Pith. sign in

REVIEW 2 major objections 3 minor 41 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.3

Every transpose-invariant spectral diagnostic on attention operators is structurally blind to information-flow direction.

2026-07-01 00:04 UTC pith:E5IJG7GY

load-bearing objection The blindness theorem and Cheeger floor for causal attention are the actual new pieces, but the forced-scoring transport modeling is a real soft spot that needs checking. the 2 major comments →

arxiv 2605.04893 v2 pith:E5IJG7GY submitted 2026-05-06 cs.LG cs.CLstat.ML

Self-Attention as Transport: Limits of Symmetric Spectral Diagnostics

Dominik Dahlem , Diego Maniloff , Mac Misiura This is my paper
classification cs.LG cs.CLstat.ML
keywords self-attentionspectral diagnosticstransport operatorsasymmetry coefficienthallucination detectionbipartite Cheeger constantorientation blindnessdegree-normalized operator
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proves that any spectral diagnostic relying only on the symmetric component of a degree-normalized attention matrix cannot distinguish the matrix from its transpose. This structural blindness means such diagnostics cannot detect the direction of information flow. A converse result bounds how much any Lipschitz diagnostic can sense the transpose using an asymmetry coefficient G. The authors pair this with a closed-form bipartite-Cheeger expression that gives an n-independent capacity floor of 1/5 for uniform causal attention and O(w/n) for window attention. The resulting two-axis view (capacity via φ, direction via G) produces a polarity prediction that is tested on hallucination benchmarks across model families.

Core claim

Every transpose-invariant spectral diagnostic of the degree-normalized attention operator is orientation-blind and cannot detect information-flow direction; a converse bounds the transpose sensitivity of any Lipschitz diagnostic by the asymmetry coefficient G. When this is combined with the closed-form bipartite-Cheeger landscape for canonical causal architectures, uniform causal attention meets an n-independent floor φ ≥ 1/5 while window attention falls as O(w/n). The resulting diagnostic (φ for transport capacity, G for direction) yields the falsifiable claim that bottleneck- and diffuse-dominated benchmarks must exhibit opposite polarity.

What carries the argument

The asymmetry coefficient G, which upper-bounds the transpose sensitivity of any Lipschitz diagnostic applied to the symmetric part of the degree-normalized attention operator.

Load-bearing premise

Attention matrices obtained under forced scoring of benchmark-labeled responses behave as transport operators whose symmetric component governs transport capacity according to the closed-form bipartite-Cheeger landscape derived for canonical causal architectures.

What would settle it

A direct numerical check showing that some transpose-invariant spectral quantity on real attention matrices differs from the same quantity on the transposed matrices would falsify the blindness theorem.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Uniform causal attention satisfies an n-independent floor φ ≥ 1/5 on the bipartite-Cheeger constant.
  • Window attention pierces that floor at rate O(w/n).
  • Bottleneck-dominated and diffuse-dominated benchmarks must show opposite polarity under the two-axis diagnostic.
  • Transport features retain 0.62-0.84 length-controlled AUROC across decoder-only, encoder-only, and encoder-decoder models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Direction-aware diagnostics will need to incorporate explicit asymmetry measures rather than relying solely on symmetric spectra.
  • The fraction of heads that pierce the idealized floor can serve as a static architectural signature independent of any benchmark.
  • The same transport framing could be applied to cross-attention or encoder-decoder attention patterns to test whether the blindness result generalizes.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 3 minor

Summary. The paper claims that every transpose-invariant spectral diagnostic of the symmetric component of the degree-normalized attention operator is structurally orientation-blind and cannot detect information-flow direction, with a converse theorem bounding any Lipschitz diagnostic's transpose sensitivity by the asymmetry coefficient G. It derives a closed-form bipartite-Cheeger landscape for canonical causal architectures, establishing that uniform causal attention satisfies an n-independent floor φ ≥ 1/5 while window attention scales as O(w/n), with failure modes being shape-different. The resulting two-axis diagnostic (φ for capacity, G for direction) is applied to attention matrices obtained via forced scoring of benchmark-labeled responses, producing length-controlled AUROC values of 0.62-0.84 across decoder-only, encoder-only, and encoder-decoder models and confirming the predicted polarity reversal between HaluEval and MedHallu.

