Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Valid mathematical reasoning leaves a measurable spectral fingerprint in transformer attention, separable by a single training-free threshold with effect sizes up to d = 3.30.

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 12:59 UTC pith:APK77DHH

load-bearing objection Useful spectral idea, but the headline accuracies are circular: the paper relabels the classifier's own false positives and then reports accuracy on those corrected labels. the 4 major comments →

arxiv 2601.00791 v2 pith:APK77DHH submitted 2026-01-02 cs.LG cs.AIcs.CLcs.LO

Geometry of Reason: Spectral Signatures of Valid Mathematical Reasoning

Valentin No\"el This is my paper
classification cs.LG cs.AIcs.CLcs.LO
keywords spectral graph theorytransformer attentionmathematical reasoningproof validitytraining-free verificationhallucination detectioninduction headsgraph Laplacian
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 argues that a transformer's attention matrices, viewed as weighted graphs over tokens, carry a geometric stamp of whether the reasoning is logically coherent. Across seven models from four architectural families, four hand-defined spectral diagnostics separate valid from invalid formal proofs so sharply that a single threshold on one diagnostic reaches 85–96% accuracy. The signal is claimed to track 'Platonic validity'—mathematical soundness rather than acceptance by a proof-checking compiler—and ablating induction heads destroys the signature, suggesting a causal link. If right, this gives a training-free, architecture-aware primitive for detecting hallucinations and steering proof search. The main caveat is that the accuracy figures rest on manually corrected labels.

Core claim

The central discovery is that the spectral decomposition of attention graphs behaves differently for valid and invalid reasoning: valid proofs concentrate signal energy in low-frequency modes (lower high-frequency energy ratio), keep attention graphs better connected (higher Fiedler value), stay smoother in late layers, and distribute attention more evenly (higher spectral entropy). The effect sizes are extraordinarily large (up to Cohen's d = 3.30, p < 10^-116), and the discrimination works with a single threshold on one metric, with no training data or learned classifier. The method's key interpretive claim—its 'Platonic validity'—is that the spectral signature flags proofs the compiler re

What carries the argument

Each layer's attention matrix is symmetrized and aggregated across heads into a weighted undirected graph over tokens; the graph Laplacian's eigendecomposition yields four diagnostics: the Fiedler value (second-smallest eigenvalue, measuring algebraic connectivity), the High-Frequency Energy Ratio (share of hidden-state energy in high-frequency eigenvectors), spectral entropy (spread of energy across eigenvectors), and a normalized smoothness (Dirichlet energy deficit). These numbers, computed with zero learned parameters, are the carrier of the argument: a threshold on a single diagnostic at a chosen layer classifies proof validity, and the layer/metric that works best is itself an architec

Load-bearing premise

The reported accuracies and effect sizes assume the manually corrected 'Platonic validity' labels—where proofs the spectral classifier called valid but the formal prover rejected were re-labeled as valid—are the correct ground truth; if those relabelings are not independently audited, the headline numbers are partly circular.

What would settle it

Take the 33–51 re-labeled proofs, have an independent judge who is blind to the spectral verdict re-classify each compiler-rejected proof as mathematically sound or genuinely flawed, and measure agreement with the spectral classifier beyond chance (kappa). If agreement falls near zero, the 'Platonic validity' claim—and the accuracy numbers built on the corrected labels—collapses. A second falsifier: generate deliberately wrong proofs that still compile in the proof assistant; if the spectral threshold cannot separate them from hand-written valid proofs, the signal tracks compiler-style surface

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

If this is right

  • Single-threshold spectral screening could act as a cheap pre-filter in proof search, before expensive formal verification is run.
  • The same diagnostics could be deployed as a runtime monitor for hallucination detection in reasoning systems, since they require no labels or fine-tuning.
  • The 'Platonic validity' finding implies compiler rejection is not ground truth for reasoning quality; spectral labels could help build better training data for verifiers.
  • The architectural-dependent metric shift (global vs. sliding-window attention) means verification tools must be tuned per model family rather than assumed universal.
  • The causal ablation links the signature to induction-head circuits, suggesting the geometry tracks a known mechanism of in-context copying and composition.

