Skip to content

Refresh the arXiv note to the current results (dose-fragility 3-vs-3, MoE, methods lesson) - #42

Merged
bamdadd merged 1 commit into
mainfrom
paper/refresh
Jul 20, 2026
Merged

Refresh the arXiv note to the current results (dose-fragility 3-vs-3, MoE, methods lesson)#42
bamdadd merged 1 commit into
mainfrom
paper/refresh

Conversation

@bamdadd

@bamdadd bamdadd commented Jul 20, 2026

Copy link
Copy Markdown
Owner

The committed docs/note.tex/note.pdf had gone stale against docs/note.md. This rebuilds them from the current source so the paper is submission-ready (submission itself still awaits an explicit go).

What's new in the paper

  • Cross-model dose-fragility (§4). The pinned norm-relative k=2 dose collapses coherence out-of-family: DeepSeek-Coder-33B-Instruct (injected coherent 0.042, random-direction 0.093 vs 1.000; 0/216 strict) and CodeLlama-34B-Instruct (0.056 / 0.060 vs 1.000; 0/216 strict). A three-versus-three split — collapse on {Coder-7B, DeepSeek-33B, CodeLlama-34B}, hold on {Coder-32B 0.773, MoE 0.750, dense Qwen}. Both are broken-model nulls (the matched-norm control collapses alongside the injected condition), invalid tests that enter nothing into the headline.
  • Corrected methods lesson. A cheap rule-based coherence gate passed on both collapsed models, but on identical transcripts rated 0.79–0.89 coherent while the Bedrock judge rated 0.04–0.06 — the proxy is too lenient. Length is ruled out (CodeLlama's gate sampled at the full 200-token scoring length). A gate must use the scoring judge.
  • Cross-architecture MoE probe (§5). Qwen1.5-MoE-A2.7B-Chat: STEP-1 confirms the injection hook is live (routing changed at 0.786 of positions); STEP-2 is a genuine null (0/216, coherence holds at 0.750). The null is predetermined by scale and post-training, not an architecture verdict.
  • Total spend corrected to ~$23 ($6.6 GPU + ~$16 judge) to match RESULTS.md.

Unchanged

Core 7B/14B/32B trend table, the 32B three-way, the hero figure, and the logit-lens mechanism — same numbers (Coder-32B 5/216, n=1).

Verification

  • pdflatex compiles clean, 7 pages; both new tables render within margins.
  • Every number cross-checked against live RESULTS.md (independent adversarial audit, 6/6 pass).
  • No AI attribution in source or PDF metadata; author Bamdad Dashtban. Lindsey et al. arXiv:2601.01828 cited.

Rebuild: cd docs && make. Not submitted to arXiv; endorser email not sent — both await an explicit go.

Rebuild docs/note.tex + docs/note.pdf from the updated docs/note.md so the
paper matches main. Adds the cross-model face of the dose-fragility result
(the pinned norm-relative dose collapses coherence out-of-family on
DeepSeek-Coder-33B and CodeLlama-34B, a three-versus-three split against
Coder-32B, the MoE probe, and the dense Qwen rungs; broken-model nulls that
enter nothing into the headline) and the corrected methods lesson (a cheap
rule-based coherence gate is too lenient versus the Bedrock judge; length is
ruled out). Adds the cross-architecture MoE section (Qwen1.5-MoE-A2.7B: the
injection hook is live, routing perturbed at 79% of positions, yet the
behavioural result is a genuine null predetermined by scale and post-training,
not an architecture verdict). Corrects total spend to ~$23 to match RESULTS.md.

The core 7B/14B/32B trend table, the 32B three-way, the hero figure, and the
logit-lens mechanism are unchanged.
@bamdadd
bamdadd merged commit 4d69d64 into main Jul 20, 2026
1 check passed
@bamdadd
bamdadd deleted the paper/refresh branch July 20, 2026 13:14
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