DNAlign confines safety perturbations to a harmful subspace via null-space projection, cutting harmful outputs across LLM backbones while preserving fluency and factual utility

It proposes DNAlign, a lightweight alignment framework that integrates control-theoretic optimization with null-space projection, treating the LLM as a dynamic system and introducing controllable perturbations to steer generation toward safe behavior. Compared with existing safety alignment approaches, the framework pursues safety guidance through control signals rather than costly computation or changes that disrupt core knowledge, targeting the persistent trade-off between safety and utility. Framework design as stated at the abstract level; the text notes existing approaches either incur high computational cost or unintentionally disrupt the model's core knowledge, degrading fluency and factual accuracy on benign tasks.

It designs a projection module that restricts perturbations to a harmful-related subspace derived from neutral hidden states, thereby preserving general knowledge and response quality. This is the framework's key component, scoping the safety intervention to a subspace tied to harmful content rather than altering model behavior globally. The abstract describes it as a key component and attributes its mechanism to a subspace derived from neutral hidden states; implementation details are not expanded in the provided text.

It uses a value function trained on human preference data to adaptively optimize the control signals so alignment matches human safety preferences. Preference learning is brought into the control-signal optimization loop, letting perturbation strength and direction adapt to the preference objective. The abstract states the value function is trained on human preference data; training scale, data composition, and optimization details are not given in the provided text.

Extensive evaluations across multiple LLM backbones show the framework consistently reduces harmful outputs while maintaining fluency, coherence, and factual utility, with superior overall performance over prior alignment baselines and no sacrifice in generation diversity. Relative to prior alignment baselines, it reports better overall performance and explicitly emphasizes that generation diversity is not sacrificed. Abstract-level conclusion from multi-backbone evaluation; specific benchmarks, metric values, and statistical details are not listed in the provided text.

The framework targets settings that require safe and reliable LLM deployment, for practitioners who want to reduce harmful outputs without sacrificing fluency, coherence, factual utility, or generation diversity; it is designed to be general across multiple LLM backbones and to replace costly alignment pipelines with a lightweight approach. The provided text is abstract-level and does not give specific benchmark names, metric values, model lists, or hyperparameter settings, so the applicable boundary should be read as a consistent trend across multi-backbone evaluation rather than as a claim about one specific model or one specific harm category.

The provided text contains only the abstract and arXiv page navigation, with no figures, benchmark breakdowns, metric values, or ablation results, so the magnitude of differences across backbones, the construction details of the projection subspace, and the scale and provenance of the value-function training data cannot be judged here. Readers interested in these should return to the full text and accompanying code; likewise, the specific comparison conditions and evaluation protocol behind the abstract's claim of superiority over prior alignment baselines need confirmation in the body.