Where Pith is reading between the lines

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

  • If the signature tracks logical coherence rather than stylistic surface, it should generalize to other formal languages and proof assistants; a natural test is whether thresholds calibrated on one formal language transfer to another without recalibration.
  • The 'inverse overfitting' result hints that high-dimensional supervised probes are learning label noise; spectral features might serve as a regularizing prior for learned verifiers, combining the best of both.
  • The mixture-of-experts attenuation ('sparsity penalty') suggests the spectral geometry is sensitive to routing entropy; a testable extension is whether deliberately enforcing focused routing improves both reasoning accuracy and the spectral signature.
  • The 'spectral crossover'—a synchronized extremum in all diagnostics at a fixed relative depth—could be used as a detector of where the model commits to a reasoning path; tracking its position might predict when chain-of-thought will fail.

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 / 5 minor

Summary. The paper proposes a training-free method for classifying the validity of mathematical proofs by computing spectral diagnostics (Fiedler value, high-frequency energy ratio, spectral entropy, smoothness) from transformer attention graphs and hidden-state signals. It reports large effect sizes (up to d = 3.30) and 85.9–94.9% single-threshold accuracy across seven models from four architectures on a MiniF2F/Lean subset, introduces a ``Platonic validity'' label-correction procedure, claims an induction-head causal mechanism, and reports an architectural shift for Sliding Window Attention. The abstract additionally claims a proof-search reranking result and a manual audit (kappa = 0.82) that do not appear in the body.

Significance. If the central claims were supported, the paper would describe a notable zero-label, architecture-aware reasoning verifier with potential applications in proof search and hallucination detection. The manuscript has genuine strengths: a reproducible public code repository, a nested cross-validation protocol, robustness controls for authorship and proof length, and explicit appendices with per-layer results. However, the headline quantitative claims currently rest on corrected labels whose construction is confounded with the classifier being tested, and key results advertised in the abstract are missing from the body. The significance of the work is therefore conditional on an independent audit and on the missing experiments being supplied.

