Pith. sign in

REVIEW 4 major objections 5 minor 37 references

From Embedding Geometry to Spectral Search: Energy Dispersion Networks For Vector Retrieval

T0 review · 4 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Embedding spaces can be searched more coherently by mixing cosine geometry with a spectral energy signal taken from a feature-space graph Laplacian.

desk verdict Useful open-source spectral re-ranker with modest TREC gains; complementarity claim is still unisolated. read the letter →

arxiv 2606.21535 v2 pith:IAQRVIOD submitted 2026-06-19 cs.IR

Lorenzo Moriondo , Ilias Azizi This is my paper · ORCID
classification cs.IR
keywords spectralindexinggraphwiringfeature-spaceLaplacianenergydispersionnetworksvectorretrievalRAGtau-modulationepiplexity
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

The paper argues that high-dimensional embedding spaces are not only clouds of points related by angles, but also energy networks whose feature-space topology carries corpus-level semantic structure that pure geometric scores miss. It introduces Graph Wiring to build a Laplacian over feature dimensions after a centroid reduction of the data matrix, then Spectral Indexing (SPIN), which ranks each item by a linear mix of cosine similarity and a bounded Rayleigh-quotient energy score. That mix is controlled by a single parameter tau, so a system can trade local geometric affinity for topological coherence—exactly the adaptive behavior modern retrieval-augmented generation pipelines need. On an industrial vulnerability corpus and on the human-labeled TREC-COVID benchmark, the spectral mix flattens score tails, raises tolerant and graded relevance metrics, and produces positive semantic uplift on most queries while preserving near-perfect tolerant recall. The authors treat the extra structural information in the Laplacian as certified by an information-theoretic notion they call epiplexity.

What carries the argument

Spectral Indexing (SPIN): the lambda-aware distance dist_SPIN = alpha · sim_cos(q,i) + (1-alpha) · sim_lambda(q,i), where sim_lambda is the difference of bounded Rayleigh quotients on the feature-space Laplacian and alpha (called tau) is the geometric-versus-spectral mix weight.

What would settle it

Hold the embedding matrix fixed, replace the feature-space Laplacian with a randomized or degree-matched null graph that preserves degree sequence but destroys topology, and test whether the reported gains in tolerant recall, NDCG, and tail-shape metrics vanish; if they survive, the spectral claim fails.

Watch

Extended reading notes

Core claim

A graph Laplacian built on the feature-space of an embedding corpus yields a per-vector energy-dispersion signal (the Rayleigh quotient mapped into a bounded lambda-tau score) that encodes semantic correlations invisible to geometric search alone. Linearly combining that spectral signal with cosine similarity therefore produces ranked lists that are more head–tail coherent and more semantically aligned, while the mixing weight itself becomes a controllable handle for adaptive retrieval.

Load-bearing premise

The feature-space graph built after clustering the embeddings still carries semantic structure that is independent of, and complementary to, the geometry already present in those same embeddings.

Editorial extensions

If this is right

  • Cosine retrieval can be augmented rather than replaced, keeping local geometry while adding corpus topology.
  • RAG and agent pipelines can modulate tau mid-session—high for broad frame-setting, lower for fine semantic sieving.
  • Long ranked lists need not collapse into noisy low-affinity tails; spectral re-ranking can keep relevance stable deeper into the list.
  • Queries that land in high-energy or irregular regions of the feature manifold become visible as spectral outliers instead of silent geometric neighbors.
  • The same feature-space Laplacian construction can serve as a multi-purpose structural index for data curation, drift detection, and generative retrieval identifiers.

