发现大模型事实知识在最后层突然涌现，可用来提升干预效果。
MechLens: Late Crystallization of Factual Knowledge Explains Intervention Effectiveness in Language Models
- 知识不是逐步出现，而是在最后一层突然结晶。
- 93.4%的正确答案在中间层从未进入前10预测，跨模型跨任务均成立。
- 适合想改进幻觉、优化模型干预的研究者和工程师。
理解大语言模型如何存储事实知识对缓解幻觉至关重要。本文系统量化了‘晚期结晶’现象：事实知识并非渐进式浮现，而是在最终层突然形成。在五个模型家族（Pythia、Gemma、Qwen2.5、Llama-3.1、Mistral；0.5–14B）中，26.8%–93.4%的正确答案从未在任何中间层进入前10预测，且晚出现（>80%深度）在不同架构间一致。跨规模（Qwen2.5-14B）与跨基准（MMLU: 98.2%）结果验证其普遍性；调优透镜排除了探测伪影。情感分类对照实验（Qwen: 0.5% vs. 85.9%事实；Mistral: 2.0% vs. 26.8%）确认该现象仅针对事实召回。晚期结晶催生出结晶引导的干预原则：在中等结晶模型上（如Llama、Mistral），CAA优于DoLa（p<0.001），而在高结晶模型Qwen上则反向显著（MC1 +25.4% vs. +15.5%，p=0.069）。LayerNorm消融表明结晶是残差流内在特性；LN缩放（x1.2）实现+11.8% MC1，无推理开销。进一步揭示计算型知识（第22.1/28层）比记忆型事实（28.0/28）更早结晶。已开源MechLens，支持五类模型。
原文摘要 · Abstract (English)
Understanding where LLMs store factual knowledge is critical for hallucination mitigation. We systematically quantify Late Crystallization: factual knowledge does not gradually emerge across layers but "crystallizes" abruptly at the final layers. Across five model families (Pythia, Gemma, Qwen2.5, Llama-3.1, Mistral; 0.5--14B), 26.8%--93.4% of correct answers never enter top-10 predictions at any intermediate layer, with late emergence (>80% depth) consistent across architectures. Cross-scale (Qwen2.5-14B) and cross-benchmark (MMLU: 98.2%) results confirm generality; tuned lens rules out probe artifacts. A sentiment-classification control (0.5% for Qwen vs. 85.9% factual; 2.0% for Mistral vs. 26.8%) confirms the phenomenon is specific to factual recall. Late Crystallization yields a crystallization-guided intervention principle: CAA outperforms DoLa on moderate-crystallization models (Llama, Mistral; p<0.001), with a directionally consistent reversal on high-crystallization Qwen (+25.4% vs. +15.5% MC1, p=0.069). LayerNorm ablation shows crystallization is intrinsic to the residual stream; LN scaling (x1.2) yields +11.8% MC1 with zero inference overhead. We further reveal a Computability-Memorization Spectrum: computable knowledge crystallizes earlier (layer 22.1/28) than memorized facts (28.0/28). We release MechLens supporting five model families.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。