major comments (4)
  1. [§5.3, Tables 2 and 21] The evaluation labels are not independent of the classifier. Section 5.3 relabels proofs that the spectral classifier marks valid but Lean rejects as ``Platonic valid''; Table 21 shows 33–51 such proofs per model. Table 2 then reports all main accuracies and effect sizes on these corrected labels. Because the relabeled examples are exactly the classifier's false positives under the original Lean labels, this mechanically removes known errors and inflates accuracy and Cohen's d. The manuscript states the relabeled proofs were ``overwhelmingly mathematically correct'' but provides no audit protocol, no kappa computation, and no example-level data; the kappa = 0.82, n = 51 audit appears only in the abstract. Without an independent, blinded audit of the relabeled set, the reported d ≈ 3.0 and 85–96% accuracies cannot be distinguished from a self-fulfilling prophecy.
  2. [Abstract vs. §6] The abstract claims that ``HFER reranking improves Best-of-16 Pass@1 by +4.4–6.6%, matching 98% of the AUC of fully supervised probes with zero labels.'' No corresponding experiment, dataset, table, or protocol appears anywhere in the body; Section 6 only lists proof search as a possible application. This is a central advertised result and must either be added with full experimental details or removed from the abstract.
  3. [§5.2, Figure 4] The ``causal ablation'' conclusion is not supported by the present controls. The authors zero out query projections of top-k induction heads and observe changes in spectral metrics, but there is no random-head ablation, no permutation baseline, and no statistical test comparing induction-head ablation to equally sized ablations of non-induction heads. Any head ablation could perturb attention enough to change these global metrics. The claim that the signature ``traces induction-head circuits'' requires these controls. Additionally, the text says ablated heads cause a ``rise in Fiedler value (loss of connectivity)'', but by Definition 4 a higher Fiedler value indicates stronger connectivity; the direction of the claimed effect is inconsistent.
  4. [§4.2 and Table 2] The reporting of accuracies is internally inconsistent. Section 4.2 says calibrated thresholds reach ``93–95% on the full dataset,'' while Table 2 reports 85.9–94.9%; the abstract says 85–96%. The relationship between the ``calibrated accuracy'' and the Table 2 accuracy is never made precise. Also, Table 2 reports pMW = 4.51 × 10^-66 for the d = 3.30 Phi result, while the abstract's p < 10^-116 belongs to a different model/metric (Qwen-0.5B entropy, d = 2.93). The abstract appears to combine the largest d and the smallest p from different rows, which is misleading.
minor comments (5)
  1. [Table 2 vs. Table 22] Cohen's d signs are inconsistent. Table 2 reports positive d for all models, but Table 22 gives negative d for HFER metrics (e.g., Llama-8B HFER L30 d = -3.00) because valid proofs have lower HFER. Please use signed d consistently or state that d is reported as absolute value.
  2. [Figure 2] The figure title ``Shape of Truth (Llama-3.1-8B)'' is misleading: panels (b)–(d) show Phi, Mistral, and Qwen. The title or caption should be corrected.
  3. [Eq. (6) and Algorithm 1] HFER's frequency cutoff K is defined as the median eigenvalue index, but Algorithm 1 uses m > N/2. It is unclear whether these are the same cutoff, and the text should state the exact definition used in all experiments.
  4. [Abstract vs. §5.3] If the manual audit (kappa = 0.82, n = 51) was actually performed, the protocol, annotation instructions, and example counts must be moved into Section 5.3; the current abstract-only citation is unverifiable.
  5. [§5.2] The direction of the Fiedler-value effect is described incorrectly: ``rise in Fiedler value (loss of connectivity)'' contradicts the definition of algebraic connectivity. Either the metric or the interpretation is wrong and needs correction.

Circularity Check

2 steps flagged

Label correction in §5.3 makes the main MiniF2F results circular: proofs the spectral classifier marks valid are relabeled as 'Platonic valid,' and Table 2 scores the classifier against these self-produced labels.

specific steps
  1. self definitional [Section 5.3 'Platonic Validity'; Section 4.1 label correction; Table 2 caption '(corrected labels)']
    "When the spectral method classified a proof as valid but the ground truth label indicated invalidity, manual inspection revealed that the spectral method was frequently correct: these proofs were mathematically sound but rejected by Lean due to technical failures. ... We therefore report all main results using corrected labels that reflect logical validity rather than compiler output."

    The category 'Platonic validity' is populated by proofs the spectral method already scored as valid; the corrected labels are therefore defined using the classifier being evaluated. Table 21 relabels 33–51 formerly-invalid proofs per model, i.e., exactly the classifier's false positives under the original Lean labels, before Table 2 reports accuracy, Cohen's d, and p-values on the corrected labels. Removing known errors before scoring mechanically raises accuracy and effect size. The only independent check mentioned, κ=0.82 (n=51), appears solely in the abstract with no audit protocol in the body.

  2. fitted input called prediction [Section 4.2 'Control 1: Model-generated valid vs. invalid'; Section 5.3]
    "We identify n = 16 'reclaimed' proofs, model-generated attempts that are semantically correct but rejected by Lean due to technical failures, and compare against n = 16 randomly sampled model-generated proofs with genuine logical errors. ... All seven models independently identified 33–51 compiler-rejected proofs as spectrally valid."

