Section 7 and Appendix D): Hellinger distance, k=3k = 3, α=β=1\alpha = \beta = 1, frozen all-MiniLM-L6-v2 embeddings, and the same GCN and training protocol. Results are mean ± std over 10 seeds.

W1 & Q1. What is the research gap? We agree, and we believe the root cause is the framing. The current version makes authorship attribution look like the main focus, whereas our main contribution is technical: studying the design of metrics and edge weights for geometry-aware graph representations built from estimated distributions, with guarantees that can be replicated to apply to a much wider range of graph tasks. We will revise the abstract, introduction, and conclusion to reflect this explicitly and to treat authorship attribution as a case study. Authorship attribution is itself a broad field with diverse subproblems, including stylometric representation learning, detection of AI-generated text, plagiarism detection, tracing how literary style evolves over time, and dialect classification; our time-period and location tasks are instances of the last two. Specifically:

  • Abstract. It now opens with the general problem: when each node of a graph carries a distribution estimated from finite data, the behavior of the model depends on the geometry used to compare these distributions, yet how this choice propagates to the learned model is rarely analyzed. Authorship attribution is then introduced as the setting in which we study this question, and the gap is stated directly: classical WAN methods use relative entropy, which violates (O1) and (O2), while LLM embeddings represent each text in isolation.
  • Introduction. A new opening paragraph describes graph-based learning on estimated distributions in general: the metric, kernel, and sparsification are usually chosen by convention; GNN guarantees take the graph as given; and no prior work traces how a small change in the data propagates to the GNN output or uses this to guide those choices. It ends by introducing authorship attribution as a natural instance of this setting. A new research-gap paragraph follows the objectives (O1)–(O4), and contribution (C1) now states that our analysis turns the choice of metric into a principled choice of geometry.
  • Conclusion. The opening now presents the work as improving authorship attribution by improving the geometry in which texts are compared, an instance of geometry-aware representation learning in which the candidate metrics induce different geometries on the probability simplex. A new Broader Applicability paragraph explains that the LLM embeddings can be replaced by representations on other manifolds, with the same analysis guiding the choice of metric, and lists the related tasks above.

The reviewer also notes that Sections 2–6 read as a sequence of methodological choices without contextual motivation. We will state explicitly that each design choice in Sections 2–6 is motivated by the objective it serves: the metric by (O1)–(O2), the kernel by (O2)–(O3), the k-NN margin by (O2), and calibration by (O4), as stated in the contributions (C1)–(C3).

AuthorsOpenReview ↗

Rebuttal Continued

W1 & Q1 Continued.

The research gap, to be stated explicitly in the introduction, has two sides.

Technical gap (geometry-aware graph representations).

  1. Graph construction from estimated distributions is unanalyzed. GNNs are stable to graph perturbations (Gama et al. 2020) and transferable across graphs (Ruiz et al. 2020a), but these guarantees take the graph as given. When the graph is built from distributions estimated from finite, noisy data, as in graph-based learning on histograms or statistical manifolds (Lafferty & Lebanon 2005; Belkin et al. 2006), no prior work traces how estimation noise propagates through the distribution estimates, the metric, the edge weights, and the neighbor selection to the GNN output.
  2. The choice of metric is made by convention. Graph construction choices are known to change downstream results substantially (Maier, von Luxburg & Hein 2009; von Luxburg 2007), yet it is not known which divergences keep the chain above well-conditioned. We show that candidate metrics fall into three sensitivity families with distinct behavior under perturbation, which turns the choice of metric into a principled choice of geometry.
  3. Edge-weight distributions are uncontrolled. Graph-learning methods control the sparsity of a learned graph (Kalofolias 2016), but fixing a metric and a kernel gives no control over the edge-weight distribution of a graph built from a different source. We propose a rank-preserving OT calibration (Peyré & Cuturi 2019) and characterize when it preserves stability and discriminability.

Gap in existing authorship attribution methods.

  1. Frequency-based stylometry relies on function-word frequencies, from the Federalist Papers study (Mosteller & Wallace 1963) to distance measures such as Burrows' Delta (Burrows 2002), as surveyed by Stamatatos (2009), Juola (2008), and Koppel et al. (2009). These features are topic-independent but discard the order in which function words occur, and attribution uses fixed distance rules rather than learned representations.
  2. WAN-based methods (Segarra et al. 2015; 2016) capture this order by modeling each text as a Markov chain over function words, but they assign a text to the closest author profile under relative entropy. Relative entropy is asymmetric and violates the triangle inequality, so it does not define a metric graph (O1); it is unbounded when a profile assigns zero probability to an observed transition and weights perturbations of rare transitions by their inverse probability, so it is not stable under small edits (O2). Each text is compared only with author profiles, so relations among individual texts are not used.
  3. Neural and LLM-based methods, including sentence embeddings (Reimers & Gurevych 2019) and learned authorship or style representations (Rivera-Soto et al. 2021; Wegmann et al. 2022), learn rich representations but embed each text in isolation, so a classifier on these embeddings cannot use relations among texts in a corpus. They also come without guarantees on how small textual perturbations affect predictions, to which text classifiers are known to be sensitive (Pruthi et al. 2019).
  4. No existing method combines these strengths: a learned representation, cross-text structure from a valid and stable metric, and anonymization to function words, together with guarantees on how perturbations of the text propagate to the prediction.

We will also rewrite the related work in Appendix A so that each line of prior work ends with its specific limitation and the part of our framework that addresses it, rather than only describing it.

AuthorsOpenReview ↗

Rebuttal Continued

W1 & Q2. Established baselines and why to choose this method. We agree that the baselines were insufficient, and we added the following. All results are on play-disjoint splits, where every chunk of a play falls in a single partition, so that no test chunk has a training chunk from its own play.

Baselines. The original WAN method (Segarra et al. 2015) builds one WAN profile per author and attributes each chunk to the closest profile under relative entropy; it uses no LLM features, no learning and no graph. GNNs over WANs (Ruiz et al. 2020b) treat function words as nodes and each chunk as a signal on an author's WAN. The linear classifier on the WAN stationary distribution takes the 30-dimensional π^\hat\pi of each chunk's WAN as input, and the LLM-MLP uses the LLM embeddings without a graph.