Significance. If the modeling of forced-scoring attention matrices as transport operators holds, the blindness theorem and converse bound provide a rigorous limit on symmetric spectral methods for direction detection, while the closed-form Cheeger landscape supplies an idealized-architecture benchmark that is falsifiable. The empirical polarity prediction and interpretable signal in transport features add practical value for hallucination diagnostics in LLMs. The work earns credit for the parameter-free derivation of the landscape and the explicit falsifiability of the two-axis diagnostic.

major comments (2)
  1. [Theory (Cheeger landscape derivation) and Experiments (forced-scoring setup)] The blindness theorem and Lipschitz converse are general operator results, but the specific claims that uniform causal attention satisfies φ ≥ 1/5 (n-independent) and window attention is O(w/n), together with the polarity prediction, rest on the load-bearing assumption that forced-scoring attention matrices behave as the transport operators whose symmetric component governs capacity according to the derived bipartite-Cheeger landscape. The manuscript must supply explicit justification or ablation showing that real heads from the tested models match the canonical architectures sufficiently closely; otherwise the floor and shape-difference claims do not transfer to the empirical diagnostic.
  2. [Converse theorem statement and application] The converse bound on transpose sensitivity by G assumes the diagnostic is Lipschitz; the manuscript does not report or bound the Lipschitz constant of the empirical φ diagnostic, which is required to make the bound operational for the two-axis method.
minor comments (3)
  1. [Empirical evaluation] Dataset details, number of heads/models, and exact length-control procedure should be stated explicitly rather than summarized.
  2. [Preliminaries] Notation for the asymmetry coefficient G and the precise definition of the degree-normalized operator should be introduced before the blindness theorem.
  3. [Results] The abstract reports AUROC ranges without error bars or statistical tests; these should be added to the results tables or text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and the recommendation of major revision. We address each point below with the strongest honest defense of the manuscript.

read point-by-point responses

  1. Referee: [Theory (Cheeger landscape derivation) and Experiments (forced-scoring setup)] The blindness theorem and Lipschitz converse are general operator results, but the specific claims that uniform causal attention satisfies φ ≥ 1/5 (n-independent) and window attention is O(w/n), together with the polarity prediction, rest on the load-bearing assumption that forced-scoring attention matrices behave as the transport operators whose symmetric component governs capacity according to the derived bipartite-Cheeger landscape. The manuscript must supply explicit justification or ablation showing that real heads from the tested models match the canonical architectures sufficiently closely; otherwise the floor and shape-difference claims do not transfer to the empirical diagnostic.

    Authors: The blindness theorem and the closed-form bipartite-Cheeger landscape, including the n-independent floor φ ≥ 1/5 for uniform causal attention and the O(w/n) scaling for window attention, are derived as exact results for the specified canonical architectures; these derivations do not invoke empirical attention matrices. The polarity prediction is a direct consequence of the shape-difference established in the landscape for those architectures. We agree, however, that the manuscript would benefit from explicit bridging analysis to support transfer of the benchmark interpretation to the forced-scoring empirical setting. We will add a new subsection that compares the degree-normalized symmetric components of real heads (from the tested models) against the canonical uniform-causal and window forms via total-variation distance and eigenvalue alignment, thereby providing the requested justification. revision: yes

  2. Referee: [Converse theorem statement and application] The converse bound on transpose sensitivity by G assumes the diagnostic is Lipschitz; the manuscript does not report or bound the Lipschitz constant of the empirical φ diagnostic, which is required to make the bound operational for the two-axis method.