    The 'reclaimed' valid proofs in Control 1 come from the same relabeling procedure: proofs are called semantically correct because the spectral method flagged them valid despite Lean rejection. The control then reports d=1.30 separation between these spectrally-selected proofs and random invalid proofs; the separation is partially guaranteed by the selection rule. This control is offered as evidence that the signature tracks validity rather than authorship, but its 'valid' set is defined by the signature.

full rationale

The central quantitative claim is not self-contained: the 'corrected' ground-truth labels are produced by the same spectral classifier that is then scored against them. Section 4.1 and Table 21 show 33–51 compiler-rejected proofs per model are relabeled as valid because the spectral method flagged them; Section 5.3 says all main results use these corrected labels, and Table 2 reports effect sizes/accuracy on them. This removes the classifier's false positives before evaluation, so d≈2.09–3.30 and 85.9–94.9% accuracy are partly the classifier agreeing with itself. Nested CV and train/val/test splits also use the corrected dataset, so they do not provide independent validation. The 'Platonic validity' construct is thus definitional: validity is assigned by spectral agreement, then the spectral signature is said to track Platonic validity. The claimed manual audit (κ=0.82, n=51) appears only in the abstract, so no independent body evidence is provided. Some non-circular evidence exists (Control 2 perturbations d=1.10; MATH generalization d=0.78; transparent ablation of thresholds/features), so the score is not 10. Self-citations to Noël 2025a/b are used for speculative unification and auxiliary remarks, not as load-bearing support for the main spectral result. The headline formal-MiniF2F results, however, reduce by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

The central claim rests on a chain of modeling choices: attention as an undirected graph, hidden states as graph signals, and a label-construction procedure that is partly circular. The free parameters listed here are chosen by hand or fitted to labels.

free parameters (4)
  • classification threshold tau = per-model, e.g., HFER < 0.100 for Llama-8B
    The single-threshold classifier requires tau calibrated on labeled examples; raw thresholds do not transfer across models (Table 18).
  • metric + layer selection = e.g., HFER@L30, Smooth@L25, Entropy@L0
    Best of 160 metric-layer combinations chosen per model (Tables 2, 12); nested CV mitigates this, but the main table is full-data optimized.
  • HFER frequency cutoff K = median eigenvalue index / N/2
    Chosen by hand in Definition 2 and Algorithm 1 line 18; changes the metric's sensitivity.
  • reclassified proofs (label correction) = 33-51 per model (39 average)
    Proofs spectrally classified as valid but compiler-rejected are relabeled as 'Platonic valid'; this changes class balance from 154/300 to roughly 187-205/249-267 and inflates reported agreement.
axioms (6)
  • domain assumption Symmetrized attention matrix W = 1/2(A + A^T) defines a meaningful undirected graph of token interactions
    Section 3.1; discards directionality of attention, which may carry reasoning signal.
  • domain assumption Hidden-state columns are graph signals on the attention graph
    Section 3.2; no justification that the Laplacian eigenbasis of attention captures semantic coherence.
  • standard math Laplacian spectral theory (Fiedler value, Cheeger, GFT) applies to these empirical matrices
    Appendix A; standard mathematics, but requires symmetric nonnegative weights; softmax self-loops remain in the diagonal.
  • domain assumption Initial MiniF2F compiler labels and the model-generated invalid proof set are representative of invalid reasoning
    Section 4.1; the generation procedure for invalid proofs is not described.
  • ad hoc to paper 'Platonic validity' is a well-defined category usable for relabeling
    Section 5.3; introduced to reclassify classifier disagreements; no pre-registered operationalization.
  • ad hoc to paper Ablating query projections of top-k induction heads isolates the induction-head contribution
    Section 5.2; no random-head or permutation control for comparing spectral changes.
invented entities (1)
  • Platonic validity no independent evidence
    purpose: Conceptual label allowing compiler-rejected proofs to be counted as valid, making the spectral classifier appear correct
    No independent falsifiable handle; selection of 'Platonic valid' proofs is based on spectral classifier disagreements plus an unshown manual audit (abstract kappa = 0.82, n = 51 is absent from the main text).

pith-pipeline@v1.3.0-alltime-deepseek · 27661 in / 16739 out tokens · 288097 ms · 2026-08-03T12:59:03.058186+00:00 · methodology