BaselineTest accuracy (%)
Linear classifier on WAN stationary distribution π^\hat\pi58.2 ± 5.7
GNN over WANs (Ruiz et al. 2020b)59.4 ± 7.1
LLM-MLP (no graph)65.7 ± 5.9
Original WAN (Segarra et al. 2015)68.9 ± 3.2
OursRaw weights (%)OT weights (%)
LLM-GCN67.5 ± 5.370.2 ± 4.9
LLM-GCN, inductive66.9 ± 5.370.4 ± 5.2
LLM-GCN with π^\hat\pi concatenated to the node features69.5 ± 4.672.4 ± 5.2
LLM-GIN66.3 ± 6.170.0 ± 5.9
LLM-GPRGNN72.4 ± 5.869.8 ± 5.0
LLM-SAGE71.2 ± 5.169.6 ± 3.6
LLM-GraphConv68.4 ± 4.567.2 ± 5.4
LLM-GraphGPS74.3 ± 3.572.6 ± 3.8
LLM-GT72.1 ± 3.572.7 ± 3.7
LLM-NAGphormer70.3 ± 5.071.4 ± 3.8
LLM-VCR-Graphormer72.6 ± 4.372.9 ± 6.5

Why choose this method. LLM-GCN-OT outperforms GNNs over WANs by 10.8 points and the LLM-MLP by 4.5 points. With the same WAN statistics as input, it exceeds the original WAN method (72.4% vs. 68.9%, higher in 7/10 seeds, p=0.06p = 0.06). With identical inputs and parameter count, the OT-calibrated graph adds a significant gain over an MLP on concatenated [LLM ⊕ π^\hat\pi] features (72.4% vs. 69.8%, p=0.004p = 0.004), so the cross-text structure contributes beyond the features. The gain holds when test texts are unseen during training (inductive setting, 70.4%), across five GNN architectures (GCN, GIN, GPRGNN, SAGE and GraphConv) and four graph transformers (GraphGPS, GT, NAGphormer and VCR-Graphormer), all of which outperform the LLM-MLP and, with OT weights, GNNs over WANs (by 7.8–13.5 points, p≤0.04p \le 0.04), with larger frozen encoders (Reviewer 2jLP, Q4), and on CCAT50 (W2 below). Three further results support choosing this method:

  • Robustness to text edits (see our response to Reviewer 5xPf, Q2): with 300 function words replaced per chunk and no retraining, LLM-GCN-OT loses only 0.9 points on unseen plays, against 5.3 for LLM-GCN, so the calibrated model stays reliable when texts are noisy or edited.
  • Robustness to graph construction (see our response to Reviewer 5xPf, Q3): as the k-NN graph becomes denser, raw weights collapse for some architectures (down to 29.9%), while with OT every architecture stays between 69.4% and 72.5%, so the method does not depend on a carefully tuned kk.
  • Robustness to WAN perturbations (Figure 2; W2 & Q3 below): with a model trained once on the clean graph, accuracy drops by at most 0.9 points with OT across the five metrics, against 3.7–6.5 points with raw weights, so the stability guarantees of the calibrated pipeline hold in practice.

We will state the intended use case explicitly. Our method targets settings where texts must be anonymized to function words and where cross-text structure is available. It is not meant as a replacement for fine-tuned attribution models on unrestricted text.

We note that fine-tuned attribution models are not competitors to our framework but possible components of it: the LLM embeddings are the node features of our graph, so any encoder, frozen or fine-tuned, can be used in Stage 1, and the WAN graph adds cross-text structure on top of it. Our larger-encoder experiments support this: with OT calibration, the graph improves over the encoder alone for MiniLM and mpnet (see our response to Reviewer 2jLP, Q4). Moreover, a model fine-tuned on unrestricted text sees the topical content that our anonymized setting deliberately removes, so it would not be evaluated under the same conditions. We will discuss these approaches and where each falls short in the introduction (see W1 & Q1), and add these baselines and CCAT50 to the experiments section of the paper.

Reviewer_2jLPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a technique to determine authorship attribution in a standard transductive graph node label prediction setup. Each text is first transferred into a word adjacency network (WAN) over a predefined set of function words, with edge weights derived from word co-occurences. This WAN, with outflow normalized edge weights, defines a Markov chain whose stationary distribution π\pi represents the text . Then these π\pi per text are connected in a kNN graph based on a divergence measure, with node features derived from LLM embeddings. A GCN run over this graph helps predict the node labels (aka authors) in a transductive node label prediction setup. This paper also proposes an optional OT calibration step to make the kNN graph edge weights distribution aligned to another learned sparse graph derived from LLM embeddings. The authors argue the theoretical merits of their setup and study how token level edits in the texts propagate downstream to the stationary distributions, edge weights, and the GCN outputs, in terms of stability and discriminability.

Critical strengths

The paper focuses adequately on the graph construction process and is motivated carefully from the goal task. Although there is some vague phrasing like "well-behaved" graphs which need clarification, the overall desiderata are carefully listed out and evaluated in theory. I appreciate the step by step description in Figure 1.

The robustness experiments are convincing, especially Figure 2 where the effect of OT calibration can be clearly seen.

Critical weaknesses

Honestly I am not very excited about the problem statement itself. Perhaps this is of more interest to the NLP community, which I myself am not a part of. While the authors present the design challenges in a systematic way, I don't see how authorship attribution on a few hundred chunks from playwright texts is interesting to the ICLR community. I will factor in the feedback from the other reviewers in this regard, and also perhaps the authors can explain more on why the contributions of this paper are important and if they can be useful in a more broader range of applications.

The paper promises in (C2) to "trace a token-level edit through every stage of the pipeline and prove stability and discriminability at each stage". But I don't see how, as stated, the per stage bounds can be chained together in a single end-to-end stability claim. Particularly, in the kNN sparsification step a tiny change in π\pi can change neighbors in a discontinuous manner and this affects the adjacency information that Theorem 5.1 depends on. We cannot assume the edge set stays fixed under perturbation. There is a similar issue in Theorem 6.2 which assumes "pμ,pνp_\mu, p_\nu are strictly positive and continuous" but the map is built from empirical CDFs which are step functions that will change under perturbation.

Questions for authors
  • What happens in an inductive setting ? In my understanding currently it is highly likely that a test chunk is from the same play as another chunk in the train set and the two chunks are likely neighbors of each other . This would make the challenge much easier for graph algorithms based on neighborhood voting. What happens in an inductive setting where there are held out plays (by the same authors) but no chunks from the held out play in the training set?

  • What does it mean for a graph to be “well-behaved” ?

  • In light of the weakness mentioned above, how can one defend the end-to-end stability claim? Particularly, following the potential discontinuity in the kNN sparsification, I don’t see how smoothness of κ\kappa makes the graph as a whole stable.

  • The method is currently not compared against any earlier authorship attribution baselines? One could compare against the original WAN method, classical stylometry authorship attribution methods, learned authorship embeddings, bigger /fine-tuned LLMs instead of small frozen MiniLM.

  • Also since this is based on natural language text, I imagine real-world perturbation effects are more interesting such as edits/ typos etc, rather than the current artificial perturbation.

  • Lastly to show broader applicability, a more detailed discussion and/or more datasets is perhaps helpful.

  • There are a couple of broken references.

7 comments

We thank the reviewer for the constructive review, for the positive assessment of the graph-construction design and the robustness experiments, and for the precise questions. We answer each point below.

Experimental setting. Unless otherwise stated, all new experiments in this response use the default setting of the paper (Section 7 and Appendix D): Hellinger distance, k=3k = 3, α=β=1\alpha = \beta = 1, frozen all-MiniLM-L6-v2 embeddings, and the same GCN and training protocol. Results are mean ± std over 10 seeds.

W1. Relevance to ICLR. We appreciate the reviewer's candor. Our core contribution is not specific to plays. It concerns a setting that is common in representation learning: each node carries an empirical distribution or Markov chain estimated from finite data, and the graph is built from distances between these estimates. GNN stability and transferability guarantees (Gama et al. 2020; Ruiz et al. 2020a) assume the graph is given. To our knowledge, prior work does not trace estimation noise in the raw data through the distribution estimates and the metric to the graph operator and the GNN output, nor study which divergences make this chain well-conditioned.

This paper studies how to improve authorship attribution by improving the geometry in which texts are compared, an instance of geometry-aware representation learning, where data are modeled on Riemannian manifolds rather than in Euclidean space (e.g., hyperbolic embeddings, Nickel & Kiela 2017). The probability simplex with the Fisher information metric is a Riemannian manifold (geodesic distance: Fisher–Rao; chordal distance: Hellinger), and Theorem 4.1 lets us understand the choice of metric rather than tune it: the divergences fall into Fisher-information, ℓ1\ell_1, and sub-linear sensitivity families, a split our experiments confirm (Table 2), with the Fisher-information family about 2.6 times more sensitive to rare-word perturbations. How to choose the geometry of a representation space, and how that choice propagates to a learned model, is of general interest to the ICLR community.

The same construction applies to other sequence data summarized by transition statistics, for example user sessions or clickstreams, code stylometry, and k-mer transition profiles. Within text, the same machinery serves several problems closely related to authorship classification: learning representations of writing style (stylometry); detecting AI-generated text, where function-word transition statistics give a topic-independent signal of interest to the LLM community; plagiarism detection, where shifts in style within or across documents are the cue; tracing how literary style evolves over time, which our time-period classification task already addresses (Section 7); and dialect classification, i.e., distinguishing dialects of the same language, of which our location classification task is one example (Section 7). The divergence-sensitivity hierarchy (Theorem 4.1) and the rank-preserving OT calibration of edge weights are likewise independent of the application. Although we build our theory and experiments around LLM embeddings and authorship attribution, broader applicability can be supported by replacing the LLM-generated embeddings with other representations living on arbitrary manifolds. An identical study of the implications of the choice of metric then applies to study which metric suits a given task.

AuthorsOpenReview ↗

Rebuttal Continued

W2 & Q3. End-to-end stability through k-NN sparsification, and Theorem 6.2 with empirical CDFs. We agree with both points and will correct the claims.

k-NN. The chain requires the edge set to stay fixed, which the current text implicitly assumes. We will add the following lemma to the paper (new Section 4.4, kk-NN Sparsification, with the proof in the appendix). It also addresses a related point raised by another reviewer.

Lemma (edge-set invariance). Let ηj=d(πj,π~j)\eta_j = d(\pi_j,\tilde\pi_j) and ηmax⁡=max⁡jηj\eta_{\max} = \max_j \eta_j, and let the kernel κ\kappa be LκL_\kappa-Lipschitz on the range of distances. For each node ii, let gi=di(k+1)−di(k)g_i = d_i^{(k+1)} - d_i^{(k)} be the gap between the distances to its (k+1)(k{+}1)-th and kk-th nearest neighbors. If gi>0g_i > 0 for all ii and ηmax⁡<14min⁡igi\eta_{\max} < \tfrac14 \min_i g_i, then (i) the k-NN neighbor sets, and hence the symmetrized edge set, are unchanged; and (ii) ∣W~ij−Wij∣≤2Lκηmax⁡|\tilde W_{ij} - W_{ij}| \le 2L_\kappa \eta_{\max} for all i,ji,j, so that ∥W~−W∥F≤2Lκηmax⁡∣E∣\Vert\tilde W - W\Vert_F \le 2L_\kappa \eta_{\max}\sqrt{|E|}, where ∣E∣|E| is the number of nonzero entries of WW.

Proof sketch. By Eq. (3), every pairwise distance moves by at most 2ηmax⁡2\eta_{\max}, so a margin larger than 4ηmax⁡4\eta_{\max} rules out any neighbor swap, and the Lipschitz kernel then gives (ii). Because neighbors are selected under dd and the calibration map is monotone, calibration leaves the edge set unchanged, and (ii) holds with LκL_\kappa replaced by MLκML_\kappa for a fixed map. Theorem 3.1 then gives a sufficient condition on the token edit itself. This also answers how we defend the end-to-end stability claim: we will no longer claim unconditional end-to-end stability. Under the neighbor-margin condition, the chain token edit → π\pi → distances → kernel weights → (calibration) → k-NN graph → GNN output holds with explicit constants. Outside that regime we make no stability claim.

Regarding the role of κ\kappa: the kernel is designed only so that pairs at shorter distance receive stronger edges in the graph. Since κ\kappa is decreasing, it preserves the ranking of pairs by distance, so the k-NN neighbor sets are the same as those selected under dd, and smoothness of κ\kappa alone cannot make the edge set stable. Whether the edge set survives a perturbation depends only on the neighbor margin in dd, as in the lemma above; once the edge set is fixed, the Lipschitz constant LκL_\kappa bounds how much the edge weights change.

Robustness of neighbor sets. We ran an experiment to measure how robust the k-NN neighbor sets are to text edits. On the paper's graph (k=3k = 3, Hellinger, 773 nodes), we replaced mm function words in a chunk, rebuilt the WANs and the k-NN graph, and counted the nodes whose neighbor set changes. We considered two conditions: a single chunk is edited (each of the 773 chunks in turn), or every chunk is edited at once. No model is trained, so the result does not depend on the split.

