EVENT DOSSIER
A study on a 0.6B parameter language model found that when verifying 1,200 logical conclusions, its behavior could not distinguish between true and false (accuracy was only 50%). However, linear probes could accurately read the correct judgments in the hidden states (AUC 0.96). The study indicated that the dominant factor causing judgment loss was a single scalar: the judgment information along the alignment direction in the model’s logits was intact (AUC 0.89), but it was erased by a saturation decision threshold with an offset of +4.6 standard deviations. This diagnosis is universal; in experiments with 90 semantic label configurations and across models, behavior accuracy collapsed in a single-function relationship with the threshold offset (Spearman -0.93), while marginal rankings were less affected. The study proposed an actionable solution: by correcting with a single parameter, calibrating marginal decoding, or adjusting prompts, the threshold can be re-centered, improving the behavior accuracy of the 0.6B model from 50% to 81%, and restoring the 8B model to 94%.
一项针对 0.6B 语言模型的研究发现，其在验证 1,200 个逻辑结论时虽行为上无法区分真伪（准确率 50%），但线性探针可准确读取隐藏状态中的正确判决（AUC 0.96）。研究指出，导致判决丢失的主导因素是单个标量：模型输出 logits 中沿对齐方向的判决信息完整（AUC 0.89），却被一个偏移 +4.6 个标准差的饱和决策阈值擦除。该诊断具有普适性：在 90 种语义标签配置及跨模型实验中，行为准确率随阈值偏移呈单函数关系崩溃（Spearman -0.93），而边际排名受影响较小。研究提出可操作方案：通过单一参数校正、校准边际解码或调整提示词，将阈值重新中心化，可将 0.6B 模型的行为准确率从 50% 提升至 81%，使 8B 模型恢复至 94%。