Transformer 中证据效应的因果定位
摘要
论文提出受控证据聚合（CEA）与平衡效用因果选择（BUC）方法，通过符号探针定位出可影响Transformer自然语言决策的紧凑注意力头集合，并在Pythia-410M上验证其因果效应。
arXiv:2610.36524v1 Announce Type: new
Abstract: Can controlled symbolic inputs reveal attention heads that influence a language model's decisions from natural language evidence? We introduce controlled evidence aggregation (CEA), which varies the counts and order of two symbols in otherwise fixed prompts and measures changes in a frozen Transformer's preference between two answer tokens. The preference is measured by their logit difference; the prompts prescribe no correct answer. Balanced utility causal selection (BUC) identifies compact head sets by jointly evaluating how much ablating their outputs removes the count effect and how closely patching mean outputs between count conditions recovers the effect in both directions. We derive conditional bounds linking head attribution to intervention effects and specifying when the attribution shortlist can be recovered from finite samples. In Pythia-410M, ablating sets of 1-12 heads selected separately for four ordering regimes removes 65.5-91.6% of the count effect. The sets retain their influence on count comparisons not used for selection. To test relevance beyond symbolic inputs, we evaluate a natural language task asking which state is supported by more reports. Ablating eight heads selected only on symbolic yes/no prompts lowers mean accuracy by 8.2 percentage points, versus 0.8 for random sets matched by size and layer. Our results demonstrate that controlled symbolic probes can identify compact attention head sets whose causal influence extends to natural language decisions.
查看缓存全文
缓存时间: 2026/10/28 08:04
# Causal Localization of Evidence Effects in Transformers Source: [https://arxiv.org/abs/2610.36524](https://arxiv.org/abs/2610.36524) [View PDF](https://arxiv.org/pdf/2610.36524)[HTML \(experimental\)](https://arxiv.org/html/2610.36524v1) > Abstract:Can controlled symbolic inputs reveal attention heads that influence a language model's decisions from natural language evidence? We introduce controlled evidence aggregation \(CEA\), which varies the counts and order of two symbols in otherwise fixed prompts and measures changes in a frozen Transformer's preference between two answer tokens\. The preference is measured by their logit difference; the prompts prescribe no correct answer\. Balanced utility causal selection \(BUC\) identifies compact head sets by jointly evaluating how much ablating their outputs removes the count effect and how closely patching mean outputs between count conditions recovers the effect in both directions\. We derive conditional bounds linking head attribution to intervention effects and specifying when the attribution shortlist can be recovered from finite samples\. In Pythia\-410M, ablating sets of 1\-12 heads selected separately for four ordering regimes removes 65\.5\-91\.6% of the count effect\. The sets retain their influence on count comparisons not used for selection\. To test relevance beyond symbolic inputs, we evaluate a natural language task asking which state is supported by more reports\. Ablating eight heads selected only on symbolic yes/no prompts lowers mean accuracy by 8\.2 percentage points, versus 0\.8 for random sets matched by size and layer\. Our results demonstrate that controlled symbolic probes can identify compact attention head sets whose causal influence extends to natural language decisions\. ## Submission history From: Jipeng Li \[[view email](https://arxiv.org/show-email/1e1509d9/2610.36524)\] **\[v1\]**Tue, 27 Oct 2026 02:27:36 UTC \(288 KB\)
相似文章
相同证据，不同目标：从语言模型状态解码诊断证据如何关联因果问题
本文研究语言模型是否能正确判断诊断证据如何支持或挑战不同的因果主张，并引入了仅改变因果目标的配对提示。在Qwen2.5-7B-Instruct等模型的倒数第二个Transformer隐藏状态上进行的线性读出显示，平衡准确率适中（0.654-0.659），并在49对中恢复了18-21对，表明因果相关性具有一定程度的线性可解码性。
各有不足：在不同证据状态下校准语言模型
提出证据状态校准（evidence-state calibration）概念与SuffiCE方法，通过微调模型下一token分布中的置信度，使其随证据变化而准确反映答对概率，在临床诊断和多跳问答上将Brier分数降低最高60%、AUROC提升最高0.16。
对智能体因果发现进行基准测试与增强
论文提出 Heuresis，一个使用匿名模拟数据的因果发现基准，用于评估 LLM 智能体在数据驱动下进行细粒度、证据驱动的因果发现调整能力。结果显示 GPT-6.1 Sol 在全部六个数据族上超过最佳算法选择，最难族上邻接 F1 提升 14%。
Evidence-RL：迈向证据密集型视觉推理
本文介绍了反事实证据解耦（CED），这是一种训练时方法，使视觉语言模型依赖具体的图像证据而非语言先验或捷径，从而提升跨基准的视觉推理grounding能力。
面向循环 Transformer 的因果潜在工作区
该论文提出因果潜在工作区 Transformer（causal latent workspace Transformer），通过在共享深度的循环块之间加入可读写的前缀专属潜在槽位，提升了循环 Transformer 的语言建模困惑度，并在三个字节级语料上的多数设置中优于五个基线对照。