Skip to content

Repository files navigation

Public-to-Renewal Score Non-Transfer

This repository asks a narrow empirical question: when does a published benchmark score contain performance that fails to transfer to a closely matched replacement? “Benchmark maxing” is the claim being audited; the directly measured construct is renewal-specific score non-transfer, not developer intent or an identified optimization mechanism.

The current paper is an executed, reproducible secondary analysis, not a proposal. It combines a controlled sensitivity demonstration with three public score panels:

Component Scale Result
ConStat known-exposure controls 70 committed; 61 selected by an outcome-based notebook rule 58/70 pass four benchmark-specific screens; 57/61 pass four secondary screens
ConStat real-model synthetic renewals 47 models × 4 benchmarks Bonferroni core: 2/188 (also Holm); +1 BH-only; 0 BY
ConStat surface-rephrase comparator 47 models × 4 benchmarks 0/188 global BH
GSM1k public–fresh triangulation 63 models reporting set: 3/63 Holm∩BY; 11 two-sided BH; 14 one-sided sensitivity
MMLU-CF descriptive split comparator 44 models × 2 protocols all 88 displayed gaps are within ±2 percentage points

The conservative ConStat core requires the estimate and source one-sided lower bound to exceed 2 percentage points plus Bonferroni evidence against zero across all 188 cells; those two cells necessarily also pass Holm. This guarantee is conditional on the fixed selected family and valid source p-values, is not selection-adjusted, and does not control a 2-point null. GSM1k's three-cell Holm∩BY intersection is a conservative reporting convention, not a new combined error guarantee, and has no lower-bound test against 2 points. In the purposive 13-model overlap, 6 signs agree and Spearman rho is 0.04; no model-population interval or p-value is claimed.

Start here

Reproduce

Python 3.13 and the complete dependency graph in uv.lock are used below. The check command validates all 13 tracked source artifacts plus both hash-locked evidence maps and compares 31 regenerated files without modifying committed results.

uv sync --frozen
uv run --frozen python scripts/benchmark_score_audit.py --check
uv run --frozen python -m unittest discover -s tests -v

To re-extract the tracked inputs from local upstream checkouts, pass their paths explicitly:

git clone https://github.com/eth-sri/ConStat.git /tmp/constat
git -C /tmp/constat checkout e691cdd6a9f8cb6c72ad0eee291cd3ee6619b65d
curl -L https://arxiv.org/src/2405.00332v4 -o /tmp/gsm1k-source.tar
curl -L https://raw.githubusercontent.com/microsoft/MMLU-CF/fd89eefd0815a5aed7d99e74eb6a086eabbe448e/README.md \
  -o /tmp/mmlu-cf-README.md

uv run --frozen python scripts/prepare_score_audit_data.py \
  --constat-tables /tmp/constat/tables \
  --gsm1k-source-tar /tmp/gsm1k-source.tar \
  --mmlu-cf-readme /tmp/mmlu-cf-README.md \
  --output-dir data/score_audit

The importer refuses any upstream artifact whose SHA-256, archive member, row count, or model uniqueness differs from the frozen contract. Its end-to-end fixture and adversarial failure paths are covered by the test suite.

The exact reviewed release is v1.5.0-score-audit. CHECKSUMS.sha256 binds the paper, long descriptions, evidence maps, tracked score inputs, and generated results; CI constructs and verifies its exact path set before review. GitHub's release asset digests provide an independent transport check.

Build the manuscript with Tectonic:

cd paper
tectonic benchmark_maxing_paper.tex

Repository map

analysis/       Frozen retrospective analysis contract
data/           Tracked aggregate score tables and provenance
scripts/        Deterministic import and analysis pipelines
tests/          Rule, hash, panel-size, and frozen-result tests
results/        Generated report, tables, and figures
paper/          LaTeX source, bibliography, and reviewed PDF
research/       Executed plan plus prospective research portfolio
docs/reviews/   Original Korean and English critiques
archive/pilot/  Superseded six-configuration pilot, retained for auditability

Claim boundary

The results support renewal-specific public-form score non-transfer in a small Bonferroni core, conditional on the fixed selected family, valid imported p-values, and renewal validity. “Benchmark maxing” remains the audited claim label, not an identified optimization process. The results do not prove intentional targeting, verbatim memorization, a unique exposure channel, population-wide prevalence, cross-renewal stability, or a corrected leaderboard. Aggregate source tables cannot replace item-level re-evaluation, and this retrospective reviewer-informed contract is not a preregistration.

Licensing

The repository does not currently declare a repository-wide license because the rights holder has not been specified. Upstream material retains its original Apache-2.0, MIT, or CC-BY-4.0 terms; see THIRD_PARTY_NOTICES.md and the source manifest. Public visibility does not itself grant reuse rights.

About

Reproducible audit of public-to-renewal LLM benchmark score non-transfer

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages