发表机构
University of Wisconsin-Madison（威斯康星大学麦迪逊分校）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究针对LLM作为评判者的摘要评估机制，通过多组实验分析了Themis和Prometheus等模型的评估流水线结构及微调的作用，并发布了相关代码与数据。
AI 中文摘要
基于大语言模型（LLM）的自然语言生成（NLG）质量评估器被广泛用作评分工具和自动训练信号，但它们分配评分的内部过程仍未被充分理解。我们通过针对NLG质量的可读性和充分性维度的八项攻击扰动分类法、生成具有可控错误强度和显式token级修改映射的成对干净与损坏摘要的生成流水线，以及应用于Themis（Llama-3-8B）和Prometheus（Mistral-7B）的因果追踪、logit-lens词汇投影和注意力头 knockout 的四项实验，从机制上研究了这一过程。两种评估器都实现了一个结构化、连贯的两阶段评估流水线：在第15层以下，注意力执行局部错误比较并将结果路由到最终输入位置；在第15层以上，MLP级联整合信号并写入评分，决策在较晚的层（Themis为L=26，Prometheus为L=25）的残差流中明确形成。此外，相同规模的基础模型（Llama-3-8B）重现了路由架构和决策形成，但未重现阶段分离，从而分离出微调专门引入的两种机制：抑制第15层以下MLP在最后位置的贡献，以及决策形成深度提前两层，表明微调是在现有基础上塑造而非从头构建流水线。我们在该https URL发布了源代码和数据。
英文摘要
LLM-based evaluators of natural language generation (NLG) quality are widely deployed as scoring tools and as automated training signals, yet the internal procedure by which they assign a rating remains poorly understood. We investigate this procedure mechanistically through an eight-attack perturbation taxonomy across the Readability and Adequacy dimensions of NLG quality, a generation pipeline that produces paired clean and corrupt summaries with controlled error intensity and explicit token-level modification maps, and a four-experiment battery of causal tracing, logit-lens vocabulary projection, and attention-head knockout applied to Themis (Llama-3-8B) and Prometheus (Mistral-7B). Both evaluators implement a structured, coherent evaluation pipeline operating in two stages: below layer 15, attention performs local error comparison and routes the result to the final input position; above it, the MLP cascade integrates the signal and writes the rating, with the decision crystallizing in the residual stream at a sharp late layer (L = 26 on Themis, L = 25 on Prometheus). Furthermore, a base-model control at the same scale (Llama-3-8B) reproduces the routing architecture and crystallization but not the stage separation, isolating the two mechanisms that fine-tuning specifically installs, suppression of below-L15 MLP contribution at the last position and a two-layer advance of the crystallization depth, indicating that fine-tuning sculpts an existing substrate rather than building the pipeline from scratch. We release the source code and data at https://github.com/himil-v/judge-mech
CommentsAccepted at EMNLP 2026 Main Conference