    Authors: The converse theorem is stated for any Lipschitz diagnostic and supplies the general bound in terms of G. The empirical φ is the normalized bipartite Cheeger constant obtained from the landscape derivation; because the Cheeger constant is defined via a minimization over indicator functions and the underlying operator is degree-normalized, it is Lipschitz continuous with respect to the operator norm with an explicit constant that follows from the same variational characterization used to obtain the landscape. We will revise the manuscript to state this Lipschitz constant explicitly and to verify that the two-axis method remains within the regime where the converse bound is operational. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivations are self-contained with external empirical checks

full rationale

The paper derives a general blindness theorem for transpose-invariant spectral diagnostics of the symmetric component and a Lipschitz converse bound by the asymmetry coefficient G. It then derives a closed-form bipartite-Cheeger landscape specifically for canonical causal architectures, yielding architecture-specific floors (uniform causal φ ≥ 1/5, window O(w/n)). These are presented as mathematical results applied to the modeled attention operators. The resulting two-axis diagnostic produces a polarity prediction that is stated as falsifiable and is directly tested on held-out benchmarks (HaluEval vs. MedHallu) under length-controlled evaluation, returning 0.62-0.84 LC-AUROC with the predicted reversal. No self-definitional reductions, fitted parameters renamed as predictions, load-bearing self-citations, or ansatz smuggling are exhibited in the text. The modeling choice of forced-scoring matrices as transport operators is an explicit assumption whose downstream predictions are externally validated rather than true by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; the ledger is therefore minimal and provisional.

axioms (2)
  • domain assumption Attention matrices under forced scoring can be treated as transport operators
    The entire analysis begins from this modeling choice stated in the abstract.
  • domain assumption The bipartite-Cheeger landscape applies to canonical causal architectures
    Used to obtain the n-independent floor φ ≥ 1/5.

pith-pipeline@v0.9.1-grok · 5858 in / 1328 out tokens · 34249 ms · 2026-07-01T00:04:43.710531+00:00 · methodology

0 comments
read the original abstract

When a language model processes a hallucinated response, its attention routing tends to fail in one of two shapes: over-concentrating on a narrow set of positions, or spreading so diffusely that relevance is diluted, and the shape of the failure carries diagnostic signal. We study these shapes as a diagnostic characterization, computed from attention matrices under \emph{forced scoring} of benchmark-labeled responses rather than during live generation. A widely used family of spectral methods analyzes the symmetric component of the degree-normalized attention operator, which governs transport \emph{capacity}; we prove that every transpose-invariant spectral diagnostic of this operator is structurally \emph{orientation-blind} (it cannot distinguish an operator from its transpose, and therefore cannot detect information-flow direction), with a converse to the blindness theorem bounding any Lipschitz diagnostic's transpose sensitivity by the asymmetry coefficient $G$. Pairing this with a closed-form bipartite-Cheeger landscape for canonical causal architectures, we show that uniform causal attention satisfies an $n$-independent floor $\phi \ge 1/5$, while window attention pierces the floor as $O(w/n)$; failure modes are shape-different, not just value-different. This floor is an idealized-architecture benchmark, not an empirical attractor: the fraction of real attention heads that pierce it is itself an architectural signature. The resulting two-axis diagnostic ($\phi$ for capacity, $G$ for direction) yields a falsifiable polarity prediction: bottleneck- and diffuse-dominated benchmarks should exhibit opposite polarity. Under length-controlled evaluation, transport features retain interpretable signal (0.62-0.84 LC-AUROC) across the tested decoder-only, encoder-only, and encoder-decoder models, with polarity reversing as predicted between HaluEval and MedHallu.

Figures

Figures reproduced from arXiv: 2605.04893 by Diego Maniloff, Dominik Dahlem, Mac Misiura.