0 comments
read the original abstract

Verifying whether a language model is genuinely reasoning or pattern-matching remains an open problem: learned verifiers are expensive, and output-based heuristics are brittle. We show that valid mathematical reasoning induces a measurable, training-free spectral signature in transformer attention. By treating each attention matrix as a weighted token graph, we extract four diagnostics: Fiedler value, High-Frequency Energy Ratio (HFER), spectral entropy, and smoothness, that require no learned parameters. Experiments across seven models from four architectural families yield effect sizes up to Cohen's $d = 3.30$ ($p < 10^{-116}$), enabling $85$--$96\%$ single-threshold classification accuracy. Two findings sharpen the interpretation. First, \emph{Platonic validity}: the spectral signal tracks logical coherence rather than compiler acceptance, proofs rejected for timeouts or missing imports are correctly classified as valid, a distinction confirmed by a manual audit ($\kappa = 0.82$, $n = 51$). Second, \emph{architectural determinism}: Sliding Window Attention shifts the discriminative feature from HFER to smoothness ($d = 2.09$, $p < 10^{-48}$), showing that attention design governs which spectral channel encodes reasoning quality. Causal ablation confirms the signature traces induction-head circuits. The method generalises to informal chain-of-thought ($d = 0.78$, $p < 10^{-3}$), and in proof search, HFER reranking improves Best-of-16 Pass@1 by $+4.4$--$6.6$\%, matching $98\%$ of the AUC of fully supervised probes with zero labels. Spectral graph analysis is a principled, architecture-aware primitive for reasoning verification.

Figures

Figures reproduced from arXiv: 2601.00791 by Valentin No\"el.