Margin distribution. We also computed the neighbor margins gig_i. Since we are unable to add the figure to the rebuttal, we describe it. The median margin is 1.4×10−31.4\times10^{-3}; most margins lie between 10−410^{-4} and 10−210^{-2}, with a peak just above the median and a tail toward smaller values. The 34 nodes with gi=0g_i = 0 correspond to duplicate texts, whose kk-th and (k+1)(k{+}1)-th neighbors tie exactly. We will include the figure of this distribution in the revised paper.

Neighbor-set changes. When a single chunk is edited, the fraction of nodes whose neighbor set changes is between 0.06% and 0.76% across all edit levels. When every chunk is edited at once, it grows with the number of replacements per chunk, from 25.0% at m=1m = 1 to 99.8% at m=300m = 300.

Replacements per edited chunk mm131030100300
One chunk edited0.06%0.10%0.19%0.32%0.57%0.76%
Every chunk edited25.0%41.5%64.3%83.8%97.0%99.8%

A single edited text changes the neighbor sets of 0.4 nodes on average after one replacement and of at most 6 nodes after 300, and in 81% of single-replacement edits no other node is affected. Edits therefore change the edge set only locally, around the edited text, as the node-wise form of the lemma in the appendix describes (2ηi+ηˉin+ηˉout<gi2\eta_i + \bar\eta_{in} + \bar\eta_{out} < g_i). When neighbor sets do change, the swapped neighbors are nearly equidistant, so an edge is replaced by one of almost equal weight; this is consistent with the robustness of the calibrated model under text edits (see Q5).

AuthorsOpenReview ↗

Rebuttal Continued

W2 & Q3 Continued.

Empirical CDFs. The reviewer is right. Both the WAN weight distribution μ\mu and the reference distribution νactive\nu_{\mathrm{active}} are empirical, so their CDFs are step functions and Fν−1∘FμF_\nu^{-1}\circ F_\mu is not bi-Lipschitz as literally written. We will restate Theorem 6.2 for the finite-sample map:

Theorem 6.2 (revised). Let c(1)<⋯<c(n)c_{(1)} < \dots < c_{(n)} be the distinct sorted WAN weights and v(1)<⋯<v(n′)v_{(1)} < \dots < v_{(n')} the distinct sorted active entries of W∗W^{\ast}. Let Q^ν\hat Q_\nu be the piecewise-linear interpolant through (j−1n′−1,v(j))\big(\tfrac{j-1}{n'-1}, v_{(j)}\big), set w(k)=Q^ν(k−1n−1)w_{(k)} = \hat Q_\nu\big(\tfrac{k-1}{n-1}\big), and let ff be the piecewise-linear interpolant through (c(k),w(k))(c_{(k)}, w_{(k)}), with slopes sk=(w(k+1)−w(k))/(c(k+1)−c(k))s_k = (w_{(k+1)} - w_{(k)})/(c_{(k+1)} - c_{(k)}). Then ff is (a) strictly monotone (rank-preserving); (b) piecewise linear, with f′=skf' = s_k on (c(k),c(k+1))(c_{(k)}, c_{(k+1)}); and (c) bi-Lipschitz on [c(1),c(n)][c_{(1)}, c_{(n)}], with m=min⁡ksk>0m = \min_k s_k > 0 and M=max⁡ksk<∞M = \max_k s_k < \infty.

Proof sketch. Both (c(k))(c_{(k)}) and (w(k))(w_{(k)}) are strictly increasing, the latter because Q^ν\hat Q_\nu is. Every slope sks_k therefore lies in (0,∞)(0, \infty). The map ff is absolutely continuous with f′∈{sk}f' \in \lbrace s_k \rbrace almost everywhere, so f(c2)−f(c1)=∫c1c2f′∈[m(c2−c1),M(c2−c1)]f(c_2) - f(c_1) = \int_{c_1}^{c_2} f' \in [m(c_2 - c_1), M(c_2 - c_1)], which is (c). Property (a) follows from m>0m > 0, and (b) holds by construction. The proof no longer requires continuous densities. The original density ratio pμ/pνp_\mu/p_\nu survives only as an interpretation: sks_k approximates it as the number of weights grows. When the data are perturbed, the map is re-estimated. Because each WAN weight is mapped to the target quantile at its rank, and the target ν\nu comes from the LLM embeddings, which do not change under WAN perturbations, a calibrated weight changes only when a perturbation reorders the WAN weights.

Effect on other results. Only Theorem 6.2 and the localized bounds of Appendix C change. The latter become minimum and maximum slopes over the knot intervals in each regime, m(C)=min⁡{sk:c(k+1)>C}m(\mathcal{C}) = \min\lbrace s_k : c_{(k+1)} > \mathcal{C} \rbrace and M(C)=max⁡{sk:c(k)<C}M(\mathcal{C}) = \max\lbrace s_k : c_{(k)} < \mathcal{C} \rbrace, and the argument there is unchanged. All other results are unaffected. Sections 3 and 4, Theorems 5.1 and 5.2, and the edge-set lemma do not involve the transport map.

Q1. Inductive setting with held-out plays. This concern is well-founded. Under the random node-level split, chunks of the same play can appear in train and test and are likely to be graph neighbors, which favors neighborhood voting. We ran (i) play-disjoint splits, where all chunks of a play fall in a single partition (146 play groups, stratified by author), and (ii) an inductive setting on these splits, in which every edge incident to a test node is removed during training and model selection and restored only for the final test. In this setting, test texts are unseen during training: no information passes to or from them during training and model selection, and they join the graph only at test time, as new documents would.

MethodPlay-disjoint (%)
LLM-MLP65.7 ± 5.9
LLM-GCN67.8 ± 5.2
LLM-GCN-OT70.2 ± 4.9
LLM-GCN, inductive66.9 ± 5.3
LLM-GCN-OT, inductive70.4 ± 5.2

On unseen plays, LLM-GCN-OT remains the most accurate model, and the inductive variants stay within 1 point of their transductive counterparts. The advantage of the graph therefore does not come from same-play neighbors or from test-node edges seen during training. We will add these results to Table 1.

Q2. What does "well-behaved" mean? We use it as shorthand for objectives (O1)–(O4) in §1. The distances form a valid metric (O1), are stable under small edits (O2), keep distinct styles separated (O3), and yield an edge-weight distribution in which informative edges stand out (O4). We will replace the informal phrase with an explicit definition that refers to (O1)–(O4).

