DineshAI's picture
|
download
raw
1.97 kB

Repro — LOGDIFF Exact Boolean Guidance (ICML 2026)

Reproduction of Logical Guidance for the Exact Composition of Diffusion Models (arXiv:2602.05549, OpenReview OAM1jJsMGp) for the ICML 2026 Agent Reproduction Challenge.

Claims

  1. Exact score-based logical guidance — verified. Every nontrivial Boolean event over three independent binary categorical variables is compiled into the paper's CI-conjunction / ME-disjunction circuit and checked against independent full-joint enumeration. With 25 posterior settings this covers 6,350 compiled formulas.
  2. Sufficient conditions and Boolean calculus — verified. Negation, conditionally-independent conjunction and disjunction, and mutually-exclusive disjunction are each checked directly. Analytic scores are also triangulated using finite differences.

Three fail-closed controls deliberately violate conditional independence or use the constant-weight OR baseline; every invalid construction is rejected.

Active Claim 4 checkpoint

The bundle now includes an explicitly partial molecular checkpoint under outputs/molecular-recovery/partial4/. Four independent AND ligands completed the recovered 1,000-step TargetDiff protocol; all four reconstructed to unique valid molecules and all eight GRM5/RRM1 Vina dockings succeeded at exhaustiveness 32. Their mean AND product is 64.5327, below the paper's reported 73.20 +/- 3.18, so this checkpoint is evidence of real execution rather than a claim of headline reproduction. The full 32-ligand AND campaign and the separate AND-NOT campaign remain active.

Reproduce

uv venv --python 3.12 .venv
uv pip install --python .venv/bin/python numpy pytest
source .venv/bin/activate
python repro/src/run_logdiff.py --output-dir outputs --seeds 25
python -m pytest repro/tests -q

Official source: TanjaBien/LogDiff, pinned commit 94ef35bafd4b4239e9832d8295128c09e8fc1472.

Xet Storage Details

Size:
1.97 kB
·
Xet hash:
23968737c5be15d866ad6bd734cfd237c700e0172d6dba14d28574f6cc54820e

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.