Staged source for a planned arXiv submission of Semiotic Taps: Lightweight Adapter Modules for Bifurcation Detection in Frozen Language Models.
Scope of this folder. Everything under
arxiv/is the Stage 3 SRT-Adapter manuscript only (v5 through v22c_a050 on a frozen Qwen-2.5-7B backbone, with a Stage 3 Phase 1 prelude on TinyLlama-1.1B). It does not cover the Stage 4 NLA work, hidden-state verbalisation, or the Llama-3.2-3B / Gemma-2-2B cross-backbone replications. Those live in../paper_nla.mdand will, when prepared for submission, be staged in a sibling folder (arxiv_nla/), not merged here. Mentions of "LLaMA" and "Gemma" insidepaper.md/paper.texrefer either to TinyLlama (the Stage 3 Phase 1 backbone) or to the boilerplate "the adapter works with any HuggingFaceAutoModelForCausalLM(LLaMA, Qwen, Mistral, Phi, Gemma)" sentence, not to the Stage 4 NLA evaluations on those backbones.
Status: not yet submitted. This folder is forward-looking. The manuscript is currently repository-hosted only. The only Lancaster preprints actually posted at the time of writing are the two SSRN entries (SSRN 5987495 and SSRN 6349978); see
../paper_nla.md§14.
paper.md— canonical Markdown source. Edit this.paper.tex— single-file LaTeX source (pandoc-generated frompaper.md; regenerated bybuild.sh).paper.pdf— locally built reference PDF (xelatex, ~43 pages).abstract.txt— standalone abstract for the arXiv submission form.00README.XXX— arXiv processing hints (declares xelatex toplevel).build.sh— rebuild script (runs pandoc + xelatex twice).
No .bib or external figure files. References are inline author-year prose.
./build.shOr manually:
pandoc paper.md -o paper.tex --pdf-engine=xelatex --standalone
xelatex -interaction=nonstopmode paper.tex
xelatex -interaction=nonstopmode paper.tex # second pass for longtable + refs- Upload
paper.texplus00README.XXXas a tarball:tar czf srt-adapter-arxiv.tar.gz paper.tex 00README.XXX
- arXiv's autotex driver detects
fontspec/unicode-mathand runs xelatex automatically. No special declarations needed beyond that. - Suggested categories:
- Primary:
cs.CL(Computation and Language) - Cross-list:
cs.LG(Machine Learning),cs.AI(Artificial Intelligence)
- Primary:
- Title and abstract: full abstract is in
abstract.txt(~5000 chars). arXiv's submission form limits the abstract field to ~1920 characters, so it will need trimming before paste. The full abstract remains intact in the PDF. - License: arXiv's default (non-exclusive license to distribute) or CC-BY 4.0.
- Source uses
unicode-mathfor symbols (Δ, ×, →, ≈, ±). pdflatex will reject these directly; theiftexswitch in the preamble handles engine detection. - 41 pages, ~200 KB PDF, no external assets.
- Companion artifacts (model weights, benchmarks, code) live at