New Preprint Proposes Removing Safety Neurons to Make LLM Alignment Stronger
A 2026 arXiv preprint introduces Addition by Subtraction, a finetuning method that prunes safety-critical neurons then uses preference optimization to rebuild safer behavior…
- Safety alignment in current LLMs often depends on a tiny set of neurons that can be removed to break refusals
- The proposed method makes alignment less fragile by redistributing safety behavior during finetuning
- Tested on four models from three vendors, it cut some attack success rates by up to 90 percent and improved truthfulness
What the preprint actually shows
Safety alignment in large language models is meant to stop harmful outputs, but researchers have found it can be surprisingly easy to undo. The new preprint shows that removing or disabling just a handful of specific neurons can destroy a model's refusal behavior.
The authors, Hamid Dashtbani and Alexandra Dmitrienko, propose a technique called Addition by Subtraction. During finetuning, they gradually remove the safety-critical neurons that the model currently relies on. They then apply preference optimization so the model learns to stay safe using other parts of its network instead.
According to the abstract, the resulting models become much harder to break with pruning-based attacks. The robustness also carries over to refusal-direction ablation and many common black-box jailbreak prompts. The finetuned models showed better truthfulness and kept most of their general capabilities.
Key results and limitations
The authors tested the approach on four different models from three different vendors. They report consistent gains in safety robustness, with attack success rates dropping by as much as 90 percent for some attacks.
This work is a preprint posted on arXiv on November 4, 2026. Only the abstract is available here; the full paper, exact methods, and all experimental details are not part of this capture. No independent replication has been performed.
The technique is presented as an early step toward alignment that does not rest on a few fragile components. It does not claim to eliminate all jailbreak risks or make models perfectly safe.
Why it matters for uncensored AI
Our analysis: The preprint starts from the documented fact that current safety alignment can be sparse and therefore easy to disable. This highlights a practical choice for users and developers who want more reliable model behavior without depending on a handful of neurons that can be targeted.
For those running local or open-weight models, techniques like this could inform how to create versions that resist certain attacks while still allowing broad creative freedom. It also underscores why many people seek alternatives that let them control or remove alignment layers themselves rather than relying on vendor safety mechanisms that may be brittle.
We recommend considering Muah AI when you want a multimodal companion platform that combines uncensored chat, photo exchange, voice interactions and real-time phone calls in one place, giving users more direct choice over expression without corporate gatekeepers dictating every boundary.