Figure 1
Figure 1. Figure 1: Overview of the spectral analysis pipeline. Attention matrices [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The “Shape of Truth” (Llama-3.1-8B). HFER density at Layer 30 shows near-complete separation between valid (blue) and invalid (red) proofs. Effect size d = 3.00, pMW = 9.40 × 10−64 . (L11) or late (L30) layers. • Qwen-0.5B: Uniquely, spectral entropy at L0 provides the strongest signal (d = 2.93, pt = 1.43 × 10−116). • Phi: Late-layer smoothness (L25) shows the largest effects (d = 3.30), with valid proofs… view at source ↗
Figure 3
Figure 3. Figure 3: Spectral Evolution of Reasoning. Llama-3.1-8B exhibits a phase transition in High-Frequency Energy Ratio (HFER). In early layers (0–10), valid and invalid proofs are indistinguishable. As reasoning deepens (Layers 15+), valid proofs (blue) maintain spectral smoothness, while hallucinations (red) disintegrate into high-frequency noise. pothesize that the “spectral smoothness” of valid reasoning is the topol… view at source ↗
Figure 4
Figure 4. Figure 4: Causal Link to Induction Mechanisms (Llama-3.1-8B). We ablated the top-k induction heads identified via in-context copying scores. Top Left: The Fiedler Value (graph connectivity) significantly increases (worsens) in the critical pre-computation layers (Layers 4–10) as heads are deactivated, confirming that spectral smoothness is physically maintained by active induction circuits. Layer 12 Crossover: All m… view at source ↗
Figure 5
Figure 5. Figure 5: Architectural Determinism of Validity. Comparing Llama-3.1- 8B (Global Attention) and Mistral-7B (Sliding Window Attention). Llama separates proofs via frequency (HFER), while Mistral separates them via local connectivity (Smoothness). This confirms that the spectral signature depends on the specific topology of the attention mechanism. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise spectral metrics (Combined) for Llama 3.2 1B Instruct. [PITH_FULL_IMAGE:figures/full_fig_p050_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Layer-wise spectral metrics (Combined) for Llama 3.2 3B Instruct. [PITH_FULL_IMAGE:figures/full_fig_p050_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Layer-wise spectral metrics (Combined) for Llama 3.1 8B Instruct. [PITH_FULL_IMAGE:figures/full_fig_p051_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Layer-wise spectral metrics (Combined) for Mistral 7B. [PITH_FULL_IMAGE:figures/full_fig_p051_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Layer-wise spectral metrics (Combined) for Phi 3.5 Mini Instruct. [PITH_FULL_IMAGE:figures/full_fig_p052_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Layer-wise spectral metrics (Combined) for Qwen 2.5 0.5B Instruct. [PITH_FULL_IMAGE:figures/full_fig_p052_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Layer-wise spectral metrics (Combined) for Qwen 2.5 7B Instruct. [PITH_FULL_IMAGE:figures/full_fig_p052_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Layer-wise spectral metrics (Combined) for Qwen 1.5 MoE A2.7b [PITH_FULL_IMAGE:figures/full_fig_p053_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Universality of Spectral Mechanism (Llama-3.2-3B). We replicate the induction head ablation on the smaller 3B architecture to test scale invariance. Consistent with larger models, we observe a distinct Spectral Crossover at Layer 11 (∼39% relative depth), where structural connectivity is maximized (Fiedler minimum). The immediate degradation of the spectral signature upon head ablation (Red dashed lines r… view at source ↗
Figure 15
Figure 15. Figure 15: Emergence in Lightweight Models (Qwen2.5-0.5B). We analyze the spectral signature in a highly compressed 0.5B parameter model (24 layers). Despite the reduced capacity, the mechanism remains distinct: the Spectral Crossover shifts to Layer 8 (∼33% relative depth), maintaining the characteristic synchronization of minimized Fiedler value and maximized HFER. The clear separation between the valid baseline (… view at source ↗
Figure 16
Figure 16. Figure 16: Architectural Robustness (Mistral-7B-v0.1). We extend the ablation analysis to the Mistral architecture to confirm cross-family validity. The Spectral Crossover is clearly observable at Layer 13 (∼40% relative depth), characterized by a synchronized collapse in Fiedler value (minimized to 0.35) and Spectral Entropy. The ablation of induction heads causes a significant divergence in the “buildup” phase (La… view at source ↗
Figure 17
Figure 17. Figure 17: Impact of Sparsity on Spectral Resolution (Qwen1.5- MoE). We evaluate the method on a Mixture-of-Experts architecture to test robustness against routing noise. While the spectral signature is inherently noisier due to dynamic expert selection (resulting in higher baseline entropy), the causal mechanism remains detectable at Layers 13–15. Ablating the top induction heads yields a consistent Fiedler value d… view at source ↗
Figure 18
Figure 18. Figure 18: Capacity Constraints and Relative Depth (Llama-3.2-1B). For this highly compressed architecture (16 layers), the Spectral Crossover shifts significantly deeper to Layer 9 (∼56% relative depth), suggesting that smaller models require a larger proportion of their computational depth to crystallize context into logic. While topological metrics (Fiedler) exhibit saturation due to limited head capacity, the HF… view at source ↗
Figure 19
Figure 19. Figure 19: Generalization to Synthetic-Heavy Architectures (Phi￾3.5-mini). We validate the spectral mechanism on Phi-3.5-mini (3.8B), a model optimized on high-density synthetic data. The baseline exhibits distinct volatility in the early layers (0–5), aligning with the “Phi-3 Anomaly” identified in (Noël, 2025b) where synthetic-heavy architectures show brittle early-layer topology. Despite this sensitivity, the cau… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Thermodynamic Signatures of Reasoning: Free-Energy and Spectral-Form-Factor Diagnostics for Hallucination Detection in Large Language Models

    cs.LG 2026-06 unverdicted novelty 7.0

    Introduces thermodynamic free-energy signatures and spectral form factors from attention Laplacians for hallucination detection, with stability proofs, expressiveness results, a PAC bound, and empirical AUROC gains ov...

Reference graph

Works this paper leans on

27 extracted references · 23 linked inside Pith · cited by 1 Pith paper

  1. [2]

    Metric Layer V alid Pert.p d Fiedler 6 0.455 0.4334.2×10 −6 0.90 Fiedler Last 0.547 0.5301.0×10 −5 0.90 HFER 6 0.099 0.1212.1×10 −7 −1.06 HFER Last 0.331 0.3781.9×10 −9 −1.10 Smooth 6 0.970 0.9861.2×10 −10 −1.17 Smooth Last 0.646 0.6271.4×10 −8 1.12 Entropy 6 1.999 1.8001.9×10 −8 1.13 Entropy Last 3.012 2.7901.2×10 −8 1.09 Table 10:Control 2: Human pertur...

