Event ·Official first-hand

用目标冲突样本微调，提升模型指令层级遵循能力

指令层级（IH）要求 LLM 在用户请求与系统策略冲突时优先遵循系统策略。研究者将任务分解为目标和约束后发现一类现有训练集缺失的冲突类型——目标-目标冲突，OpenAI 和 NVIDIA 的 IH 训练集中都不包含该类型。在 Qwen3.5-9B 和 Gemma4-12B-it 上用这类样本微调后，模型在多个训练中未见的真实任务上 IH 表现明显提升，并入现有 IH 训练集微调也能带来更好效果。在八个基准的无冲突任务上，模型的平均效用变化微乎其微。

Current read

OpenAI 和 NVIDIA 的指令层级训练集都漏了目标-目标冲突这一类，这篇论文补上了，微调 Qwen3.5-9B 和 Gemma4-12B-it 就能明显提升效果，做对齐训练的可以看看。

11 sources38° AI heatLast updated 2026/10/26 09:33:17

Evidence

11 sources

Primary sourcearXiv: OpenAI

2026/10/26 09:33:17View source

Related source 1Decoder

2026/10/25 10:08:49View source

Related source 2theverge

2026/10/26 23:34:58View source

Related source 3The Information

2026/10/24 18:30:03View source

Related source 4IT之家

2026/10/24 23:27:45View source

Related source 5Nathan Lambert

2026/10/25 11:19:36View source

Related source 6TestingCatalog

2026/10/25 17:06:12View source

Related source 7Greg Brockman

2026/10/25 22:23:33View source

Related source 8爱范儿

2026/10/26 00:16:16View source

Related source 9Andrew Curran

2026/10/26 02:00:18View source

Related source 10Detik Inet

2026/10/26 04:40:27View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story