Pith. sign in

REVIEW 4 major objections 4 minor 24 references

This paper proves that classifying LLM responses by token-embedding dynamics becomes exponentially more reliable with response length, and that this reliability transfers across embedding models when the embedding spaces are approximately l

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 00:44 UTC pith:WZLYMDJW

load-bearing objection The paper targets a genuine gap but the advertised guarantees do not hold as written: Result 1's inequality is false, Result 2's rate is overclaimed, and Result 3's proof has a missing step. the 4 major comments →

arxiv 2607.28667 v1 pith:WZLYMDJW submitted 2026-07-25 cs.LG cs.AImath.DS

Guarantees on Dynamical System Distinguishability for LLM Token Generation

Mohamed Akrout , Dan Wilson This is my paper
classification cs.LG cs.AImath.DS
keywords dynamical systemsLLM token embeddingsKoopman operatorhallucination detectionbinary hypothesis testingexponential concentrationcross-embedding transfersingle-response classification
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 formalizes LLM response classification as a binary hypothesis test between two stochastic linear dynamical systems fitted to token embeddings. It establishes three results: static marginal statistics of embeddings can be nearly identical across classes even when dynamics differ, so static classifiers hit an accuracy floor; a residual-based classifier has misclassification error that decays exponentially with response length, with the rate set by a dynamical discriminability quantity; and this discriminability transfers across embedding models when the embedding spaces are approximately linearly related. The practical upshot is a theoretical explanation for why single-response, dynamics-based hallucination detection works and how many tokens are needed for a target error rate.

Core claim

The central discovery is that the distinguishability of two LLM response regimes is governed by a single scalar, the dynamical discriminability delta^2, defined as the expected squared norm of the model-mismatch signal P(A_c - A_h)z_k under the stationary distribution. Given this quantity, the paper proves that the residual-based classifier's error decays like exp(-c L delta^2 / sigma^4), that any classifier ignoring token dynamics is bounded below by a near-random accuracy floor, and that cross-embedding transfer is possible with accuracy bounded by the smallest singular value of the intertwining map minus a penalty proportional to the intertwining defect. If correct, these results turn an

What carries the argument

The key object is the dynamical discriminability delta^2 = E[||P(A_c - A_h)z_k||^2], which measures the expected per-step disagreement between the two fitted linear dynamical system models when one is used to predict the other's trajectories. It combines the operator difference between the two Koopman-style approximations with the state distribution they act on. The proofs also rely on an approximate intertwining condition, G^(theta2)(q) approximately T G^(theta1)(q), which quantifies when two embedding spaces are affinely related and carries the cross-embedding transfer bound.

Load-bearing premise

The cross-embedding transfer guarantee rests on the assumption that the two embedding models are approximately linearly related (G2 approximately T G1 + r with small squared residual chi^2), which is stated without evidence that real LLM embeddings satisfy it.

What would settle it

Measure the residual-based classifier's error on a fixed LLM pair as a function of sequence length; if, for a model pair with estimated delta^2 > 0 and bounded noise, the error does not decay exponentially with L, or fails to reach the predicted L*, the central exponential-separability claim is falsified. Alternatively, find two embedding models with measured chi^2 near zero but chance-level transfer accuracy, which would refute Result 3.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If correct, classification error for token-dynamics classifiers falls exponentially with response length, so longer generations are disproportionately easier to classify.
  • The token complexity L* = O(sigma^4_max,c (delta^2 + beta_c)^-2 log(1/epsilon)) gives a quantitative rule for how many tokens are needed to reach a target error, explaining why larger embedding models need shorter sequences.
  • Static, marginal-distribution classifiers cannot beat the accuracy floor derived in Result 1, so the dynamics information is necessary, not optional.
  • Cross-embedding transfer works only when the two embedding spaces are approximately affinely related; the bound predicts which transfer directions will fail based on the intertwining defect chi^2.
  • The proof framework suggests a principled objective: choose the SVD projection and lifting that maximize delta^2, rather than variance.

Where Pith is reading between the lines

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

  • A natural extension the authors leave implicit is an algorithm to jointly optimize the embedding projection and Koopman operators to maximize delta^2; the paper's bounds make this objective well-defined.
  • The intertwining assumption is strong: real embedding spaces may be nonlinearly related, and the residual r(q) is likely to depend on the token q. A testable prediction is that transfer accuracy should degrade as the measured chi^2 grows, and that pairs with similar geometric density may transfer in only one direction.
  • Since the bounds are stated for linear dynamics and Gaussian noise, the exponential rate may not hold for nonlinear Koopman approximations or heavy-tailed noise; the scaling of error with L under those relaxations is an open question.
  • One could use the bound to design a cheap pretest: fit the two systems on a small sample, estimate delta^2 and sigma^2, and predict whether a given embedding pair will transfer before committing to full training.

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

4 major / 4 minor

Summary. The paper formalizes LLM response classification by modeling token embeddings as trajectories of two stochastic linear dynamical systems (H_c and H_h). It defines a dynamical discriminability quantity δ² based on the operator difference and the stationary state distribution, and states three results: (1) Result 1 gives a total-variation bound (Eq. 16) intended to show that marginal-distribution classifiers have a fundamental accuracy floor; (2) Result 2 gives an exponential bound on the misclassification probability of the residual-based classifier (Eq. 19), with Corollary 1 providing a √L separability growth and a corresponding minimum sequence length (Eqs. 20–21); (3) Result 3 gives a cross-embedding transfer lower bound under an approximate intertwining assumption (Eq. 25). Empirical illustrations on the HaluEval dataset are provided in Figs. 2, 4, and 5.

Significance. If the results were correct, the paper would supply a useful theoretical justification for single-response, dynamics-based hallucination detection and would predict how classification error scales with response length and embedding geometry. The conceptual distinction between marginal and dynamical information (Result 1) and the decomposition of the residual gap into a signal term and a noise cross-term (Appendix B, Eq. 36) are clean and pedagogically valuable. The empirical √L scaling in Fig. 4 is suggestive. However, the central theorems are not established: Result 1's inequality is false, Result 2's exponential rate is overstated by an order of magnitude, Corollary 1's derivation is not rigorous, and Result 3's proof relies on an undefined residual and an unjustified spectral bound. The paper does not currently deliver the advertised guarantees.

