English(EN) The Alignment Tax Is Graded: Task-Selective Capability Loss Under Length-Normalized DPO with SFT Retention

Qwen3.5-4B 模型在奖励边际增加时表现出任务选择性能力损失

作者 PulseAugur 编辑部 · [1 个来源] · 2026-11-04 13:18

一项针对 Qwen3.5-4B 模型的研究，采用了带有 SFT 保留的直接偏好优化（DPO），发现增加奖励边际会导致能力损失呈现等级化和任务选择性。具体而言，模型执行提取任务的能力基本保持不变，而其在 JSON 任务上的表现则略有下降，其路由器能力受到严重影响。研究表明，这种行为损害主要是由于被拒绝的补全被抑制，而不是选择的奖励发生了变化。 AI

影响这项研究强调了特定的训练目标如何选择性地降低模型的能力，从而影响路由等复杂任务的可靠性。

排序理由学术论文，详细描述了在特定训练目标下模型能力的经验表征。[lever_c_demoted from research: ic=1 ai=1.0]

在 Hugging Face Daily Papers 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: 0 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 学术论文，详细描述了在特定训练目标下模型能力的经验表征。[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: 6 days old; Aged out of breaking-news scoring windows; ranking reflects the durable signal from the full source set.

完整方法见我们的编辑标准。

报道来源 [1]

Hugging Face Daily Papers TIER_1 English(EN) · 2026-11-04 13:18

对齐税的等级：长度归一化DPO与SFT保留下的任务选择性能力损失

Setting. This is an empirical characterization study; we propose no new method and run no comparisons against alternative objectives. We fine-tune Qwen3.5-4B with sparse LoRA and length-normalized DPO plus an SFT retention term (the RPO objective) on structured financial-filing t…

本文如何被排名

报道来源 [1]

对齐税的等级：长度归一化DPO与SFT保留下的任务选择性能力损失

相关实体

相关话题