Reading between the lines

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

  • If the spectral score is truly a cheap one-dimensional complement, existing approximate-nearest-neighbor indexes could store a single lambda scalar per vector and reuse their infrastructure with almost no extra memory.
  • The method is likely to degrade when the embeddings themselves lack a stable feature manifold (random projections, very sparse bags-of-words), giving a clean stress test that the paper leaves open.
  • Tau-modulation naturally suggests a closed feedback loop that watches context entropy or generation quality and retunes the mix automatically, which the authors flag only as future work.
  • Treating feature space as the semantic metadata layer of a corpus opens a route to spectral unique identifiers that could drive generative retrieval without a separate discrete codebook.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper argues that embedding spaces should be treated not only as geometric point clouds but as spectral energy networks induced by a feature-space graph Laplacian. It introduces Graph Wiring (GW) to build that Laplacian after centroid reduction of the embedding matrix, and Spectral Indexing (SPIN) which mixes cosine similarity with a bounded Rayleigh-quotient score λτ via a convex weight τ (tau). Empirically, hybrid and taumode settings improve head–tail score coherence on an industrial CVE corpus and yield modest gains in Relevance Recall@10 (~2.8%) and NDCG@10 (~2.6%) plus higher semantic uplift on the human-labeled TREC-COVID benchmark. The authors release the arrowspace library and frame the construction via an epiplexity / two-part MDL argument.

Significance. If the complementarity claim holds, SPIN would give RAG and domain retrieval a practical, tunable corpus-structure signal that pure cosine lacks, together with an open implementation. The feature-space (rather than item-space) Laplacian and the explicit tau-modulation idea are useful contributions. Strengths that should be credited include the open-source arrowspace library, dual evaluation on an industrial CVE workload and human-qrel TREC-COVID, extensive qualitative appendices, and an explicit limitations section that acknowledges circular Traditional Recall and the frozen encoder. The absolute gains remain small and the theoretical isolation of the spectral signal is incomplete, so significance is currently provisional rather than decisive.

major comments (4)
  1. The central complementarity claim (that LF + Rayleigh λτ encodes semantic structure independent of the geometry already present in the same embeddings) is never isolated experimentally. All reported gains use the convex mix dist_SPIN = α·sim_cos + (1-α)·sim_λ with τ and graph hyperparameters (ε, k, p, σ) chosen on the evaluation sets. Because LF is constructed from the identical embedding matrix that defines cosine, improvements could be re-weightings of existing geometry. A destroy-spectral control (e.g., degree-preserving edge randomization of LF, or permutation of the λτ marginals while holding embeddings fixed) is required to substantiate that the spectral term is complementary rather than redundant.
  2. §5.1 and Table 3: Traditional Recall@25 is defined against cosine-ranked neighbors, so cosine scores 1.0 by construction (explicitly admitted). Semantic Uplift and Tolerant Recall soften but do not remove reference dependence on the CVE experiment; many negative-uplift queries are then re-interpreted qualitatively rather than against independent labels. The load-bearing external evidence is therefore the TREC-COVID human-qrel results (§4.4, Fig. 3), whose absolute gains are modest (~0.508 o0.522 Relevance Recall@10; ~0.455 o0.467 NDCG@10) on only 50 queries with high per-query variance. Stronger claims of “fully-featured semantic search” and “SOTA performance” should be scaled to these effect sizes or supported by larger labeled sets and additional baselines (e.g., learned re-rankers, graph-based ANN).
  3. §1–§2 and the epiplexity citations: the theoretical foundation is asserted via two-part MDL / epiplexity pointing primarily to the authors’ contemporaneous notes ([9], [13], [28], [31]) rather than a self-contained derivation or an experimental MDL certificate computed on the same corpora used for retrieval. The paper states that the Laplacian “carries structural information beyond the original geometric scores” but does not report the compression test MDL_T(X) < N·F·b (or an equivalent diagnostic) for the CVE or TREC embeddings. Either supply that certificate on the evaluation data or demote epiplexity from “theoretical foundation” to related work / motivation.
  4. Free parameters and tuning: τ, graph hyperparameters (ε, k, p, σ), the ad-hoc embedding scale 1.12, centroid count C (JL), and the Rayleigh ε are all free. Pareto tuning is mentioned but not reported with held-out selection or sensitivity analysis that would show the gains survive reasonable hyperparameter perturbation. Without this, it is hard to separate method contribution from search over a flexible hybrid score.
minor comments (5)
  1. Notation for the mixing weight is overloaded: the text uses both α and τ (and “tau”) for the same convex coefficient; unify early (e.g., §3.1).
  2. Several appendix tables and figures (e.g., Fig. 5, Table 6) are referenced as available with code; for journal form they should be self-contained or moved to a permanent archive with stable DOIs.
  3. Typos and polish: “ODD item” (§3.1) should be OOD; “ttau” in §4.4; “SPectral INdexing” capitalization is inconsistent; arXiv id in header vs body.
  4. Related work on graph signal processing and spectral clustering is cited, but comparison to existing hybrid retrieval / learned sparse+dense or manifold-aware ANN methods is thin; a short positioning paragraph would help.
  5. CVE qualitative appendix is valuable but very long; consider moving the full 11-query walkthrough to supplementary material and keeping a compact summary of verdict categories in the main text.