Figure 1
Figure 1. Figure 1: Conductance has a bounded healthy range; asymmetry detects temporal isolation. Each attention head defines bipartite transport between queries (Q) and keys (K). (a)–(c) Conductance spectrum: healthy attention occupies an optimal band; too low indicates bottleneck (a), too high indicates diffuse dilution (c). (a) Bottleneck: concentrated attention yields low ϕb, missing relevant context. (b) Healthy: select… view at source ↗
Figure 2
Figure 2. Figure 2: Conductance landscape: theory and empirics. (a) Theoretical ϕ(St) vs. cut location t/n for canonical causal architectures at n = 100, derived in closed form from Theorems 12 and 13. Uniform causal (blue, solid) is U-shaped with worst case at t ∗/n ≈ 0.32 and asymptotic floor u∞/(2+u∞) ≈ 0.36, strictly above the 1/5 Cheeger floor (Theorem 16). Window attention (red and orange, dashed) follows ϕ ≤ w/(n−t) → … view at source ↗
Figure 3
Figure 3. Figure 3: Per-layer transport profiles reveal distinct failure signatures. Conductance ϕˆ (a) and spectral norm σ2 (b) averaged across heads within each layer (Pythia-160M). Shaded bands show population ±1 std around dotted mean lines (green: factual, red: hallucinated from HaluEval). Bottleneck (HaluEval, hallucinated): uniformly depressed ϕˆ and elevated σ2 across all layers—the spectral gap is small everywhere, i… view at source ↗
Figure 4
Figure 4. Figure 4: Conductance–spectral norm scatter reveals regime-dependent polarity. Each point is one sample; axes show conductance ϕˆ and spectral norm σ2 averaged across all heads and layers (Pythia-160M, 500 subsampled per class for HaluEval). The near-perfect anti-correlation (ρ=−0.99) confirms the Cheeger inequality: low ϕˆ implies high σ2 (small spectral gap). (a) HaluEval: hallucinated samples cluster at low ϕˆ / … view at source ↗
Figure 5
Figure 5. Figure 5: Temporal isolation (G) is sparse but architecture- and position-encoding￾dependent. (a) LC-AUROC for G std across all 15 model-dataset combinations with 95% bootstrap CIs. Most configurations cluster near chance; Flan-T5 decoder/HaluEval (0.78) and Pythia/HaluEval (0.82) are exceptions. (b) Per-layer G profiles on HaluEval (mean ± 1 std ribbons). Flan-T5 decoder (solid): clear class separation; factual G d… view at source ↗
Figure 6
Figure 6. Figure 6: Scaling validation and aggregation crossover (HaluEval). (a) LC-AUROC across Pythia 70M–1.4B (same training data, varying parameter count) plus LLaMA 3.1 8B (cross-architecture; GQA, RoPE). Conductance features (σ2 std, ϕˆ CVaR75) retain signal at all scales; G std drops to chance above 410M. (b) Conductance aggregation profile: the dominant aggregation shifts from location (mean) at 70M to spread (std) at… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 1 internal anchor

  1. [1]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

    work page

  2. [2]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

    work page

  3. [3]

    ]xGT|!&rC, O G6Mߒ \)G ܧx־> y6Ŷ8Zc؜,F^ *R : j . Yl7 Qzд / X bZ8.JgRc `Z Q ƪmZ)QZ _ M5yi6ӻa ޙ W ^ brE^N flO 8슉<

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    work page arXiv 2025

  4. [4]

    Hallucination detection in LLMs using spectral features of attention maps

    Jakub Binkowski, Denis Janiak, Albert Sawczyn, Bogdan Gabrys, and Tomasz Jan Kajdanowicz. Hallucination detection in LLMs using spectral features of attention maps. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 24354--24385, Suzhou, China, 2025. Association for Computational Linguistics. doi:10.18...

    work page doi:10.18653/v1/2025.emnlp-main.1239 2025

  5. [5]

    A Lower Bound for the Smallest Eigenvalue of the Laplacian

    Jeff Cheeger. A Lower Bound for the Smallest Eigenvalue of the Laplacian . In Robert C. Gunning, editor, Problems in Analysis : A Symposium in Honor of Salomon Bochner , pages 195--199. Princeton University Press, Princeton, NJ, 1970. Princeton Legacy Library reprint: 2015, ISBN 978-1-4008-6931-2

    work page 1970

  6. [6]