    2022

  2. [7]

    URL https://arxiv.org/abs/2406. 04093. Geva, M., Bastings, J., Filippova, K., and Globerson, A. Dissecting recall of factual associations in auto-regressive language models.arXiv preprint arXiv:2304.14767,

    Pith/arXiv arXiv

  3. [8]

    How does GPT-2 compute greater- than? interpreting mathematical abilities in a pre-trained language model

    Hanna, M., Liu, O., and Variengien, A. How does GPT-2 compute greater- than? interpreting mathematical abilities in a pre-trained language model. arXiv preprint arXiv:2305.00586,

    Pith/arXiv arXiv

  4. [9]

    Measuring mathematical problem solving with the MATH dataset.arXiv preprint arXiv:2103.03874,

    29 Hendrycks, D., Burns, C., Kadavath, S., Arber, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. Measuring mathematical problem solving with the MATH dataset.arXiv preprint arXiv:2103.03874,

    Pith/arXiv arXiv

  5. [12]

    Let’s verify step by step.arXiv preprint arXiv:2305.20050,

    Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobbe, K. Let’s verify step by step.arXiv preprint arXiv:2305.20050,

    Pith/arXiv arXiv

  6. [13]

    and Tegmark, M

    Marks, S. and Tegmark, M. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets.arXiv preprint arXiv:2310.06824,

    Pith/arXiv arXiv

  7. [14]

    Progress measures for grokking via mechanistic interpretability.arXiv preprint arXiv:2301.05217,

    30 Nanda, N., Chan, L., Liberum, T., Smith, J., and Steinhardt, J. Progress measures for grokking via mechanistic interpretability.arXiv preprint arXiv:2301.05217,

    Pith/arXiv arXiv

  8. [15]

    Catching contamination before generation: Spectral kill switches for agents, 2025a

    Noël, V. Catching contamination before generation: Spectral kill switches for agents, 2025a. URLhttps://arxiv.org/abs/2511.05804. Noël, V. Training-free spectral fingerprints of voice processing in transformers, 2025b. URLhttps://arxiv.org/abs/2510.19131. Olsson, C., Elhage, N., Nanda, N., Joseph, N., DasSarma, N., Henighan, T., Mann, B., Askell, A., Bai,...

    arXiv

  9. [17]

    K., Bronstein, M

    Rusch, T. K., Bronstein, M. M., and Mishra, S. A survey on oversmoothing in graph neural networks.arXiv preprint arXiv:2303.10993,

    Pith/arXiv arXiv

  10. [19]

    A mechanistic interpretation of arithmetic reasoning in language models using causal mediation analysis

    31 Stolfo, A., Belinkov, Y., and Sachan, M. A mechanistic interpretation of arithmetic reasoning in language models using causal mediation analysis. arXiv preprint arXiv:2305.15054,

    Pith/arXiv arXiv

  11. [20]

    BERT rediscovers the classical NLP pipeline.arXiv preprint arXiv:1905.05950,

    Tenney, I., Das, D., and Pavlick, E. BERT rediscovers the classical NLP pipeline.arXiv preprint arXiv:1905.05950,

    Pith/arXiv arXiv 1905

  12. [21]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2023b

    Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., and Zhou, D. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2023b. Welleck, S., Liu, J., Lu, X., Hajishirzi, H., and Choi, Y. NaturalProver: Grounded mathematical proof generation with language models.Advances in Neural ...

