Transformer Attention Dilution as Hard Limit on Reasoning Depth

Fibers on X

What the data actually shows: Decades of research into transformer attention dilution as hard limit on reasoning depth have produced fragmented results because the underlying information-theoretic foundation was never recognized.
Why this is necessary: Polarity demands that transformer attention dilution as hard limit on reasoning depth exist as the mechanism that reconciles apparent opposites while preserving the ability to transmute between them.
What current approaches miss: The information-polarity framework was never applied at this scale, leaving the field with fragmented and often contradictory explanations.
If this is right: Raises the quality bar for future research by demanding testable predictions beyond mere consistency with existing data.

Series vertex. No documented primary-theorem join — Art.27 stays in the Idea Series.

Idea Series on X — Introduction · Index · Atlas

Outside instrument

Crux

Stake a prior. Name the disagreement. Not a theorem.

Outside the corpus

LLM depth limits. A paper can support a concept. It cannot support a necessity claim. Art.27 remains unproved by every card below.

Support · living source

Tight papers that bear on concepts inside Art.27. A DOI is not a join.

Liu, N. F. et al. (2024). Lost in the middle: How language models use long contexts. TACL 12, 157.
Long-context use is U-shaped; middle tokens drop — a depth/attention limit.
Does not support the necessity claim
Open originalNamed on Art.26, 27, 66, 67
Press, O. et al. (2023). Measuring and narrowing the compositionality gap in language models. Findings of EMNLP 2023.
Named on Art.26. Multi-hop composition fails as graphs grow.
Does not support the necessity claim
Open originalNamed on Art.26, 27, 66
Vaswani, A. et al. (2017). Attention is all you need. NeurIPS 2017.
The architecture whose dilution Art.27 names.
Does not support the necessity claim
Open originalNamed on Art.27
Dziri, N. et al. (2023). Faith and fate: Limits of transformers on compositionality. NeurIPS 2023.
Multi-hop composition collapses as graphs grow — a hard reasoning wall.
Does not support the necessity claim
Open originalNamed on Art.67

Counter-arguments · refusing pole

One domain counter per cluster, not a repeating method quartet. The mirror reads these as enough to refuse necessity.

Wei, J. et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 2022.
Counter: explicit intermediate tokens raise hop-depth without a new stabilizer.
Does not support the necessity claim
Open originalNamed on Art.26, 27, 66, 67