    INSIDE : LLMs ' internal states retain the power of hallucination detection

    Chao Chen, Kai Liu, Ze Chen, Yi Gu, Yue Wu, Mingyuan Tao, Zhihang Fu, and Jieping Ye. INSIDE : LLMs ' internal states retain the power of hallucination detection. In The Twelfth International Conference on Learning Representations (ICLR 2024), 2024

    work page 2024

  7. [7]

    Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps

    Yung-Sung Chuang, Linlu Qiu, Cheng-Yu Hsieh, Ranjay Krishna, Yoon Kim, and James Glass. Lookback lens: Detecting and mitigating contextual hallucinations in large language models using only attention maps. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 1419--1436, Miami, Florida, USA, 2024. Association for...

    work page doi:10.18653/v1/2024.emnlp-main.84 2024

  8. [8]

    Fan R. K. Chung. Spectral Graph Theory, volume 92 of CBMS Regional Conference Series in Mathematics. American Mathematical Society, 1997. ISBN 978-0-8218-0315-8

    work page 1997

  9. [9]

    Fan R. K. Chung. Laplacians and the Cheeger Inequality for Directed Graphs . Annals of Combinatorics, 9 0 (1): 0 1--19, April 2005. ISSN 0219-3094. doi:10.1007/s00026-005-0237-z

    work page doi:10.1007/s00026-005-0237-z 2005

  10. [10]

    Attention is not all you need: Pure attention loses rank doubly exponentially with depth

    Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning Research, pages 2793--2803. PMLR, 2021

    work page 2021

  11. [11]

    Ettori, N

    Davide Ettori, Nastaran Darabi, Sina Tayebati, Ranganath Krishnan, Mahesh Subedar, Omesh Tickoo, and Amit Ranjan Trivedi. EigenTrack : Spectral activation feature tracking for hallucination and out-of-distribution detection in LLMs and VLMs , 2025. URL https://arxiv.org/abs/2509.15735

    work page arXiv 2025

  12. [12]

    Nature630, 625–630 (06 2024)

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy. Nature, 630 0 (8017): 0 625--630, June 2024. ISSN 1476-4687. doi:10.1038/s41586-024-07421-0

    work page doi:10.1038/s41586-024-07421-0 2024

  13. [13]

    Asymptotic fringe distributions for general families of random trees,

    James Allen Fill. Eigenvalue bounds on convergence to stationarity for nonreversible M arkov chains, with an application to the exclusion process. The Annals of Applied Probability, 1 0 (1): 0 62--87, 1991. doi:10.1214/aoap/1177005981

    work page doi:10.1214/aoap/1177005981 1991

  14. [14]

    The emergence of clusters in self-attention dynamics

    Borjan Geshkovski, Cyril Letrouit, Yury Polyanskiy, and Philippe Rigollet. The emergence of clusters in self-attention dynamics. In Advances in Neural Information Processing Systems, volume 36, pages 57026--57037, 2023

    work page 2023

  15. [15]

    Transformer Feed-Forward Layers Are Key-Value Memories

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5484--5495, 2021. doi:10.18653/v1/2021.emnlp-main.446

    work page internal anchor Pith review doi:10.18653/v1/2021.emnlp-main.446 2021

  16. [16]

    Golub and Charles F

    Gene H. Golub and Charles F. van Loan. Matrix Computations. Johns Hopkins University Press, 4th edition, 2013. ISBN 978-1-4214-0794-4

    work page 2013

  17. [17]

    Graham, Donald E

    Ronald L. Graham, Donald E. Knuth, and Oren Patashnik. Concrete Mathematics: A Foundation for Computer Science. Addison-Wesley, 2nd edition, 1994. ISBN 978-0-201-55802-9

    work page 1994

  18. [18]

    Higher dimensional discrete Cheeger inequalities

    Anna Gundert and May Szedl\'ak. Higher dimensional discrete Cheeger inequalities. Journal of Computational Geometry, 6 0 (2): 0 54--71, 2015. doi:10.20382/jocg.v6i2a4

    work page doi:10.20382/jocg.v6i2a4 2015

  19. [19]

