Skip to content

Seqev/dcr-attention-paper

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DOI

DCR-Attention — single-paper assembly (paper-v1.0)

Title. DCR-Attention: Exact Q-Axis Sparse Decode at Production Scale — A Measured Scaling Law and Retention of Long-Range Dependence, under Matched-Control Discipline.

Authors. Evgeny Vyaltsev, Daniil Vyaltsev. ORCID 0009-0004-3712-6798.

Release route. GitHub tag paper-v1.0 → Zenodo DOI (priority record). Not arXiv. Same material may later be submitted to TMLR (rolling) or a 2027 conference cycle.

What this is

One paper assembled from the existing project corpus (Candidate α of the synthesis). It unifies four research threads around a single object — softmax attention concentration — and validates every quantitative claim under an explicit matched-control discipline.

Contributions: exact Q-axis ranking (Theorem 1); a measured, descriptively-stated scaling law ΔPPL = O(N^{-(1-α)}), α≈0.31 (Theorem 3); retention of long-range dependence to N=128K via true selection, with a leakage dissociation; a per-query correctness certificate; a multi-batch memory-wall systems result with a fused ABKV kernel (measured 1.21× over SDPA); and the matched-control methodology, with an explicit retraction ledger.

Layout

paper/main.tex     LaTeX source (compiles with pdflatex, 3 passes)
paper/main.pdf     compiled preprint (13 pp.)
PLAN.md            assembly plan (section→source map, deferred items, ledger)
REPRO.md           registered-run provenance and environment
docs/              synthesis note (rationale for Candidate α)
LICENSE            CC-BY-4.0

Build

cd paper && pdflatex main.tex && pdflatex main.tex && pdflatex main.tex

How to cite

@misc{dcr_attention_2026,
  author    = {Vyaltsev, Evgeny and Vyaltsev, Daniil},
  title     = {{DCR-Attention: Exact Q-Axis Sparse Decode at Production Scale}},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20479271},
  url       = {https://doi.org/10.5281/zenodo.20479271}
}

Scope honesty

Five items are explicitly deferred (stated as in-progress / future, not claimed): multi-seed retention boundary; causal boundary mechanism (starvation vs mis-selection, inject running); TTFT decomposition; causal upgrade of Theorem 3 (Gap-1); cross-model α. Eight retracted hypotheses are listed in the paper's retraction ledger (§9).

About

Training-free Q-axis sparse attention for long-context LLM decode — exact ranking, a measured entropy-concentration scaling law, long-range retention to 128K, and a per-query accuracy certificate, validated under matched-control discipline (Llama-3.2-1B).

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors

Languages