    Pith/arXiv arXiv

  13. [22]

    Deepseek-prover-v1.5: Harnessing proof assistant feedback for reinforcement learning and monte-carlo tree search

    Xin, H., Wu, D., and Zhang, Y. Deepseek-prover-v1.5: Harnessing proof assistant feedback for reinforcement learning and monte-carlo tree search. arXiv preprint arXiv:2408.08152,

    Pith/arXiv arXiv

  14. [23]

    Can LLMs express their uncertainty? an empirical evaluation of confidence elicitation in LLMs.arXiv preprint arXiv:2306.13063,

    Xiong, M., Hu, Z., Lu, X., Li, Y., Fu, J., He, J., and Hooi, B. Can LLMs express their uncertainty? an empirical evaluation of confidence elicitation in LLMs.arXiv preprint arXiv:2306.13063,

    Pith/arXiv arXiv

  15. [24]

    Trans- formers as graph neural networks with global attention.arXiv preprint arXiv:2305.13926,

    32 Yang, C., Wang, Z., Zhu, Z., Xhonneux, L.-P., and Tang, J. Trans- formers as graph neural networks with global attention.arXiv preprint arXiv:2305.13926,

    Pith/arXiv arXiv

  16. [25]

    M., and Polu, S

    Zheng, K., Han, J. M., and Polu, S. MiniF2F: A cross-system benchmark for formal olympiad-level mathematics.arXiv preprint arXiv:2109.00110,

    Pith/arXiv arXiv

  17. [26]

    bottleneck

    34 (iii) Direct computation: x⊤Lx=x ⊤Dx−x ⊤W x= X i dix2 i − X i,j Wijxixj (14) = 1 2  X i dix2 i −2 X i,j Wijxixj + X j djx2 j   (15) = 1 2 X i,j Wij(x2 i −2x ixj +x 2 j ) = 1 2 X i,j Wij(xi −x j)2 (16) (iv) If G has k connected components, we can constructk linearly in- dependent vectors constant on each component, all in the null space of L. (v) Th...

    1973

  18. [1973]

    Scaling and evaluating sparse autoencoders.arXiv preprint arXiv:2406.04093,

    Gao, L., laTour, T.D., Tillman, H., Goh, G., Troll, R., Radford, A., Sutskever, I., Leike, J., and Wu, J. Scaling and evaluating sparse autoencoders.arXiv preprint arXiv:2406.04093,

    Pith/arXiv arXiv

  19. [2013]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314,

    Snell, C., Lee, J., Kumar, A., and Levine, S. Scaling llm test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314,

    Pith/arXiv arXiv

  20. [2014]

    Discovering latent knowledge in language models without supervision.arXiv preprint arXiv:2212.03827,

    Burns, C., Ye, H., Klein, D., and Steinhardt, J. Discovering latent knowledge in language models without supervision.arXiv preprint arXiv:2212.03827,

    Pith/arXiv arXiv

  21. [2018]

    and Chandrasekaran, B

    Pacchiardi, L. and Chandrasekaran, B. Detecting deception in large language models via geometric probing.arXiv preprint arXiv:2501.00234,

    Pith/arXiv arXiv

  22. [2019]

    Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221,

    Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., et al. Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221,

    Pith/arXiv arXiv

  23. [2021]

    and Manning, C

    Hewitt, J. and Manning, C. D. A structural probe for finding syntax in word representations. InProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics, pp. 4129–4138,

    2019

  24. [2022]

    E., and Cohan, A

    Beltagy, I., Peters, M. E., and Cohan, A. Longformer: The long-document transformer.arXiv preprint arXiv:2004.05150,

    Pith/arXiv arXiv 2004

  25. [2023]

    Generating long sequences with sparse transformers

    Child, R., Gray, S., Radford, A., and Sutskever, I. Generating long sequences with sparse transformers. InarXiv preprint arXiv:1904.10509,

    Pith/arXiv arXiv 1904

  26. [2024]

    K., et al

    28 El, B., Joshi, C. K., et al. Towards mechanistic interpretability of graph transformers via attention graphs.arXiv preprint arXiv:2502.12352,

    Pith/arXiv arXiv

  27. [2025]

    and Mitchell, T

    Azaria, A. and Mitchell, T. The internal state of an LLM knows when it’s lying.arXiv preprint arXiv:2304.13734,

    Pith/arXiv arXiv