    Horn and Charles R

    Roger A. Horn and Charles R. Johnson. Matrix Analysis . Cambridge University Press, 2nd edition, 2012. ISBN 978-0-521-83940-2. doi:10.1017/CBO9781139020411

    work page doi:10.1017/cbo9781139020411 2012

  20. [20]

    Improved Cheeger's inequality: Analysis of spectral partitioning algorithms through higher order spectral gap

    Tsz Chiu Kwok, Lap Chi Lau, Yin Tat Lee, Shayan Oveis Gharan, and Luca Trevisan. Improved Cheeger's inequality: Analysis of spectral partitioning algorithms through higher order spectral gap. In Proceedings of the Forty-Fifth Annual ACM Symposium on Theory of Computing (STOC), pages 11--20. ACM, 2013. doi:10.1145/2488608.2488611

    work page doi:10.1145/2488608.2488611 2013

  21. [21]

    Frustration index and Cheeger inequalities for discrete and continuous magnetic Laplacians

    Carsten Lange, Shiping Liu, Norbert Peyerimhoff, and Olaf Post. Frustration index and Cheeger inequalities for discrete and continuous magnetic Laplacians . Calculus of Variations and Partial Differential Equations, 54 0 (4): 0 4165--4196, 2015. doi:10.1007/s00526-015-0935-x

    work page doi:10.1007/s00526-015-0935-x 2015

  22. [22]

    Lee, Shayan Oveis Gharan, and Luca Trevisan

    James R. Lee, Shayan Oveis Gharan, and Luca Trevisan. Multiway spectral partitioning and higher-order Cheeger inequalities. Journal of the ACM, 61 0 (6): 0 1--30, 2014. doi:10.1145/2665063. Conference version in STOC 2012

    work page doi:10.1145/2665063 2014

  23. [23]

    Levin, Yuval Peres, and Elizabeth L

    David A. Levin, Yuval Peres, and Elizabeth L. Wilmer. Markov chains and mixing times . American Mathematical Society, 2006

    work page 2006

  24. [24]

    H alu E val: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models

    Junyi Li, Xiaoxue Cheng, Wayne Xin Zhao, Jian-Yun Nie, and Ji-Rong Wen. HaluEval : A large-scale hallucination evaluation benchmark for large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 6449--6464, Singapore, 2023. Association for Computational Linguistics. doi:10.18653/v1/2023.emnlp-main.397

    work page doi:10.18653/v1/2023.emnlp-main.397 2023

  25. [25]

    T ruthful QA : Measuring How Models Mimic Human Falsehoods

    Stephanie Lin, Jacob Hilton, and Owain Evans. TruthfulQA : Measuring how models mimic human falsehoods. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3214--3252, Dublin, Ireland, 2022. Association for Computational Linguistics. doi:10.18653/v1/2022.acl-long.229

    work page doi:10.18653/v1/2022.acl-long.229 2022

  26. [26]

    Lov\'asz

    L. Lov\'asz. Random walks on graphs: A survey. In D. Mikl\'os , V. T. S\'os , and T. Sz o nyi , editors, Combinatorics, Paul Erd o s is Eighty , volume 2, pages 353--398. J\'anos Bolyai Mathematical Society, Budapest, 1996

    work page 1996

  27. [27]

    Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9004--9017, 2023. doi:10.18653/v1/2023.emnlp-main.557

    work page doi:10.18653/v1/2023.emnlp-main.557 2023

  28. [28]

    Clustering by weighted cuts in directed graphs

    Marina Meila and William Pentney. Clustering by weighted cuts in directed graphs. In Proceedings of the 2007 SIAM International Conference on Data Mining, pages 135--144, 2007. doi:10.1137/1.9781611972771.13

    work page doi:10.1137/1.9781611972771.13 2007

  29. [29]

    Locating and editing factual associations in GPT

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in GPT . In Advances in Neural Information Processing Systems, volume 35, 2022

    work page 2022

  30. [30]

