Translation in progress Some entries still include text in the original language.
BackSource OpenAI Alignment Research
Tom Dupre la Tour and Dan Mossing, in collaboration with the Interpretability team
Indexed by this site 2026-10-01 03:58 12 days ago
Originally published Dec 2, 2025
Selected 2026-10-01 03:58
Category Alignment OpenAI Alignment Research · Tom Dupre la Tour and Dan Mossing, in collaboration with the Interpretability team · Indexed by this site 2026-10-01 03:58 · Originally publishedDec 2, 2025 SelectedAttention53
OpenAI 用 SAE 潜变量归因调试失准输出 Debugging misaligned completions with sparse-autoencoder latent attribution
AI summary
OpenAI 对齐研究团队提出用稀疏自编码器(SAE)潜变量归因来定位与失准行为因果相关的特征,替代此前依赖两个模型对比的模型差分方法。该方法用一阶泰勒展开近似每个潜变量对补全交叉熵损失的贡献,在同一提示的正负补全之间计算归因差值,再通过激活引导和 GPT-5 评分单独验证因果性。在涌现性失准案例中,团队从同一提示采样 35 对对齐与失准补全,选出归因差值最大的 top-100 潜变量,其可引导模型远离或走向失准的潜变量数量与平均变化幅度均高于按激活差值选出的 top-100。在不当认可案例中,团队使用 148 对补全,得到类似结果。
Why it was selected
OpenAI 对齐团队用 SAE 潜变量归因替代模型差分,定位到同时驱动两类失准行为的单一特征,为机制可解释性审计提供了可迁移方法。
Read original alignment.openai.com Why it was selected SelectedAttention53
Attention 0–100: content value contributes up to 70 and circulation up to 30; circulation decays over time. How it was rated
OpenAI 对齐团队用 SAE 潜变量归因替代模型差分,定位到同时驱动两类失准行为的单一特征,为机制可解释性审计提供了可迁移方法。