major comments (4)
  1. [Result 1 / Appendix A, Eq. (16)] The stated total-variation bound is false. Take d=1, P=I, Γ_c=1, Γ_h=0.1. Then Γ_c^{-1/2}(Γ_h-Γ_c)Γ_c^{-1/2} = -0.9, so Eq. (16) gives dTV ≤ 0.45. The actual total variation between N(0,1) and N(0,0.1) is approximately 0.50, contradicting the inequality. The error originates in the bound DKL_cov ≤ 1/4||Γ_c^{-1/2}(Γ_c-Γ_h)Γ_c^{-1/2}||²_F at Eq. (28), which fails when the standardized covariance difference has negative eigenvalues: for the example above the left side is 3.35 while the right side is 0.2025. Since Eq. (17) is the formal basis for the claimed classification accuracy floor, Result 1 is invalid as stated. The existence claim in the final sentence ('there exist pairs ... with ∥∆∥_F bounded away from zero ... arbitrarily small') may still be true, but it does not follow from the false bound.
  2. [Result 2 / Appendix B, Eqs. (43)–(44)] The exponential bound in Eq. (19) is inconsistent with the paper's own estimates. Eq. (43) lower-bounds µ_L by a quantity of order √(L-1)·(δ²+β_c) (all other factors are L-independent), so substituting into Eq. (19) gives an exponent of order L² for fixed η. But Appendix C itself shows Var(∆E|H_c)=O(1) while E[∆E|H_c]=Θ(√L), so the deviation µ_L-η is of order √L; for a statistic built from sums of L i.i.d. terms, such large deviations typically have rate Θ(L), not Θ(L²). The step 'The same concentration applied to the fluctuations of ∆E around µ_L' in Appendix B is not a derivation; the preceding inequalities (39)–(40) control Sh-Sc, not √Sh-√Sc. Moreover, Eq. (43) is not a valid statement about E[∆E|H_c]: it uses a high-probability bound on the denominator and then treats that bound as the expectation. Because Result 2 is the paper's main quantitative claim, this is a load-bearing error
  3. [Corollary 1 / Appendix C, Eqs. (20)–(21)] The claimed scaling S(L)=Θ(√L·(δ²+β_c)/σ²_max,c) is not derived by the proof. Appendix C estimates the mean as Θ(√L(δ²+β_c)/√(dσ²_max,c+||Pm_c||²)) and the standard deviation as Θ(√(σ²_max,cδ²+||Pm_c||²R²λmax(G_c))/√(dσ²_max,c+||Pm_c||²)). Taking the ratio does not yield the expression in Eq. (20) in general; even in the simple case β_c=0, P_m=0, the ratio is Θ(√(Lδ²/σ²_max,c)) rather than Θ(√L·δ²/σ²_max,c). The delta method is invoked without checking the required regularity or uniform bounds, and Eq. (52) simply asserts that the misclassification probability is exp(-c S(L)²). Thus the minimum-sequence-length formula (21) is unsupported.
  4. [Result 3 / Appendix D, Eq. (25)] The proof of Result 3 does not deliver the claimed bound. In Eq. (54) a residual ρ_k is introduced but its explicit form is never given, and the key estimate E[||ρ_k||²] ≤ C2 χ²(||∆(θ1)||²λmax(G_c^{(θ2)})+δ²(θ1)) is asserted in Eq. (55) without proof. More fundamentally, Assumption 3 relates raw embeddings via G^{(θ2)}_2(q)=T G^{(θ1)}_2(q)+r(q), while the Koopman operators act on lifted observables z_k; the paper never specifies a map between the lifted spaces or verifies that A^{(θ1)} can be applied to z^{(θ2)}_k. The derivation treats P^{(θ2)}A^{(θ1)}z^{(θ2)}_k as well-defined despite the dimension mismatch. Eq. (26) then substitutes δ²_{θ1→θ2} into Result 2 without re-deriving concentration for the cross-embedding test statistic. These gaps are load-bearing for the advertised transfer guarantee.
minor comments (4)
  1. [General] There are several typos and grammatical errors: 'fundimental', 'classifers', 'probabilty', 'A similar bound' instead of 'A similar bound holds', and 'It has been show' instead of 'It has been shown'. The manuscript should be carefully proofread.
  2. [Fig. 4] The dashed lines K√L are empirical fits with a free constant K, so the plot only checks the √L exponent. Report the fitted constants, provide error bars, and compare against alternative scalings (e.g., log L or L) to make the claimed prediction falsifiable.
  3. [Fig. 5] The reported intertwining defects (χ²≈8.8×10^6 and 4.8×10^6) are many orders of magnitude larger than the other terms appearing in Eq. (25). The qualitative interpretation is plausible, but the figure does not quantitatively validate the bound; a numerical evaluation of the entire right-hand side of Eq. (25) is needed.
  4. [Result 1, notation] The sentence 'Defining ∥·∥ as the Mahalanobis distance' is confusing because the left-hand side of Eq. (16) uses a Frobenius norm. The notation should be clarified: the first term uses a matrix norm and the second uses a Mahalanobis norm.

Circularity Check

0 steps flagged

No circular derivation: Results 1-3 are conditional theorems from explicit assumptions; self-citations are motivational, not load-bearing. The main risk is a proof gap in the exponential rate, not circularity.

