Technical Summary: Self-Attention as Transport: Limits of Symmetric Spectral Diagnostics

Problem Statement

Large language models (LLMs) hallucinate through distinct routing pathologies: bottleneck failures, where attention over-concentrates on a narrow set of positions, and diffuse failures, where attention spreads so thinly that relevance is diluted. While these modes appear quantitatively similar under standard spectral lenses, they are mechanistically opposite. Existing spectral diagnostics (e.g., LLM-Check, EigenTrack, LapEigvals) often fail to distinguish them or conflate capacity failures with directional failures. Furthermore, many spectral features inherit confounding dependencies on sequence length, obscuring the true structural signal of attention routing. This paper seeks to characterize the structural boundaries of what attention-based diagnostics can and cannot measure, specifically addressing the limitations of symmetric spectral methods in detecting information-flow direction.

Methodology

The authors formalize attention heads as degree-normalized bipartite transport operators ($M$) between queries and keys. This representation yields a scale-invariant operator whose spectral properties encode routing quality. The methodology rests on three pillars:

1. Orthogonal Decomposition: The transport operator $M$ is decomposed under the Hilbert–Schmidt inner product into a symmetric component ($M_{sym}$) governing transport capacity and an antisymmetric component ($M_{asym}$) governing transport direction.
2. Spectral Diagnostics:
   • Conductance ($\phi$): Derived from the symmetric component via the Cheeger inequality, $\phi$ measures the efficiency of information transport. Low $\phi$ indicates bottlenecks; high $\phi$ indicates diffuse mixing.
   • Asymmetry Coefficient ($G$): Defined as the normalized Frobenius distance of $M$ to the symmetric subspace ($G = \|M_{asym}\|_F / (\|M\|_F + \epsilon)$), $G$ serves as the minimal complement required to recover directional information lost by symmetric summaries.
   • Spectral Gap ($1-\sigma_2$): The second singular value of $M$ acts as a proxy for conductance.
3. Length-Controlled Evaluation: To address confounding, the authors introduce a Length-Controlled AUROC (LC-AUROC) protocol. This involves stratifying samples by response length into bins, residualizing scores against length within bins, and aggregating via pair-weighted AUROC. This isolates structural routing signals from length-dependent artifacts.
4. Theoretical Benchmarks: The paper derives closed-form conductance landscapes for canonical causal architectures (uniform causal, window attention) to establish architectural "floors" and expected failure-mode shapes.

Key Contributions

1. Structural Limits of Symmetric Diagnostics (Orientation Blindness)

The paper proves Theorem 4 (Orientation Blindness): Any spectral diagnostic depending solely on singular values or the symmetric component of the attention operator is transpose-invariant. Consequently, such diagnostics are structurally "orientation-blind" and cannot distinguish an operator from its transpose, rendering them incapable of detecting information-flow direction. A quantitative converse (Proposition 6) establishes that any Lipschitz diagnostic sensitive to direction must depend on the antisymmetric component, with sensitivity bounded by $\|M_{asym}\|_F$.

2. Two-Sided Conductance Theory

Unlike metrics detecting only one failure mode, the conductance framework provides a two-sided diagnostic:

• Bottlenecks: Low conductance ($\phi$) certifies over-concentrated routing.
• Diffuse Mixing: Anomalously high conductance (relative to architectural baselines) certifies under-concentrated routing.
  The authors derive a closed-form bipartite-Cheeger landscape for uniform causal attention, proving an $n$-independent floor of $\phi \ge 1/5$ with a worst-case cut at $t^*/n \approx 0.32$. Window attention is shown to pierce this floor as $O(w/n)$.

3. Degree Sufficiency and Coupling Structure

Theorem 2 (Degree Sufficiency) formalizes when coupling structure contributes to spectral properties beyond degree heterogeneity. It establishes that for near-regular attention ($\kappa \approx 1$), the spectral gap is fully accounted for by degree distributions. Only when degree heterogeneity deviates significantly does the coupling structure provide independent signal.

4. Empirical Polarity Prediction

The framework yields a falsifiable prediction: bottleneck-dominated and diffuse-dominated benchmarks should exhibit opposite polarity in conductance scores.

• HaluEval: Hallucinations cluster in the low-conductance (bottleneck) regime.
• MedHallu: Hallucinations cluster in the high-conductance (diffuse) regime.
  This polarity reversal is observed empirically under length-controlled evaluation, validating the two-sided theory.

Empirical Results

• Diagnostic Performance: Under LC-AUROC, transport features retain interpretable signal (0.62–0.84) on models up to 8B parameters.
  • Conductance ($\phi$): Captures capacity failures. The dominant signal shifts with model scale: at 70M parameters, location statistics (mean) dominate (uniform depression); at 1.4B, spread statistics (std) dominate (routing heterogeneity).
  • Asymmetry ($G$): Generally sparse (near-chance performance) in decoder-only models, reflecting that content-dependent directional routing is not the primary failure mode. Exceptions include Flan-T5 (where $G$ detects cross-attention grounding failures) and Pythia (where $G$ detects deviations from RoPE-induced positional decay).
• Baseline Comparison:
  • Hidden-state probes (LLM-Check): Achieve strong general detection (0.62–0.72 LC-AUROC) but are structurally agnostic to failure modes and show moderate length confounding.
  • Supervised Spectral Classifiers (EigenTrack): Show marginal signal and poor generalization across datasets, likely due to overfitting to dataset-specific spectral signatures and irreducible length dependence via aspect ratios.
• Scaling: Transport diagnostics persist across Pythia scales (70M to 1.4B) and LLaMA 3.1 8B, though the aggregation method required to capture the signal changes from mean-based to variance-based as model capacity increases.

Significance and Claims

The paper claims to establish a structural boundary for attention diagnostics rather than proposing a single "best" hallucination detector. Its primary significance lies in:

1. Decomposition: Partitioning routing failures into orthogonal axes (capacity vs. direction) with formal guarantees (Cheeger inequality, orientation blindness theorem).
2. Interpretability: Providing a mechanistic explanation for why different datasets exhibit different failure modes (bottleneck vs. diffuse) based on the shape of the conductance landscape.
3. Methodological Rigor: Demonstrating that length confounding can inflate raw AUROC by up to 0.28 points and establishing a protocol (LC-AUROC) to isolate structural signals.
4. Architectural Insight: Showing that the "sparsity" of directional signals ($G$) in decoder-only models is an architectural property (dominance of positional structure over content-dependent direction) rather than a failure of the diagnostic.

The authors emphasize that while the framework characterizes how attention routing fails, it does not claim to solve hallucination detection in all cases (e.g., knowledge gaps without routing pathology may remain undetectable). The contribution is a principled framework for understanding the structural limits of spectral attention analysis.