新的蒸馏方法利用RL反馈训练紧凑型AI重排器

作者 PulseAugur 编辑部 · [1 个来源] · 2026-09-03 04:00

研究人员开发了一种新颖的两阶段框架，用于训练紧凑型指令遵循重排器，结合了离策略教师优化和在策略学生蒸馏。第一阶段使用LLM-judge反馈在88K个示例上通过离策略GRPO来增强一个4B教师重排器。第二阶段涉及一个1B学生重排器，它对自己的排名进行采样，并接收源自教师策略的软奖励，尤其是在分布偏移下提高了性能。与传统的蒸馏技术相比，该方法在MAIR-11和MAIR-Full基准测试上取得了优越的结果，甚至超过了已发布的RL训练重排器。 AI

影响这项研究提供了一种更有效的方法来训练紧凑型AI重排器，有可能提高部署能力和在分布偏移下的性能。

排序理由该集群包含一篇详细介绍新AI模型训练方法的论文。[lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.AI 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 22 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 该集群包含一篇详细介绍新AI模型训练方法的论文。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration: Single-source cluster; Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics: paper, model release; Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance: High; Clearly on-topic for AI-industry coverage.
Story freshness: Breaking (< 6h); Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

arXiv cs.AI TIER_1 English(EN) · Vignesh Prabhakar, Jialing Pan, Anil Babu Ankisettipalli · 2026-09-03 04:00

On-Policy Distillation Meets Off-Policy GRPO: Training Compact Instruction-Following Rerankers

arXiv:2609.01947v1 Announce Type: cross Abstract: Compact instruction-following rerankers are attractive for deployment, but conventional distillation pipelines typically train students by offline imitation of teacher outputs on a fixed set of examples, constraining supervision t…

本文如何被排名

报道来源 [1]

On-Policy Distillation Meets Off-Policy GRPO: Training Compact Instruction-Following Rerankers

相关实体

相关话题