full rationale

The derivation chain is not circular. Results 1-3 are stated as theorems from explicit assumptions (stability, Gaussian noise, approximate intertwining) and definitions (δ^2, β_c, δ^2_{θ1→θ2}); no fitted parameter is renamed as a prediction. Figure 4 fits a K√L curve to held-out estimates of S(L) with a free constant K; the predicted √L scaling is being tested, not manufactured by the fit. Wilson & Akrout (2026) is cited for the empirical LLM classification phenomenon and for the cross-embedding transfer observation, but no theorem proof depends on that citation for its mathematical content; the self-citation is motivational/empirical only and therefore not load-bearing. Assumption 3 is openly stated and empirically probed in Figure 5, so the transfer result is a conditional theorem rather than a hidden reuse of the conclusion. The main substantive concern is a proof gap/inconsistency, not circularity: in Appendix B the step from Eq. (43) to Eq. (44) is asserted ('The same concentration applied to the fluctuations of ∆E around µ_L') rather than derived, and the paper's own estimates (µ_L = Θ(√L), Var(∆E) = O(1)) imply a tail exp(−Θ(L)), whereas Eq. (19)/(44) gives an exp(−Θ(L^2)) bound. That affects correctness of the advertised rate, but it does not reduce the result to its inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

No new physical entities are postulated. The paper's free parameters are limited to post-hoc empirical fit constants; the mathematical results depend on the stability, Gaussianity, and linear-DS assumptions, plus the ad hoc intertwining assumption for transfer.

free parameters (1)
  • Empirical fit constants K (Fig. 4) = not reported
    Dashed lines of the form K√L are fit to the estimated separability factor to visually confirm the √L scaling. This is post-hoc validation of a corollary, not a parameter in the theorems.
axioms (5)
  • domain assumption Assumption 1: both systems are stable, ρ(A_c) < 1 and ρ(A_h) < 1
    Stability is needed for the existence and uniqueness of stationary distributions and for the geometric mixing used in the concentration arguments (Section 3.2).
  • domain assumption Assumption 2: i.i.d. Gaussian noise with positive-definite covariance
    Gaussianity makes the state z_k Gaussian, enabling Hanson-Wright and sub-exponential concentration bounds in Appendices B and C. The empirical justification in Fig. 2 is a fit to residuals, not a proof that the true noise is Gaussian.
  • domain assumption Token embedding sequences are well described by a finite-dimensional stochastic linear dynamical system
    The entire framework replaces the actual LLM generation process with a linear Koopman approximation; the theorems are about the fitted model, not the true generative process.
  • ad hoc to paper Assumption 3: approximate intertwining with bounded defect χ²
    This is introduced solely to obtain the cross-embedding transfer bound, and there is no independent evidence that different LLM embedding spaces are affinely related.
  • standard math The observable y_k is obtained by projecting raw embeddings onto the top-d SVD modes
    This is the EDMD setup inherited from Williams et al. (2015); it fixes the observable space but is part of the model rather than an independently justified design choice.

pith-pipeline@v1.3.0-alltime-deepseek · 14858 in / 19017 out tokens · 171059 ms · 2026-08-03T00:44:21.987755+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Guarantees on Dynamical System Distinguishability for LLM Token Generation." pith.science (2026). https://pith.science/paper/WZLYMDJW

@misc{pith2026260728667,
  author       = {Pith},
  title        = {Pith review of: Guarantees on Dynamical System Distinguishability for LLM Token Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WZLYMDJW}},
  note         = {Machine review of arXiv:2607.28667}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent work has shown that classifying large language models (LLMs)' responses can be distinguished by modeling token embeddings as trajectories of a black-box dynamical system (DS) and comparing prediction residuals of two DSs. Despite the empirical success of this dynamical approach, a theoretical understanding of why it works, how well it scales as a function of the token sequence, and when it transfers across embedding models remains lacking. We address these questions by formalizing the classification task as a binary hypothesis test between two stochastic linear DSs. We show that the total variation distance between the stationary marginal distributions of the two DSs can be arbitrarily small even when the dynamics differ substantially, which provides a fundamental accuracy floor for any classifier that ignores token dynamics. We then show that the misclassification probability of DS-based classification decays exponentially in the sequence length $L$, with the decay governed by a dynamical discriminability quantity $\delta^2$ that captures the spectral distance between the two DSs. We also characterize cross-embedding generalization by introducing an approximate intertwining condition between embedding models and establishing a lower bound on the transferable discriminability in terms of the intertwining map's smallest singular value. Together, these results explain the empirical performance of DS-based classification and motivate further investigation into using DS theory to analyze AI systems, in contrast to the more common approach of using AI to model dynamical systems.

