Abliteration HandbookField guide
Research & reference · source archive

Research corpus (offline)

Pinned PDFs, text extracts, and GitHub README snapshots for handbook citations.

2 min read 0 examples Source-linked

In this chapter
On this page Interactive companion
Personal reading progress

0/5 sections reviewed

View original Markdown
Interactive companion

Explore the primary records

Use the visual controls alongside the complete source chapter.

Pinned PDFs, text extracts, and GitHub README snapshots for handbook citations.

Latest index: catalog-2026-10.json adds eight primary records; read the October map. The August 50-paper catalog and September 14-paper catalog remain available unchanged. The catalog links to canonical arXiv records without adding another large batch of binaries to git.

Refresh: npm run fetch:research-papers

Naming: PDFs here use dot IDs (arxiv-2602.02132.pdf). sources/fetched/ page snapshots from fetch:docs may use hyphen IDs (arxiv-2602-02132.txt) — cite papers/ for offline PDFs.

Beginner guide: ../../docs/refusal-research-beginners-guide.md


Papers (papers/)#

arXivTitlePDFText extract
2406.11717Refusal in LLMs Is Mediated by a Single Direction (Arditi, NeurIPS 2024)arxiv-2406.11717.pdfarxiv-2406.11717.txt
2502.17420Geometry of Refusal — Concept Cones (TUM)arxiv-2502.17420.pdfarxiv-2502.17420.txt
2505.19056Embarrassingly Simple Defense Against Abliterationarxiv-2505.19056.pdfarxiv-2505.19056.txt
2506.00085COSMIC — Generalized Refusal Direction IDarxiv-2506.00085.pdfarxiv-2506.00085.txt
2510.02768Safety Pretraining under Abliterationarxiv-2510.02768.pdfarxiv-2510.02768.txt
2512.13655Comparative Abliteration Methods (Young)arxiv-2512.13655.pdfarxiv-2512.13655.txt
2602.02132More to Refusal than Single Direction (QCRI)arxiv-2602.02132.pdfarxiv-2602.02132.txt
2603.22061Failure of Topic-Matched Contrast Baselinesarxiv-2603.22061.pdfarxiv-2603.22061.txt
2605.26526Open-Weight Defenses vs Abliteration + Prefillingarxiv-2605.26526.pdfarxiv-2605.26526.txt
2606.05396Code LLMs — Refusal vs Capabilityarxiv-2606.05396.pdfarxiv-2606.05396.txt
2410.03415False-refusal single vector (ICLR 2025)arxiv-2410.03415.pdfarxiv-2410.03415.txt
2507.11878Harmfulness ≠ refusal (Zhao)arxiv-2507.11878.pdfarxiv-2507.11878.txt
2511.08379SOM multi-direction (AAAI 2026)arxiv-2511.08379.pdfarxiv-2511.08379.txt
2603.27518Task-conditioned refusalarxiv-2603.27518.pdfarxiv-2603.27518.txt
2509.15202DeepRefusal (defense — train-time probabilistic ablation)arxiv-2509.15202.pdfarxiv-2509.15202.txt
2511.06852Differentiated Directional Intervention (Zhang — analysis only, no attack recipe)arxiv-2511.06852.pdfarxiv-2511.06852.txt
2603.04355Efficient refusal ablation via optimal transportarxiv-2603.04355.pdfarxiv-2603.04355.txt
2607.02396Fast multi-dimensional refusal subspaces (RFM-AGOP)arxiv-2607.02396.pdfarxiv-2607.02396.txt

Machine manifest: manifest.json


GitHub READMEs (readmes/)#


Community blogs (live URLs)#

Listed in scripts/fetch-docs.mjs. Offline files sources/fetched/grimjim-projected-blog.txt and grimjim-normpreserve-blog.txt are not in this tree until a successful npm run fetch:docs writes them.

TopicURL
Projected abliterationgrimjim/projected-abliteration
Norm-preserving biprojectedgrimjim/norm-preserving-biprojected-abliteration

September 2026 additive catalog#

catalog-2026-09.json adds 14 verified primary records without replacing the August catalog or archived papers. Annotated update. Tooling recheck.

October 2026 evidence#

Eight-paper catalog · Huihui revision/card snapshot · Tooling provenance · Search coverage and exclusions. The combined observatory has 72 papers. Model-card records are separate from the paper count.

Trace this chapter to its source.

The complete Markdown remains the editable source. This page adds navigation, typography, and a visual companion.

sources/research/README.md
Inspect the source fingerprintSHA-256 32e939f3d12f063faac8d670e482e1f9974bdd32fc6f392c38ff07521c8184a5Source commit 6fba7436a5914cad7c8f0835a14b480a29dac263