Circularity Check

4 steps flagged · score 5.0 of 10

Admitted cosine-as-ground-truth metrics and self-cited epiplexity foundation create partial circularity; TREC human labels keep the central empirical claim independent.

  1. self definitional [§5.1 Limitations, Circular cosine baseline]
    "The Traditional Recall metrics use a ground truth constructed from cosine-ranked neighbors; cosine therefore attains perfect recall by construction. This renders Traditional Recall an unreliable proxy for semantic quality and motivates the use of Tolerant Recall and Semantic Uplift as primary evaluation criteria"

    Traditional Recall@k is defined against cosine-ranked neighbors, so cosine scores 1.0 by definition. Any comparison that treats this as evidence of cosine quality (or of SPIN degradation relative to it) is circular. The paper admits the issue but still reports Traditional Recall tables that encode this tautology.

  2. self definitional [§4.3 Semantic uplift; Fig. 2 caption; Appendix metrics]
    "Semantic uplift is defined as the per-query difference Ruplift = Rtol − Rtrad. ... Cosine (left) provides no uplift, taumode (right) provides relevant uplift for most of queries"

    Rtrad is measured against cosine-defined ground truth, so for cosine Rtrad = Rtol = 1 and Ruplift = 0 always. Presenting mean uplift 0.00 under cosine as an empirical result (Fig. 2a, Fig. 4a) is definitional, not a measured failure of geometric search. Positive SPIN uplift partly rewards divergence from that same cosine reference within a tolerance band.

2 more flagged steps
  1. self citation load bearing [Introduction, theoretical foundation paragraph; also claims list]
    "The theoretical foundation of GW is supported through epiplexity, an information-theoretic framework showing that the graph Laplacian constructed by GW carries structural information. Following [13], where we use a two-part Minimum Description Length (MDL) criterion, together with independent structural diagnostics, to prove that the constructed spectral representation encodes information beyond the original geometric scores (details in [9])."

    The load-bearing claim that LF encodes information beyond geometric scores is not proved in this paper; it is deferred to Moriondo's own contemporaneous note [9]. Finzi et al. [13] only define epiplexity generally. The paper's assertion that SPIN provides 'semantic lift ... by leveraging the structural information made evident by epiplexity' therefore rests on an unverified same-author citation rather than an independent derivation or external check.

  2. self citation load bearing [§1 closing; §5 Discussion opening]
    "Finally, more generally, the epiplexity-based analysis [14] suggests that GW can expose additional structural information beyond purely geometric retrieval signals... The core idea of GW is that an item vector xi ∈ R^{N×F}, when blended with the graph Laplacian in the feature-space LF ∈ R^{F×F}, carries measurably more discriminative information than its geometric measurement in the item-space alone, as demonstrated in the terms of epiplexity framework and structural information in [28, 9]."

    [14] is the authors' own PyArrowSpace repository and [28]/[9] are same-author GitHub/preprint materials. The 'measurable' surplus of structural information over pure geometry is asserted via this self-citation chain rather than by an ablation or independent certificate inside the present paper.

full rationale

The method definition (feature-space Laplacian, Rayleigh energy, convex mix with tau) is not circular by construction: it is a standard spectral construction applied in feature space and evaluated empirically. Circularity appears in two supporting pillars. First, Traditional Recall and the zero-uplift baseline for cosine are definitional once cosine neighbors are taken as ground truth; the paper admits this for Traditional Recall but still presents cosine semantic uplift as 0.00 as if it were an empirical finding. Second, the claim that the Laplacian encodes structural information beyond geometry is justified by epiplexity with the load-bearing proof deferred to the first author's contemporaneous note [9], not derived or independently verified in this manuscript (Finzi et al. [13] supplies only the general framework). These do not force the TREC NDCG/Relevance-Recall gains, which rest on external human labels, so the central retrieval claim retains independent content. Score 5 reflects partial circularity in theory and some metrics without collapse of the main empirical result.

Assumptions & free parameters 5 free parameters · 4 assumptions · 4 invented entities

The central claim rests on a small set of free mixing and graph-construction parameters, standard spectral-graph identities, the domain assumption that feature-space topology encodes semantics missing from item-space geometry, and several named constructs (GW, SPIN, taumode, epiplexity) whose independent status is only partially established by the experiments.

free parameters (5)
  • tau (spectral mixing weight) = 0.42 (CVE), ~0.7 (TREC peak)
    Chosen per dataset (0.42 for CVE 'taumode', 0.65-0.80 for best TREC) to maximize reported quality; the claim that spectral information helps depends on this fitted value.
  • graph hyperparameters (eps, k, p, sigma)
    Neighborhood radius, degree, weight exponent and Gaussian bandwidth that define the feature-space adjacency; tuned via Pareto trade-offs on the same evaluation data.
  • embedding scale factor 1.12 = 1.12
    Uniform multiplicative scaling applied before graph formation 'to improve graph formation'; ad-hoc numerical choice.
  • centroid count C (via JL lemma) = C ~ O(sqrt(N))
    Dimensionality of the reduced feature matrix on which the Laplacian is built; affects both cost and spectral fidelity.
  • epsilon in lambda_tau = R/(R+eps)
    Stabilizer that maps Rayleigh energy into [0,1); changes the numerical scale of the spectral score.
assumptions (4)
  • domain assumption The combinatorial/normalized Laplacian of a graph on feature dimensions encodes corpus-level semantic connectivity not already present in pairwise cosine geometry.
    Stated as the core premise of Graph Wiring (§3) and never isolated by an ablation that destroys only topology while keeping embeddings fixed.
  • standard math Rayleigh quotient R(x)=x^T L_F x / x^T x is a meaningful normalized Dirichlet energy that ranks items from 'smooth/typical' to 'rough/anomalous' on the feature manifold.
    Classical spectral-graph identity; its semantic interpretation for retrieval is the paper's domain reading.
  • ad hoc to paper Two-part MDL / epiplexity certifies that the constructed Laplacian carries non-trivial structural information beyond geometric scores.
    Invoked via citations to the authors' own contemporaneous notes ([9],[13],[28]); not an independently established theorem in the IR literature.
  • standard math Johnson-Lindenstrauss supplies a sufficient number of centroids C for the reduced feature matrix.
    Used to justify the C x F intermediate representation (§3).
invented entities (4)
  • Graph Wiring (GW)
    purpose: General framework that builds a feature-space Laplacian and treats the embedding corpus as an energy network.
    Named construct introduced by the paper; independent evidence is the experimental lift and open-source library, not an external measurement.
  • Spectral Indexing / SPIN
    purpose: Task-specific retrieval algorithm that linearly mixes cosine with lambda-tau spectral scores under a free tau.
    The concrete search procedure whose performance is claimed; evidence is internal to the two evaluation sets.
  • taumode / lambda_tau energy-dispersion score
    purpose: Bounded 1-D positional score derived from Rayleigh energy used as the spectral half of the hybrid distance.
    Synthetic score invented for the method; no external physical or information-theoretic measurement validates it outside the paper's own metrics.
  • epiplexity (as used here)
    purpose: Information-theoretic certificate that the Laplacian encodes structural information.
    Defined and applied in the authors' concurrent preprints; treated as established theory inside this manuscript.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Embedding Geometry to Spectral Search: Energy Dispersion Networks For Vector Retrieval." pith.science (2026). https://pith.science/paper/IAQRVIOD

@misc{pith2026260621535,
  author       = {Pith},
  title        = {Pith review of: From Embedding Geometry to Spectral Search: Energy Dispersion Networks For Vector Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IAQRVIOD}},
  note         = {Machine review of arXiv:2606.21535}
}
read the original abstract

High-dimensional vector spaces, particularly embedding spaces with dense semantic structure, are often interpreted primarily leveraging solely geometric relationships. In this work, we show that they can also be viewed as spectral energy networks induced by the topology of their underlying feature-space manifold with relevant improvements for downstream tasks. Building on this perspective, we introduce Graph Wiring, a general framework for exploiting feature-space spectral structure, together with Spectral Indexing, its task-specific instantiation for vector search. By coupling geometric similarity with spectral information, the proposed method improves Head-Tail coherence and semantic alignment relative to purely geometric retrieval methods. It further supports adaptive search behavior through tau-modulation, providing the flexibility increasingly required by modern Retrieval-Augmented Generation (RAG) pipelines. We present the complete algorithmic pipeline, establish its theoretical foundation through epiplexity, and evaluate the approach across benchmark and industrial settings using the open-source arrowspace library.

Figures

Figures reproduced from arXiv: 2606.21535 by the authors.

Figure 1
Figure 1. Tail quality for 4 queries for all tested modes. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. CVE™ dataset: Semantic Uplift. Cosine (left) provides no uplift, taumode (right) provides relevant uplift for most of queries 4.4 TREC-COVID experiment We evaluate SPIN on the TREC-COVID benchmark [25], a labeled document-queries COVID-related dataset containing expert-designed queries, and human-annotated relevance judgments. Unlike standard ANN benchmarks that only provide metric nearest-neighbor ground truth, TRE… view at source ↗
Figure 3
Figure 3. Impact of the spectral mixing parameter tau on retrieval quality. Lower tau values incorporate stronger spectral information, while tau=1 corresponds to pure cosine similarity. Retrieval Quality Figure 3b shows that incorporating spectral information improves Relevance Recall@10 compared to pure cosine similarity (tau=1.0). Cosine reaches a mean recall of 0.508, while intermediate τ values achieve the best performan… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Semantic Uplift on the Trec-Covid dataset. Cosine retrieval provides limited uplift, while the spectral retrieval modes improve semantic recall across most queries. Semantic Uplift [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Head_K sweep: quality of taumode results improves according to tail metrics. Appendix: GW additional comparative tables 12 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 2 linked inside Pith

  1. [9]

    Moriondo

    L. Moriondo. Epiplexity and graph wiring: An empirical study. https://doi.org/10.22541/au.177430060.02394540/ v1, 2026

    doi:10.22541/au.177430060.02394540/ 2026

  2. [13]

    Finzi, A

    M. Finzi, A. Potapczynski, M. Goldblum, and A. G. Wilson. From entropy to epiplexity: Rethinking information for computationally bounded intelligence.arXiv preprint arXiv:2601.03220, 2026

    arXiv 2026

  3. [28]

    Graph wiring epiplexity

    tuned-org-uk. Graph wiring epiplexity. https://github.com/tuned-org-uk/graph-wiring-epiplexity, 2026. GitHub repository

    2026

  4. [31]

    Moriondo

    L. Moriondo. Graph wiring: Eigenstructures for vector datasets and llm operations. https://www.techrxiv.org/doi/ full/10.36227/techrxiv.177220780.02840438/v1, 2026

    doi:10.36227/techrxiv.177220780.02840438/v1 2026

  5. [1]