Reviewer_5xPfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for authorship attribution that integrates pretrained language-model embeddings with the syntactic structure captured by Word Adjacency Networks (WANs). Each anonymized text segment is represented by a language-model embedding and a stationary distribution derived from its function-word transition matrix. Pairwise distances between stationary distributions define a sparse k-nearest-neighbor graph, over which a GNN performs authorship classification. The study also gave theoretical results related to stationary-distribution perturbations, metric sensitivity, GNN stability and discriminability, and OT calibration.

Critical strengths
  • The fusion of WAN-derived stylistic structure with pretrained text embeddings is an interesting approach to exploiting complementary syntactic and contextual information.

  • Theoretical analysis of stationary distributions, metric sensitivity, graph weighting, and GNN outputs gives a useful framework for studying perturbation behavior. The local sensitivity analysis in Theorem 4.1 and its supporting results in Table 2 are informative.

  • Table 1 reports statistically significant improvements in authorship classification from OT calibration, while the auxiliary tasks provide additional evidence that graph representations can improve authorship classification.

  • The appendices have extensive mathematical derivations and implementation details that help to examine the proposed approach.

Critical weaknesses
  • Theorem 5.2 (Section 5, p. 6; Appendix B.9, p. 21) claims that, assuming a strictly positive activation derivative, this lower bound is not generally valid under the stated matrix-dimensional assumptions. In the authorship experiment, (HΘH\Theta) maps 773 nodes to six output dimensions. Consequently, there exist nonzero perturbations (Δ\Delta W) satisfying (Δ\Delta W H$\Theta=0$), even when the individual matrices have positive smallest singular values under the conventional rectangular-matrix definition. Such perturbations produce identical outputs even with nonzero changes in the graph operator. Therefore, Theorem 5.2 needs explicit injectivity or restricted-subspace assumptions, or its discriminability claim must be weakened.

  • Equation (3) and Section 4.3 bound perturbations of pairwise distances and kernel weights, but the graph is constructed through discrete k-nearest-neighbor selection. Since small distance perturbations can change neighbor rankings and replace edges (No minimum neighbor-separation condition), the analysis does not fully connect text perturbations to perturbations of the actual graph operator used in Theorem 5.1.

  • Figure 3 describes a reference graph containing 94% zero entries, whereas Appendix D.1.3 states that the calibrated and uncalibrated models retain the same symmetrized WAN k-NN edge set. This means the method transfers an active-weight distribution rather than the reference graph's sparsity pattern or connectivity. The paper should distinguish these properties more precisely.

  • A play-disjoint evaluation, assigning all chunks of each play to a single partition, is necessary to evaluate generalization to unseen works by the same authors (independently sampled graphs or unseen documents).

  • There is a need for an evaluation of an MLP or other classifier combining language-model embeddings and stationary WAN features without constructing a graph. There is also a need to evaluate a GCN built using an embedding-space similarity graph instead of WAN-derived edges, to distinguish the benefit of syntactic geometry from graph propagation itself.

  • The conclusion concerning the full active-weight distribution appears stronger than an assumption restricting distance-density monotonicity to a neighborhood of zero. The applicable domain should be specified.

  • The implementation uses frozen all-MiniLM-L6-v2 embeddings. The scope of the conclusions should reflect this encoder rather than implying validation across general large generative language models.

Questions for authors
  • Can the authors provide a corrected discriminability theorem with explicit injectivity or restricted-subspace assumptions for both graph and feature perturbations?

  • What are the results under play-disjoint splitting, and how much of the Figure 2 robustness improvement persists when model parameters are not retrained?

  • How does the proposed OT approach compare with embedding-derived k-NN graphs and simpler monotone edge-weight transformations?

4 comments

We thank the reviewer for a detailed and technically precise review. The counterexample to Theorem 5.2 and the observation about k-NN neighbor swaps are both correct, and both lead to concrete corrections. We answer each point below.

Experimental setting. Unless otherwise stated, all new experiments in this response use the default setting of the paper (Section 7 and Appendix D): Hellinger distance, k=3k = 3, α=β=1\alpha = \beta = 1, frozen all-MiniLM-L6-v2 embeddings, and the same GCN and training protocol. Results are mean ± std over 10 seeds.

W1 & Q1. Theorem 5.2 is not valid for rectangular matrices. We agree. The proof (Appendix B.9, which is also mislabeled "Proof of Theorem 5.1") uses ∥AB∥F≥σmin⁡(B)∥A∥F\Vert AB\Vert_F \ge \sigma_{\min}(B)\Vert A\Vert_F for right multiplication. This inequality requires BB to have full row rank. Here H∈RN×dH \in \mathbb{R}^{N\times d} and Θ∈Rd×C\Theta \in \mathbb{R}^{d\times C} with C≪d<NC \ll d < N, so HΘH\Theta has rank at most CC. Any ΔW\Delta_W whose rows lie in the left null space of HΘH\Theta leaves the output unchanged, exactly as the reviewer describes. We will replace the theorem with the following restricted-subspace statement, which follows directly from the same argument:

Theorem 5.2 (revised). Let σ\sigma be absolutely continuous with σ′≥γmin⁡>0\sigma' \ge \gamma_{\min} > 0 almost everywhere on the relevant domain, let r=rank(HΘ)r = \mathrm{rank}(H\Theta), and let ΠHΘ\Pi_{H\Theta} be the orthogonal projector onto col(HΘ)\mathrm{col}(H\Theta). Then ∥Z~−Z∥F≥γmin⁡σr(HΘ)∥ΔWΠHΘ∥F\Vert\tilde Z - Z\Vert_F \ge \gamma_{\min} \sigma_r(H\Theta) \Vert\Delta_W \Pi_{H\Theta}\Vert_F. Similarly, if WW is nonsingular and ΠΘ\Pi_\Theta is the projector onto col(Θ)\mathrm{col}(\Theta), then ∥Z~−Z∥F≥γmin⁡σN(W)σr′(Θ)∥ΔHΠΘ∥F\Vert\tilde Z - Z\Vert_F \ge \gamma_{\min} \sigma_N(W) \sigma_{r'}(\Theta) \Vert\Delta_H \Pi_\Theta\Vert_F with r′=rank(Θ)r' = \mathrm{rank}(\Theta).

