Self-Attention as Transport: Limits of Symmetric Spectral Diagnostics

The contribution here isn't a new detection method but a proof of absence: symmetric spectral tools are structurally incapable of seeing directionality, meaning any audit built on them has a blind spot baked in at the mathematical level, not just an engineering limitation that better tuning could fix.

This lands directly on top of the hallucination detection work published the same day ('Detecting Hallucinations in Large Language Models via Internal Attention Divergence Signals'), which proposes measuring attention head divergence from uniform distributions as a lightweight reliability signal. That method relies on attention pattern analysis without, as far as the summary indicates, accounting for directional routing. If the asymmetry coefficient finding holds, divergence-based detectors may be measuring the wrong axis entirely, flagging distributional oddities while missing the directional failures that actually produce false outputs. The procedural execution diagnostic work from May 1 ('When LLMs Stop Following Steps') adds further context: models losing track of intermediate state during multi-step tasks is exactly the kind of failure where information flow direction inside attention layers would matter most.

Watch whether the divergence-based detection paper from the same arXiv cycle is updated or followed by a response incorporating asymmetry coefficients into its instrumentation. If it isn't within two to three months, that suggests the two research threads aren't actually in dialogue, and the theoretical gap may sit unaddressed in production tooling longer than the proof's clarity would suggest it should.