English(EN)Poster: A Preliminary Study of LLM Distillation Inference
新的蒸馏方法旨在提高LLM的能力和安全性
作者PulseAugur 编辑部·[8 个来源]·
研究人员正在探索用于大型语言模型的高级蒸馏技术,重点是提高能力同时降低风险。一种名为ReCal的方法通过校准剪枝来保留关键的教师支持的预测,从而提高策略内蒸馏恢复能力,带来更好的数学推理和代码生成。另一项研究引入了“用于归罪的蒸馏”(DFI)和“用于能力的蒸馏”(DFC),以揭示或隐藏教师模型的失准,其中DFI在检测隐藏行为方面显示出潜力。此外,SAPD和PIVOT等方法旨在通过优化知识在模型之间的转移方式来改进策略外学习和少样本分类,从而提供计算效率和在特定任务上更好的性能。
AI
arXiv:2610.11332v1 Announce Type: new Abstract: Structured pruning reduces the deployment cost of reasoning language models, but the resulting capability degradation can hinder subsequent on-policy distillation (OPD) recovery. Because OPD relies on student-generated trajectories,…
arXiv:2610.12137v1 Announce Type: cross Abstract: Unauthorized model distillation, in which a model is trained on the outputs of a proprietary large language model (LLM), is a growing threat to model providers. We study distillation inference: determining whether a suspect model …
arXiv:2610.11012v1 Announce Type: new Abstract: Powerful misaligned AI models might recognize alignment evaluations and strategically behave well on them, rendering direct audits uninformative. However, distilling such a model into a weaker benign student places the teacher in a …
arXiv:2610.11167v1 Announce Type: cross Abstract: Vertical-domain few-shot classification remains challenging for small language models, as limited supervision makes it difficult to acquire domain-specific decision knowledge. On-Policy Distillation (OPD) can improve teacher-guide…
arXiv cs.CL
TIER_1English(EN)·Tianle Wang, Jiayu Liu, Ruizhi Zhao, Ning Miao·
arXiv:2610.09665v1 Announce Type: new Abstract: On-policy post-training can improve large language models by learning from their own trajectories, but requires costly rollout generation. We ask whether fixed demonstrations can support competitive off-policy learning through bette…
arXiv:2610.07237v1 Announce Type: new Abstract: A model that learns from its own outputs inherits more than their correctness: it inherits which solutions it produces. We formulate Looped Self-Distillation, a self-evolution framework for code generation in which a model repeatedl…
Iterative self-distillation enables LLM agents to learn from successive deployments, offering a path toward recursive self-improvement (RSI). Yet our experiments with existing methods reveal a collapse in deployment performance across cycles, while task performance with privilege…
<p><b><span style="white-space: pre-wrap;">TL;DR: We previously motivated two complementary ways to use distillation for AI safety; we’re releasing a </span></b><a href="https://arxiv.org/pdf/2610.11012"><b><span style="white-space: pre-wrap;">paper</span></b></a><b><span style="…