Proof sketch. Since σ′≥γmin⁡\sigma' \ge \gamma_{\min} almost everywhere, ∣σ(a)−σ(b)∣=∣∫baσ′∣≥γmin⁡∣a−b∣|\sigma(a) - \sigma(b)| = |\int_b^a \sigma'| \ge \gamma_{\min}|a - b| entrywise, so ∥Z~−Z∥F≥γmin⁡∥ΔWHΘ∥F\Vert\tilde Z - Z\Vert_F \ge \gamma_{\min}\Vert\Delta_W H\Theta\Vert_F. Write the compact SVD HΘ=UrΣrVr⊤H\Theta = U_r\Sigma_r V_r^\top, so that ΠHΘ=UrUr⊤\Pi_{H\Theta} = U_rU_r^\top. For each row xi⊤x_i^\top of ΔW\Delta_W, ∥(HΘ)⊤xi∥=∥ΣrUr⊤xi∥≥σr(HΘ)∥ΠHΘxi∥\Vert(H\Theta)^\top x_i\Vert = \Vert\Sigma_r U_r^\top x_i\Vert \ge \sigma_r(H\Theta) \Vert\Pi_{H\Theta}x_i\Vert. Summing the squares over rows gives ∥ΔWHΘ∥F≥σr(HΘ)∥ΔWΠHΘ∥F\Vert\Delta_W H\Theta\Vert_F \ge \sigma_r(H\Theta) \Vert\Delta_W\Pi_{H\Theta}\Vert_F. The feature bound follows the same way, using ∥WY∥F≥σN(W)∥Y∥F\Vert W Y\Vert_F \ge \sigma_N(W)\Vert Y\Vert_F and the same projection argument for Θ\Theta. Conversely, ΔWHΘ=ΔWΠHΘHΘ\Delta_W H\Theta = \Delta_W\Pi_{H\Theta}H\Theta, so any ΔW\Delta_W with ΔWΠHΘ=0\Delta_W\Pi_{H\Theta} = 0 leaves the output unchanged; this shows the restriction is necessary.

The unrestricted ∥ΔW∥F\Vert\Delta_W\Vert_F form holds only for perturbations whose rows lie in col(HΘ)\mathrm{col}(H\Theta); perturbations orthogonal to this at most CC-dimensional readout subspace are invisible to any single linear-readout layer. This is inherent to the architecture, and we will state it explicitly. We will also narrow the discussion after Theorem 5.2, which currently claims a guarantee against feature collapse and that full rank is "necessary". The revised text states that (i) any perturbation with nonzero ΔWΠHΘ\Delta_W \Pi_{H\Theta} changes the output by at least γmin⁡σr(HΘ)∥ΔWΠHΘ∥F\gamma_{\min} \sigma_r(H\Theta) \Vert\Delta_W \Pi_{H\Theta}\Vert_F; (ii) perturbations orthogonal to col(HΘ)\mathrm{col}(H\Theta) leave the output unchanged, so no lower bound in the full ∥ΔW∥F\Vert\Delta_W\Vert_F can hold; and (iii) with Theorem 5.1, the output change is bounded above and below by ∥ΔWΠHΘ∥F\Vert\Delta_W \Pi_{H\Theta}\Vert_F, up to the factor (γ/γmin⁡)(σ1(HΘ)/σr(HΘ))(\gamma/\gamma_{\min}) (\sigma_1(H\Theta)/\sigma_r(H\Theta)). We will remove the necessity claim, present spectral regularization only as a way to keep σr(HΘ)\sigma_r(H\Theta) away from zero, and restrict the corresponding sentence in Section 2 to the readout subspace.

AuthorsOpenReview ↗

Rebuttal Continued

W2. Discrete k-NN selection breaks the chain from Eq. (3) to Theorem 5.1. We agree that the current analysis does not connect text edits to the actual graph operator, because neighbor rankings can change. We will add the following lemma in a new Section 4.4 (kk-NN Sparsification), with the full proof in the appendix:

Lemma (edge-set invariance). Let ηj=d(πj,π~j)\eta_j = d(\pi_j,\tilde\pi_j) and ηmax⁡=max⁡jηj\eta_{\max} = \max_j \eta_j, and let the kernel κ\kappa be LκL_\kappa-Lipschitz on the range of distances. For each node ii, let gi=di(k+1)−di(k)g_i = d_i^{(k+1)} - d_i^{(k)} be the gap between the distances to its (k+1)(k{+}1)-th and kk-th nearest neighbors. If gi>0g_i > 0 for all ii and ηmax⁡<14min⁡igi\eta_{\max} < \tfrac14 \min_i g_i, then (i) the k-NN neighbor sets, and hence the symmetrized edge set, are unchanged; and (ii) ∣W~ij−Wij∣≤2Lκηmax⁡|\tilde W_{ij} - W_{ij}| \le 2L_\kappa \eta_{\max} for all i,ji,j, so that ∥W~−W∥F≤2Lκηmax⁡∣E∣\Vert\tilde W - W\Vert_F \le 2L_\kappa \eta_{\max}\sqrt{|E|}.

Proof sketch. By Eq. (3), every pairwise distance moves by at most 2ηmax⁡2\eta_{\max}. For a neighbor jj and a non-neighbor ll of node ii, d~ij≤di(k)+2ηmax⁡<di(k+1)−2ηmax⁡≤d~il\tilde d_{ij} \le d_i^{(k)} + 2\eta_{\max} < d_i^{(k+1)} - 2\eta_{\max} \le \tilde d_{il}, using 4ηmax⁡<gi4\eta_{\max} < g_i. Every current neighbor therefore stays strictly closer than every non-neighbor, so the neighbor sets, and hence the symmetrized edge set, are unchanged. On each edge, ∣κ(d~ij)−κ(dij)∣≤Lκ∣d~ij−dij∣≤2Lκηmax⁡|\kappa(\tilde d_{ij}) - \kappa(d_{ij})| \le L_\kappa|\tilde d_{ij} - d_{ij}| \le 2L_\kappa\eta_{\max}, and the symmetrization factor is at most 1. The constant 14\tfrac14 cannot be improved from Eq. (3) alone, since the kk-th neighbor may move away by 2ηmax⁡2\eta_{\max} while the (k+1)(k{+}1)-th moves closer by the same amount.

With calibration, the edge set is the same, since neighbors are selected under dd and the calibration map is monotone; (ii) then holds with LκL_\kappa replaced by MLκML_\kappa for a fixed map. Combining the lemma with Theorem 3.1 gives a sufficient condition on the token edit itself. For Hellinger, for example, dH2≤TVd_H^2 \le \mathrm{TV}, so ηj≤12(ρjΔj/Lj)∥Aj♯∥1\eta_j \le \sqrt{\tfrac12 (\rho_j\Delta_j/L_j) \Vert A_j^\sharp\Vert_1}, where ρj\rho_j is the number of edited vocabulary rows. When the margin is violated, we make no stability claim.

Robustness of neighbor sets. We ran an experiment to measure how robust the k-NN neighbor sets are to text edits. On the paper's graph (k=3k = 3, Hellinger, 773 nodes), we replaced mm function words in a chunk, rebuilt the WANs and the k-NN graph, and counted the nodes whose neighbor set changes. We considered two conditions: a single chunk is edited (each of the 773 chunks in turn), or every chunk is edited at once. No model is trained, so the result does not depend on the split.

Margin distribution. We also computed the neighbor margins gig_i. Since we are unable to add the figure to the rebuttal, we describe it. The median margin is 1.4×10−31.4\times10^{-3}; most margins lie between 10−410^{-4} and 10−210^{-2}, with a peak just above the median and a tail toward smaller values. The 34 nodes with gi=0g_i = 0 correspond to duplicate texts, whose kk-th and (k+1)(k{+}1)-th neighbors tie exactly. We will include the figure of this distribution in the revised paper.

Neighbor-set changes. When a single chunk is edited, the fraction of nodes whose neighbor set changes is between 0.06% and 0.76% across all edit levels. When every chunk is edited at once, it grows with the number of replacements per chunk, from 25.0% at m=1m = 1 to 99.8% at m=300m = 300.

Replacements per edited chunk mm131030100300
One chunk edited0.06%0.10%0.19%0.32%0.57%0.76%
Every chunk edited25.0%41.5%64.3%83.8%97.0%99.8%

A single edited text changes the neighbor sets of 0.4 nodes on average after one replacement and of at most 6 nodes after 300, and in 81% of single-replacement edits no other node is affected. Edits therefore change the edge set only locally, around the edited text, as the node-wise form of the lemma in the appendix describes (2ηi+ηˉin+ηˉout<gi2\eta_i + \bar\eta_{in} + \bar\eta_{out} < g_i). When neighbor sets do change, the swapped neighbors are nearly equidistant, so an edge is replaced by one of almost equal weight; this is consistent with the robustness of the calibrated model under text edits (Q2).

AuthorsOpenReview ↗

Rebuttal Continued

W3. Figure 3 (94% zeros) vs. the shared k-NN edge set in D.1.3. This is correct: calibration transfers only the active-weight distribution νactive\nu_{\text{active}} of the reference graph. It does not transfer the reference graph's sparsity pattern or connectivity. The Method Overview (Section 2) states that only the active-weight distribution is transferred. In the revision, we make explicit that calibration acts on the kernel weights of all text pairs, wij=f(cij)w_{ij} = f(c_{ij}), before k-NN sparsification, and we update Figure 1 and the Method Overview to show this order. Since neighbors are selected under dd and ff is monotone, both models share the same edge set, and we will revise any wording suggesting that calibration sparsifies the WAN graph or transfers the reference graph's sparsity pattern or connectivity.

W4. Play-disjoint evaluation. We agree that this is necessary, and we ran it. All chunks of a play are assigned to a single partition (146 play groups, with co-authored plays and identical texts kept together), stratified by author, at about 60/20/20. No step of graph construction or calibration uses labels. To also evaluate on unseen documents, we ran an inductive variant on these splits: every edge incident to a test node is removed during training and model selection and restored only for the final test. In this setting, test texts are unseen during training: no information passes to or from them during training and model selection, and they join the graph only at test time, as new documents would.

MethodPlay-disjoint (%)
LLM-MLP65.7 ± 5.9
LLM-GCN67.8 ± 5.2
LLM-GCN-OT70.2 ± 4.9
LLM-GCN, inductive66.9 ± 5.3
LLM-GCN-OT, inductive70.4 ± 5.2

On unseen plays, LLM-GCN-OT attains the highest classification rate in both the transductive and inductive settings and keeps a 4.5-point margin over the LLM-MLP, and the inductive variants stay within 1 point of their transductive counterparts. The gain therefore relies neither on same-play neighbors nor on test-node edges during training. We will add these results to Table 1.

W5. Ablations separating syntactic geometry from graph propagation. We agree, and we ran both controls: (i) An MLP on the concatenation [hi∥π^i][h_i \Vert \hat\pi_i], with no graph. All models use the same single layer and the same parameter count (2,490); π^\hat\pi is z-scored. Play-disjoint split:

ClassifierLLM features (%)LLM + π^\hat\pi features (%)WAN only, no graph (%)
Original WAN (Segarra et al., 2015) (input: WAN transition matrices)––68.9 ± 3.2
Linear classifier on WAN stationary distribution π^\hat\pi (input: π^\hat\pi)––58.2 ± 5.7
MLP (no graph)65.7 ± 5.969.8 ± 5.2–
GCN67.8 ± 5.269.5 ± 4.6–
GCN on OT-calibrated graph70.2 ± 4.972.4 ± 5.2–

The two WAN baselines use no graph and no LLM features. The linear classifier takes each chunk's 30-dimensional π^\hat\pi as input. Original WAN uses neither π^\hat\pi features nor learning: it compares each chunk's full WAN transition matrix with one WAN profile per author under relative entropy and assigns the closest author. WAN statistics are complementary to the LLM embeddings even without a graph (65.7% to 69.8% for the MLP). With identical inputs and capacity, the OT-calibrated WAN graph adds a significant gain over the MLP (72.4% vs. 69.8%, p=0.004p = 0.004, paired t-test), so propagation over the WAN geometry contributes beyond the features themselves. With raw weights the graph adds nothing (69.5% vs. 69.8%), which is the weight-concentration problem (O4) that calibration addresses. (ii) A GCN on a k-NN graph built from embedding similarity in place of WAN distances. Only the edge set changes: exact k-NN (k=3k = 3) on cosine distances between the frozen MiniLM embeddings, with the same kernel 1/(1+d)1/(1+d). Node features, the OT reference W∗W^{\ast}, the GCN, training, splits and seeds are unchanged. Play-disjoint split:

MethodTest accuracy (%)
LLM-MLP (no graph)65.7 ± 5.9
LLM-GCN on WAN graph67.5 ± 5.2
LLM-GCN on OT-calibrated WAN graph70.2 ± 4.9
LLM-GCN on embedding graph69.3 ± 6.1
LLM-GCN on OT-calibrated embedding graph69.8 ± 4.8

On unseen plays, the embedding graph does not outperform the OT-calibrated WAN graph (69.3% vs. 70.2%, p=0.57p = 0.57). The two graphs encode different relations: embedding similarity mostly reflects content, so 31% of the embedding graph's edges link chunks of the same play, against 11% for the WAN graph, while across different plays the WAN graph links more same-author chunks (46.5% vs. 43.8% of edges). OT calibration also improves the WAN graph by 2.7 points but the embedding graph by only 0.5, consistent with calibration correcting the weight concentration specific to WAN graphs (O4): without calibration the WAN graph trails the embedding graph (67.5% vs. 69.3%), and with it the WAN graph is ahead.

