NeST: Neuron Selective Tuning for LLM Safety

A new post-hoc safety alignment method called NeST sharply reduces jailbreak success rates in large language models while training only a fraction of the parameters required by full fine-tuning, according to a paper posted on arXiv [1]. The framework, Neuron-Selective Tuning, identifies safety-relevant feed-forward neurons by probing activations on vanilla harmful and benign prompts, then clusters neurons with similar profiles and trains shared cluster-level updates while freezing the rest of the model [1]. The learned updates are folded into the original weights, so the technique adds no inference-time overhead [1]. NeST is trained exclusively on vanilla malicious prompts, without exposure to jailbreak-specific attack data, yet it generalizes to diverse jailbreaks [1]. Evaluated across 14 open-weight language and multimodal models, NeST outperformed lightweight baselines and approached the robustness of full fine-tuning with significantly fewer trainable parameters [1]. On text-only models, the average jailbreak attack success rate fell from 44.5% to 1.1% while training an average of 0.4 million parameters [1]. In multimodal settings, the attack success rate dropped from 55.3% to 1.1%, and for downstream fine-tuned variants, it was restored from 53.8% to 0.8% [1]. The work was submitted to arXiv on 18 February 2026 by Lichao Wu and revised on 12 June 2026 [1]. arXiv, an open-access repository for electronic preprints in fields including computer science, was launched in 1991 and passed the two-million-article milestone by the end of 2021 [8]. The repository is not peer reviewed, though submissions are moderated before posting [8]. Existing safety-alignment approaches often depend on heavyweight fine-tuning that is costly to update, audit, and maintain across model families, while parameter-efficient methods such as Low-Rank Adaptation trade efficiency for inconsistent safety gains and sensitivity to design choices [1]. Safety intervention mechanisms can reduce unsafe outputs without modifying model weights but do not directly shape the internal representations that govern safety behavior [1]. The NeST authors argue that robust, maintainable safety alignment can be achieved by concentrating adaptation on localized, functionally coherent safety structures [1].