    Graph-based vector search: An experimental evaluation of the state-of-the-art.Proceedings of the ACM on Management of Data, 3(1):1–31, 2025

    Ilias Azizi, Karima Echihabi, and Themis Palpanas. Graph-based vector search: An experimental evaluation of the state-of-the-art.Proceedings of the ACM on Management of Data, 3(1):1–31, 2025

    2025

  6. [2]

    Results of the big ann: Neurips’23 competition.arXiv preprint arXiv:2409.17424, 2024

    Harsha Vardhan Simhadri, Martin Aumüller, Amir Ingber, Matthijs Douze, George Williams, Magdalen Dobson Manohar, Dmitry Baranchuk, Edo Liberty, Frank Liu, Ben Landrum, et al. Results of the big ann: Neurips’23 competition.arXiv preprint arXiv:2409.17424, 2024

    arXiv 2024

  7. [3]

    Steck, C

    H. Steck, C. Ekanadham, and N. Kallus. Is cosine-similarity of embeddings really about similarity? InWWW ’24 Companion, 2024

    2024

  8. [4]

    Kuffo, I

    L. Kuffo, I. Tsakalidou, R. De Viti, A. Angel, J. Isa, and R. Lenhardt. Semantic recall for vector search. InSIGIR, 2026

    2026

Show all 37 references
  1. [5]

    Bruckhaus

    T. Bruckhaus. Rag does not work for enterprises.arXiv preprint arXiv:2406.04369, 2024

    2024 arXiv

  2. [6]

    Rag problems persist

    IBM. Rag problems persist. here are five ways to fix them. https://www.ibm.com/think/insights/rag-problems, 2025

    2025

  3. [7]

    The role of local dimensionality measures in benchmarking nearest neighbor search.Information Systems, 101:101807, 2021

    Martin Aumüller and Matteo Ceccarello. The role of local dimensionality measures in benchmarking nearest neighbor search.Information Systems, 101:101807, 2021

    2021

  4. [8]

    Moriondo

    L. Moriondo. Arrowspace: introducing spectral indexing for vector search.Journal of Open Source Software, 10(113):9002, 2025

    2025

  5. [10]

    D. I. Shuman, S. K. Narang, P. Frossard, A. Ortega, and P. Vandergheynst. The emerging field of signal processing on graphs.IEEE Signal Processing Magazine, 30(3):83–98, 2013

    2013

  6. [11]

    X. Dong, D. Thanou, P. Frossard, and P. Vandergheynst. Learning laplacian matrix in smooth graph signal representations.IEEE Transactions on Signal Processing, 64(23):6160–6173, 2016

    2016

  7. [12]

    Sandryhaila and J

    A. Sandryhaila and J. M. F. Moura. Discrete signal processing on graphs.IEEE Transactions on Signal Processing, 61(7):1644–1656, 2013

    2013

  8. [14]

    PyArrowSpace

    tuned-org-uk. PyArrowSpace. https://github.com/tuned-org-uk/pyarrowspace, 2026. GitHub repository

    2026

  9. [15]

    P. M. B. Vitanyi and M. Li. Minimum description length induction, bayesianism, and kolmogorov complexity. IEEE Transactions on Information Theory, 46(2):446–464, 2000

    2000

  10. [16]

    P. D. Grunwald.The Minimum Description Length Principle. MIT Press, 2007. 9 Running Title for Header

    2007

  11. [17]

    D. A. Spielman. Spectral graph theory lecture notes. https://www.cs.yale.edu/..., 2007

    2007

  12. [18]

    von Luxburg

    U. von Luxburg. A tutorial on spectral clustering.Statistics and Computing, 17(4):395–416, 2007

    2007

  13. [19]

    Johnson and Joram Lindenstrauss

    William B. Johnson and Joram Lindenstrauss. Extensions of lipschitz mappings into a hilbert space.Contemporary Mathematics, 26:189–206, 1984