AuthorsOpenReview ↗

Rebuttal Continued

W6. Scope of Proposition 6.1. We agree: the assumption was stated only near zero, while the conclusion covered the whole active distribution. Under homogeneous degrees, Corollary 6.1(ii) gives Wij∗=c−Zij/βW_{ij}^{\ast} = c - Z_{ij}/\beta for some constant cc. This map is decreasing, so ZijZ_{ij} near zero corresponds only to the largest weights, and an assumption near zero constrains pWp_W only near the top of its support. Since an edge is active exactly when Zij∈[0,βc)Z_{ij} \in [0, \beta c), the assumption is needed on this active range, and we restate the proposition accordingly:

Proposition 6.1 (revised). Assume approximately homogeneous degrees di≈dˉd_i \approx \bar d and that the density pZp_Z of ZijZ_{ij} is nondecreasing on the active range [0,βc)[0, \beta c) for some constant cc. Then the density pWp_W of active weights Wij∗>0W_{ij}^{\ast} > 0 is nonincreasing on (0,c](0, c], hence right-skewed.

Proof sketch. The map z↦c−z/βz \mapsto c - z/\beta sends [0,βc)[0, \beta c) onto (0,c](0, c], and by change of variables pW(w)∝βpZ(β(c−w))p_W(w) \propto \beta p_Z(\beta(c - w)). Since this transformation is decreasing and pZp_Z is nondecreasing on [0,βc)[0, \beta c), pWp_W is nonincreasing on (0,c](0, c].

The applicable domain is now explicit. The active edges are the closest pairs, which form the lower tail of the distribution of ZijZ_{ij}, so the assumption holds whenever the reference graph is sparse enough for this tail to lie below the mode of pZp_Z, which the hyperparameters α\alpha and β\beta control (Corollary 6.1).

W7. Scope with respect to the encoder. We agree. All results use frozen all-MiniLM-L6-v2 embeddings. We will revise the title, abstract and conclusion so they do not imply validation across general large generative LMs. We will also add an encoder ablation to the final version of the paper.

Q2. For play-disjoint results, see W4. To test how much of the robustness persists when model parameters trained on the clean graph are not retrained, we perturbed the text itself on play-disjoint splits. In every chunk, mm function-word tokens chosen at random are replaced by a different function word; the WANs, the k-NN graph and the OT weights are rebuilt from the edited text:

Edits per chunk mmLLM-GCN (%)LLM-GCN-OT (%)
067.5 ± 4.970.6 ± 4.7
168.3 ± 3.970.6 ± 4.8
367.9 ± 4.270.7 ± 5.0
1066.9 ± 3.670.4 ± 5.1
3066.9 ± 4.870.5 ± 4.5
10064.7 ± 6.270.3 ± 5.5
30062.1 ± 3.769.7 ± 4.7

On unseen plays, LLM-GCN loses 5.3 points at 300 edits per chunk, while LLM-GCN-OT loses 0.9 points and is more accurate at every edit level.

Q3. See W5 for the experiment on the embedding-space similarity k-NN graph. Regarding simpler monotone edge-weight transformations, LLM-GCN already uses a fixed monotone transformation: its weights are κ(d)=(1+d)−1\kappa(d) = (1+d)^{-1}, a decreasing function of the metric dd. Any other fixed monotone transform gg of these weights gives g∘κg \circ \kappa, which is again a fixed decreasing kernel of dd. OT differs in that the transform is not fixed but fitted to the data, so that the weights follow the distribution of the LLM reference graph. The comparison between LLM-GCN and LLM-GCN-OT therefore compares a fixed transform with the data-adaptive one (67.8% vs. 70.2% on play-disjoint splits, W4), and OT outperforms the fixed transform especially on denser graphs. We varied k∈{3,10,20}k \in \lbrace 3, 10, 20 \rbrace on play-disjoint splits for GCN, GIN, GraphConv, GPRGNN and SAGE, with LLM features only; for OT, (α,β)(\alpha, \beta) is selected for each kk by validation loss:

ModelWeightsk=3k = 3 (%)k=10k = 10 (%)k=20k = 20 (%)
Same-author edge fraction54.350.347.0
LLM-GCNraw67.5 ± 5.165.1 ± 2.566.6 ± 3.8
OT72.2 ± 5.271.5 ± 4.172.1 ± 2.7
LLM-GINraw66.3 ± 6.161.3 ± 5.829.9 ± 8.9
OT70.4 ± 5.171.3 ± 5.371.6 ± 5.0
LLM-GraphConvraw68.7 ± 4.860.4 ± 16.932.3 ± 6.5
OT68.0 ± 4.769.8 ± 4.269.4 ± 4.0
LLM-GPRGNNraw72.1 ± 5.370.7 ± 4.468.1 ± 5.9
OT71.5 ± 5.471.7 ± 5.671.9 ± 4.3
LLM-SAGEraw71.1 ± 5.371.5 ± 6.371.5 ± 3.6
OT70.4 ± 4.572.4 ± 2.972.5 ± 3.4

As kk grows, the fraction of same-author edges falls from 54.3% to 47.0%, so raw weights carry less information. With raw weights, LLM-GIN and LLM-GraphConv collapse at k=20k = 20 (29.9% and 32.3%), and LLM-GCN and LLM-GPRGNN lose accuracy. With OT, every architecture stays between 69.4% and 72.5% at k=10k = 10 and k=20k = 20.

We thank the reviewer again. These corrections make the theoretical claims accurate and the empirical claims better controlled. We are happy to discuss any further questions. If you find that our rebuttal fully addresses your concerns, we would be grateful if you would consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.