    FA ct S core: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen - tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore : Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 12076--12100, 2023. do...

    work page doi:10.18653/v1/2023.emnlp-main.741 2023

  31. [31]

    MedHallu : A comprehensive benchmark for detecting medical hallucinations in large language models

    Shrey Pandit, Jiawei Xu, Junyuan Hong, Zhangyang Wang, Tianlong Chen, Kaidi Xu, and Ying Ding. MedHallu : A comprehensive benchmark for detecting medical hallucinations in large language models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 2858--2873, Suzhou, China, 2025. Association for Computational Li...

    work page doi:10.18653/v1/2025.emnlp-main.143 2025

  32. [32]

    Ori Parzanchevski, Ron Rosenthal, and Ran J. Tessler. Isoperimetric inequalities in simplicial complexes. Combinatorica, 36 0 (2): 0 195--227, 2016. doi:10.1007/s00493-014-3002-x

    work page doi:10.1007/s00493-014-3002-x 2016

  33. [33]

    Journal of the American Statistical Association 89, 1255–1270

    James M. Robins, Andrea Rotnitzky, and Lue Ping Zhao. Estimation of regression coefficients when some regressors are not always observed. Journal of the American Statistical Association, 89 0 (427): 0 846--866, 1994. doi:10.1080/01621459.1994.10476818

    work page doi:10.1080/01621459.1994.10476818 1994

  34. [34]

    Mind the gap: A spectral analysis of rank collapse and signal propagation in attention layers

    Thiziri Nait Saada, Alireza Naderi, and Jared Tanner. Mind the gap: A spectral analysis of rank collapse and signal propagation in attention layers. In Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research. PMLR, 2025

    work page 2025

  35. [35]

    Approximate counting, uniform generation and rapidly mixing Markov chains

    Alistair Sinclair and Mark Jerrum. Approximate counting, uniform generation and rapidly mixing Markov chains. Information and Computation, 82 0 (1): 0 93--133, 1989. ISSN 0890-5401. doi:10.1016/0890-5401(89)90067-9

    work page doi:10.1016/0890-5401(89)90067-9 1989

  36. [36]

    LLM-Check : Investigating detection of hallucinations in large language models

    Gaurang Sriramanan, Siddhant Bharti, Vinu Sankar Sadasivan, Shoumik Saha, Priyatham Kattakinda, and Soheil Feizi. LLM-Check : Investigating detection of hallucinations in large language models. In Advances in Neural Information Processing Systems, volume 37, 2024

    work page 2024

  37. [37]

    2024 , issue_date =

    Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer : Enhanced transformer with rotary position embedding. Neurocomputing, 568: 0 127063, 2024. doi:10.1016/j.neucom.2023.127063

    work page doi:10.1016/j.neucom.2023.127063 2024

  38. [38]

    Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned

    Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5797--5808, 2019. doi:10.18653/v1/P19-1580

    work page doi:10.18653/v1/p19-1580 2019

  39. [39]

    A tutorial on spectral clustering.Stat

    Ulrike von Luxburg . A tutorial on spectral clustering. Statistics and Computing, 17 0 (4): 0 395--416, December 2007. ISSN 0960-3174. doi:10.1007/s11222-007-9033-z

    work page doi:10.1007/s11222-007-9033-z 2007

  40. [40]

    Jerry Wei, Chengrun Yang, Xinying Song, Yifeng Lu, Nathan Hu, Jie Huang, Dustin Tran, Daiyi Peng, Ruibo Liu, Da Huang, Cosmo Du, and Quoc V. Le. Long-form factuality in large language models. In Advances in Neural Information Processing Systems, volume 37, 2024

    work page 2024

  41. [41]

    Stabilizing transformer training by preventing attention entropy collapse

    Shuangfei Zhai, Tatiana Likhomanenko, Etai Littwin, Dan Busbridge, Jason Ramapuram, Yizhe Zhang, Jiatao Gu, and Josh Susskind. Stabilizing transformer training by preventing attention entropy collapse. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 40770--40803. PMLR, 2023

    work page 2023