The BasePeople

Binchi Zhang

University of Virginia (internship at Visa Research)

Titles as the sources gave them. In 1 entry, with 3 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“we propose that formulaic refusal strings can function as outlier modes: narrow, weakly grounded response patterns that are easy to fit but vulnerable to subsequent updates.”
When Safety Becomes An Outlier: Understanding the Retention of LLM Safety BehaviorsarXiv, November 7, 2026. Speaking as: Other. The abstract, stating the authors' explanation for why safety training is forgotten.
“Across multiple models and alignment regimes, natural safety targets improve both immediate safety and its retention under continued fine-tuning without sacrificing helpfulness.”
When Safety Becomes An Outlier: Understanding the Retention of LLM Safety BehaviorsarXiv, November 7, 2026. Speaking as: Other. The abstract's main result for the proposed NEST method.
“fixed refusals and fixed unrelated targets are both rapidly fit on train and held-out harmful prompts, reaching near-perfect ROUGE-L within roughly 100 steps.”
When Safety Becomes An Outlier: Understanding the Retention of LLM Safety BehaviorsarXiv, November 7, 2026. Speaking as: Other. From the controlled diagnostic comparing fixed refusals, a fixed unrelated sentence, and natural responses.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    When Safety Becomes An Outlier: Understanding the Retention of LLM Safety Behaviors

    Researchers at the University of Virginia and Visa Research argue that language models lose safety training after fine-tuning because fixed refusal phrases are easy to learn and easy to overwrite. They propose training on natural, request-specific safety responses and report better retention.