Figures

Figures reproduced from arXiv: 2607.28667 by Dan Wilson, Mohamed Akrout.

Figure 1
Figure 1. Figure 1: The two approaches to analyze the properties of LLM [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Extended DMD is applied to the HaluEval dataset Li e [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Logical dependencies between the key DS variables [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Estimates of the separability factor (solid line) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Empirical validation of cross-embedding transfe [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

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

24 extracted references · 1 linked inside Pith

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  4. [4]

    Automatica , volume=

    Online change points detection for linear dynamical systems with finite sample guarantees , author=. Automatica , volume=. 2025 , publisher=

    2025

  5. [5]

    Annual Review of Control, Robotics, and Autonomous Systems , volume=

    Fault Diagnosis in Dynamical Systems: Geometric Interpretation and Tractable Algorithms , author=. Annual Review of Control, Robotics, and Autonomous Systems , volume=. 2025 , publisher=

    2025

  6. [6]

    M. O. Williams and I. G. Kevrekidis and C. W. Rowley , journal=. A data--driven approximation of the. 2015 , publisher=

    2015

  7. [7]

    2020 , publisher=

    Time series analysis , author=. 2020 , publisher=

    2020

  8. [8]

    1980 , publisher=

    Linear systems , author=. 1980 , publisher=

    1980

  9. [9]

    Canadian Journal of Mathematics , volume=

    On the maximum principle of Ky Fan , author=. Canadian Journal of Mathematics , volume=. 1957 , publisher=

    1957

  10. [10]

    arXiv preprint arXiv:2605.05134 , year=

    Low-Cost Black-Box Detection of LLM Hallucinations via Dynamical System Prediction , author=. arXiv preprint arXiv:2605.05134 , year=

    Pith/arXiv arXiv

  11. [11]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  12. [12]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  13. [13]

    Zhang, Jiaxin and Li, Zhuohang and Das, Kamalika and Malin, Bradley and Kumar, Sricharan , booktitle=

  14. [14]

    Findings of the association for computational linguistics: ACL 2024 , pages=

    Chain-of-verification reduces hallucination in large language models , author=. Findings of the association for computational linguistics: ACL 2024 , pages=

    2024

  15. [15]

    Advances in neural information processing systems , volume=

    Locating and editing factual associations in gpt , author=. Advances in neural information processing systems , volume=

  16. [16]

    1999 , edition=

    System Identification: Theory for the User , author=. 1999 , edition=

    1999

  17. [17]

    IEEE Control Systems Letters , year=

    Data-driven fault isolation in linear time-invariant systems: a subspace classification approach , author=. IEEE Control Systems Letters , year=

  18. [18]

    Conference On Learning Theory , pages=

    Learning without mixing: Towards a sharp analysis of linear system identification , author=. Conference On Learning Theory , pages=. 2018 , organization=

    2018

  19. [19]

    International Conference on Machine Learning , pages=

    Near optimal finite time identification of arbitrary linear dynamical systems , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  20. [20]

    Advances in neural information processing systems , volume=

    Neural ordinary differential equations , author=. Advances in neural information processing systems , volume=

  21. [21]

    Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=

    Applied koopmanism , author=. Chaos: An Interdisciplinary Journal of Nonlinear Science , volume=. 2012 , publisher=

    2012

  22. [22]

    Annual review of fluid mechanics , volume=

    Analysis of fluid flows via spectral properties of the Koopman operator , author=. Annual review of fluid mechanics , volume=. 2013 , publisher=

    2013

  23. [23]

    Hanson-Wright inequality and sub-Gaussian concentration , author=

  24. [24]

    Li, Junyi and Cheng, Xiaoxue and Zhao, Wayne Xin and Nie, Jian-Yun and Wen, Ji-Rong , booktitle=