Truth as a Compression Artifact in Language Model Training
真相作为语言模型训练中的压缩特征
专题命中 效率与部署 :language model(title,abstract);pretraining(abstract);分类 cs.CL、cs.AI
AI总结 研究发现语言模型在矛盾数据训练中更倾向于正确答案,这源于错误的可压缩性而非真相本身。通过小规模Transformer模型实验,发现当错误遵循一致规则时,模型无法区分虚假系统与真相,提出压缩-一致性原则解释这一现象。
Comments v3: Added Qwen3 architecture check (0.6B), ~1B experiment on FineWeb-Edu, generative eval at all scales, matched-random ablation. Formal MDL predictions. Softened claims, fixed bibliography, added NeurIPS checklist. 210+ models (was 160+)