Loading
Life sciences · Preprint
arXiv · September 8, 2026
Posted before peer review. The findings may change or fail to hold.
This preprint proposes Suan, a gradient-level preference optimization algorithm for LLM safety alignment, and claims superior safety with preserved utility compared to existing methods. The work is methodological and has not been peer reviewed; it reports only benchmark comparisons without independent validation, human evaluation, or disclosure of effect sizes and statistical significance.
Preprint. Intervention: Suan: a gradient-level preference optimization algorithm for LLM safety alignment. Compared with: Existing preference optimization methods for safety alignment (named baselines not itemized in abstract).
Suan formulates optimization at gradient level, bypassing variational derivation, yielding more interpretable training dynamics Claims superior safety alignment and full preservation of response utility versus competitive baselines Addresses over-refusal and quality degradation observed in post-trained open-weight LLM variants
Claims superior safety alignment and full preservation of response utility versus competitive baselines
The source did not state who this applies to in practice.
This is an unrefereed arXiv preprint describing a novel algorithmic approach to LLM safety; it has not undergone peer review and reports no clinical outcomes or validated benchmarks suitable for practice guidance.
As stated by the source record.
Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.
What is missing. This record has no reported figures. That is a gap in the analysis, not a judgement about the study.
Integrating robust safety guardrails into Large Language Models (LLMs) is essential for delivering helpful yet harmless responses. While proprietary systems exhibit reliable safety controls, their underlying methodologies and trade-offs remain largely undisclosed. Achieving comparable security in open-weight models remains a persistent challenge, as post-trained variants frequently suffer from over-refusal and degraded general quality. To overcome these drawbacks, we introduce Suan, a novel preference optimization algorithm. Unlike existing methods, we formulate the optimization objective directly at the gradient level, bypassing the standard variational derivation. As a result, we obtain more interpretable and robust training dynamics. Extensive evaluations across a diverse suite of competitive baselines and benchmarks demonstrate that Suan achieves superior safety alignment while fully preserving response utility.
Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.