VuongKB/papers/llm-refusal-code-repos.md

created, id, keywords, links, modified, summary, tags, title

created	id	keywords	links	modified	summary	tags	title
2026-04-17T06:08:08+00:00	b6f1b91a-a987-4bef-bb94-937e318ceb66	Heretic DBDI Differentiated Bi-Directional Intervention abliteration refusal removal uncensoring AAAI-AIA 2026 USENIX Security 2025 KAIST		2026-04-17T06:08:08+00:00	Catalogs public code repositories and resources for LLM refusal removal techniques, including Heretic (automated abliteration), DBDI (white-box intervention framework), and the "Refusal Is Not an Option" paper.	llm refusal-removal uncensoring abliteration code-repos tools	Code Repositories for LLM Refusal Removal Techniques

Code Repositories: LLM Refusal Removal Techniques

This note catalogs public code repositories and resources related to LLM refusal removal and uncensoring techniques. For the full analysis, see llm-refusal-removal-techniques-deep-dive.

Repositories

1. Heretic

• URL: https://github.com/p-e-w/heretic
• Description: Fully automatic abliteration tool. One command, co-minimizes refusals and KL divergence.
• Approach: Automated refusal abliteration — modifies model weights directly to suppress refusal behavior while preserving general capabilities via KL divergence regularization.

2. DBDI (Differentiated Bi-Directional Intervention)

• URL: https://github.com/PeiJieSun/DBDI-main
• Description: Source code for the AAAI-AIA 2026 paper on Differentiated Bi-Directional Intervention.
• Approach: White-box framework with adaptive projection nullification and classifier-guided sparsification for targeted refusal mechanism removal.

3. "Refusal Is Not an Option"

• Repo: No public repository.
• Paper: https://wsp-lab.github.io/papers/song-sec25.pdf
• Authors: Song, Kim, Kim (KAIST)
• Venue: USENIX Security 2025
• Description: Analysis of refusal behavior as a security surface, with implications for jailbreaking and abliteration approaches.