Transformer通过隐藏状态旋转分离来拒绝错误答案，但并非因果机制。
How Transformers Reject Wrong Answers: Rotational Dynamics of Factual Constraint Processing
- 模型在处理正确与错误单令牌续写时，隐藏状态向量发生角度分离，中层旋转明显。
- 错误路径的对数几率远低于随机预期，错误项概率高出约11.5倍。
- 该现象跨7个模型、3个领域、300个查询稳定存在，适合研究模型决策机制。
当解码器仅有的Transformer被强制处理事实性问题的正确与错误单令牌续写时，两条路径在隐藏状态空间中逐渐分离：从仅含问题的表示出发的位移向量保持相近幅度但相互旋转，角距离在中层逐渐增大，晚期层才出现不对称结果。错误路径的对数几率灯谱显著低于等概率先验（错误项概率约为正确项的11.5倍）。这一模式——旋转分离后晚期层的非对称承诺——可视为模型外部表现为拒绝错误续写的几何特征，但仅为观察性而非因果性：错误路径也可能反映模型顺应被强制输入的令牌，需随机令牌对照实验才能确认。该现象在六个解码器型Transformer（Llama、Mistral、Gemma、StableLM）中均成立，参数规模1B至13B；第七个模型（Qwen2 1.5B）在本协议下表现平坦，可能为分词器偏差所致，暗示涌现阈值尚存。线性探针可在中间层恢复区分能力，跨领域迁移呈现结构不对称性，金融-医疗通道优于交通类配对。单层激活修补在部分模型（LLaMA-2 13B、Mistral 7B）中可解释，但无一致恢复层带；另一模型（StableLM-2 1.6B）在所有层均超出上限，诊断为代码路径偏差并排除。在此严格控制下，晚期不对称性不局限于单一组件，支持分布式轨迹解释而非单层局部回忆。通过七模型、三领域、300个查询的强制续写探测验证了该现象。
原文摘要 · Abstract (English)
When a decoder-only transformer is forced to process matched correct and incorrect single-token continuations of a factual query, the two pathways through hidden-state space diverge: displacement vectors from the query-only representation keep near-equal magnitude but rotate apart, with angular separation growing through mid-depth before late layers resolve an asymmetric outcome. A logit-lens preference in the incorrect run falls far below the equal-probability prior (roughly 11.5x more mass on the incorrect token than the correct one). We read this pattern, rotational divergence then late-layer asymmetric commitment, as the geometric signature of the model externally appearing to reject a wrong continuation, while staying explicit that it is observational, not causal: the incorrect run could equally reflect the model conforming to the token it is forced to carry, which only a random-token control can settle. It holds across six decoder-only transformers spanning four architecture families (Llama, Mistral, Gemma, StableLM) from 1B to 13B parameters; a seventh (Qwen2 1.5B) is flat under our protocol, plausibly a tokenizer artefact, leaving an emergence threshold open. Linear probes recover the distinction at intermediate depth, and cross-domain transfer is structurally asymmetric, a financial-medical corridor transferring far better than transport pairs. Where single-layer activation patching is cleanly interpretable (LLaMA-2 13B, Mistral 7B) it yields no layer band of consistent recovery; a third model (StableLM-2 1.6B) recovers uniformly above the ceiling, which we diagnose as a code-path artefact and exclude. Under this scoped null the late-layer asymmetry is not localized to a single component, fitting a distributed-by-trajectory account rather than single-layer localized recall. We document this with forced-completion probing across seven models, three domains, and 300 queries.
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。