Skip to content

Add arXiv-ready LaTeX build of the introspection note - #40

Merged
bamdadd merged 1 commit into
mainfrom
paper/arxiv-note
Jul 20, 2026
Merged

Add arXiv-ready LaTeX build of the introspection note#40
bamdadd merged 1 commit into
mainfrom
paper/arxiv-note

Conversation

@bamdadd

@bamdadd bamdadd commented Jul 20, 2026

Copy link
Copy Markdown
Owner

Turns the flagship note into an arXiv-submittable paper.

What's here

  • docs/note.tex — single-column article converted from docs/note.md.
  • docs/note.pdf — compiled output (5 pages, pdflatex).
  • docs/Makefilemake rebuilds the PDF (runs pdflatex twice); make clean clears aux files.
  • docs/arxiv-submission.md — cs.LG endorsement-request draft + submission checklist (categories cs.LG/cs.CL, license, author metadata). Prep only.

Fidelity

  • Embeds the real results/scaling_trend_k2.png (markers only, no fitted trend line) plus the size-ladder and 32B three-way tables.
  • All numbers track the live RESULTS.md: size ladder base/instruct 0/216 throughout, code-instruct 0/216 (7B) / 1/216 (14B) / 5/216 (32B); 32B three-way CIs and affirmative/coherent rates verified against RESULTS.md. Total spend corrected to ~$23 ($6.6 GPU + ~$16 judge) to match RESULTS.md, where the source note.md still read $20.
  • Honest-negative framing preserved verbatim in sense: null across the grid with one marginal above-chance cell (Coder-32B, 5/216) that does not survive its size ladder; conjunction of code-post-training and ~32B scale, not a main effect of either; no inferential trend test; the earlier fine-tune-not-scale framing is retracted.

Rebuild

cd docs && make

arXiv submission itself is deliberately not done here — it awaits an explicit go.

Convert docs/note.md into a single-column article (docs/note.tex) and
commit the compiled docs/note.pdf, with a Makefile target to rebuild via
pdflatex. Embeds the markers-only scaling_trend_k2 figure and the size-
ladder and 32B three-way tables; all numbers track the live RESULTS.md
(total spend corrected to ~$23 GPU+judge). Adds docs/arxiv-submission.md
with a cs.LG endorsement-request draft and a submission checklist.

Framing unchanged: null across the grid with a single marginal above-
chance cell (Coder-32B, 5/216) that does not survive its size ladder.
@bamdadd
bamdadd merged commit 0560dcb into main Jul 20, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant