arXiv cs.AI· Yuejia Li, JunHeng Li, Hao Wu, Ke He, Mang Ye·· 3 天前AI 评分27
Radioactive Feedback：追踪仅靠标签式 LLM 监督训练出的模型
Radioactive Feedback: Tracing Models Trained by Label-Only LLM Supervision
AI 导读
研究提出 Radioactive Feedback，通过在 LLM-as-a-Judge 的奖励信号中编码注册好的行为偏好模式，使仅凭标量分数训练的模型也能被第三方追踪。评估者对质量达标的回答施加有界逐点调整，重复优化即可将该模式传递给他人训练的学生模型。代码与数学实验显示，这些偏好仅通过分数即可跨数据集迁移，并支持针对固定参考池的黑盒检测。
正文 · 原文
Abstract:Conventional distillation transfers capability through teacher-generated demonstrations. LLM-as-a-Judge knowledge distillation offers a distinct and increasingly practical alternative. In the label-only variant studied here, the student generates every response and improves from scalar scores alone. Prior work shows that this non-demonstrative feedback can improve student capability, but it creates a fundamentally different provenance problem. No teacher-generated answer enters the student's training set, so any evaluator-specific trace must pass through a low-bandwidth reward signal rather than copied text. Existing watermarks and distillation detectors do not cover this score-only channel. We introduce Radioactive Feedback, which encodes a registered pattern of behavioral preferences in the rewards themselves. The evaluator applies bounded pointwise adjustments to quality-qualified responses, allowing repeated optimization to transfer the pattern to a student trained by another party. Experiments on code and mathematics show that the registered preferences transfer through scores alone and across datasets, enabling black-box detection against a fixed reference pool.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.15685 [cs.AI] |
| (or arXiv:2610.15685v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.15685 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Yuejia Li [view email]
[v1]
Sun, 11 Oct 2026 15:09:58 UTC (6,065 KB)
来源：arXiv cs.AI · arxiv.org