    1984

  14. [20]

    Macmillan and Co., 1877

    John William Strutt.The Theory of Sound, volume 1. Macmillan and Co., 1877. Reprinted 1945, Dover Publications

    1945

  15. [21]

    Evans.Partial Differential Equations, volume 19 ofGraduate Studies in Mathematics

    Lawrence C. Evans.Partial Differential Equations, volume 19 ofGraduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1998

    1998

  16. [22]

    F. Chung. The heat kernel as the pagerank of a graph.PNAS, 104(50):19735–19740, 2007

    2007

  17. [23]

    PhD thesis, University of York, 2007

    Xiao Bai.Heat Kernel Analysis on Graphs. PhD thesis, University of York, 2007

    2007

  18. [24]

    Cve™: Common vulnerabilities and exposures dataset

    MITRE Corporation. Cve™: Common vulnerabilities and exposures dataset. https://www.cve.org/Legal/ TermsOfUse, 2026

    2026

  19. [25]

    V oorhees, William Hersh, Max Crammer, and Nancy Ide

    Ellen M. V oorhees, William Hersh, Max Crammer, and Nancy Ide. Trec-covid: Constructing a pandemic information retrieval test collection.SIGIR Forum, 54(1):1–12, 2021

    2021

  20. [26]

    Magesh, F

    V . Magesh, F. Surani, M. Dahl, A. Sklar, R. Tseng, R. Guha, D. E. Ho, and C. D. Manning. Hallucination-free? assessing the reliability of leading ai legal research tools.Journal of Empirical Legal Studies, 22(2):216–242, 2025

    2025

  21. [27]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 3982–3992. Association for Computational Linguistics, 2019

    2019

  22. [29]

    N. D. Matsakis and J. Stone. Rayon: A data-parallelism library for rust. https://docs.rs/rayon/latest/rayon/, 2025

    2025

  23. [30]

    Reimers and I

    N. Reimers and I. Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. InEMNLP, 2019

    2019

  24. [32]

    Spectral-aware unique identifiers for generative retrieval and vector search.Authorea Preprints, 2026

    Lorenzo Moriondo. Spectral-aware unique identifiers for generative retrieval and vector search.Authorea Preprints, 2026. 10 Running Title for Header Appendix: Retrieval Metrics We evaluate retrieval quality using both classical relevance metrics and graph-signal-aware measures...

    2026

  25. [33]

    directory traversal

    Large tail-shape improvement under spectral reranking: a substantially higher tail/head ratio or lower tail CV for Taumode than Cosine. Because test_17_CVE_neurips.py calls random.shuffle(queries) before evaluation, theeffectivequery order is the one preserved in the CSV outpu...

    2006

  26. [34]

    the τ= 0.72 crossover already lies inside the spectral regime for this corpus

    Hybrid is not always between Cosine and Taumode.Q2, Q9, Q47, and Q50 show Hybrid and Taumode agreeing with each other whilebothdiffer sharply from Cosine — i.e. the τ= 0.72 crossover already lies inside the spectral regime for this corpus. For Q27 the opposite holds: Hybrid tr...

  27. [35]

    The clearest semantic win in the selected set is Q47(arbitrary file upload → RCE), where Hybrid produces a semantic-recall jump from0.0to1.0while keeping all of cosine’s score-band leaders

  28. [36]

    (no title)

    Many “(no title)” CVEs in this corpusmake individual verdicts uncertain: across the 11 selected queries, only a minority of the spectral top-5 entries are titled. A reproducibility recommendation for the paper is 21 Running Title for Header to extend extract_text to surface CV...

  29. [37]

    race condition in temporary file creation

    The pattern is consistent with the paper’s headline claim —spectral reranking changes theshapeof the top- 25 score distribution and moves between embedding neighborhoods—while urging careful interpretation of NDCG and Kendall τ on small intersections, and acknowledging that fo...

    2008

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.