Research Papers 论文研究 4h ago Updated 21m ago 更新于 21分钟前 46

From Causal Plausibility to Causal Reliability: Evaluating LLMs as Calibrated Direct Causal-Edge Classifiers 从因果合理性到因果可靠性:评估LLM作为校准的直接因果边分类器

LLM-based causal judgments are strongly recall-dominant, producing overly dense graphs with many false-positive edges; prompting shifts the precision-recall trade-off but does not resolve overprediction LLMs capture causal relatedness but fail to reliably identify directness or orientation, misclassifying 40.0% of indirect and 36.0% of reversed non-edges as direct edges 80.8% and 84.6% of false positives from indirect and reversed non-edges receive verbalized confidence of at least 80%, revealin 系统评估12个开源LLM在6个因果图基准上的直接边分类能力,发现模型倾向于过度预测边(高召回、低精确) LLM能捕捉因果相关性但无法可靠识别直接性和方向性,40%间接边和36%反向非边被误判为直接边 传统置信度估计不可靠:logit置信度常坍缩至1.0,80.8%的误判获得≥80%口头置信度 跨提示/跨模型一致性比传统置信度更可靠,但统计显著性经Holm校正后不显著 5个模型-数据集对存在基准熟悉度问题(均涉及AsiaM),建议将LLM视为需外部验证的软因果先验而非直接证据

58
Hot 热度
74
Quality 质量
66
Impact 影响力

Analysis 深度分析

TL;DR

  • LLM-based causal judgments are strongly recall-dominant, producing overly dense graphs with many false-positive edges; prompting shifts the precision-recall trade-off but does not resolve overprediction
  • LLMs capture causal relatedness but fail to reliably identify directness or orientation, misclassifying 40.0% of indirect and 36.0% of reversed non-edges as direct edges
  • 80.8% and 84.6% of false positives from indirect and reversed non-edges receive verbalized confidence of at least 80%, revealing substantial overconfidence in structurally incorrect predictions
  • Conventional confidence estimates (logit-based) are unreliable, collapsing near 1.0 regardless of correctness, while cross-prompt and cross-model agreement show better calibration though not statistically significant after Holm correction
  • Benchmark-familiarity audit flags potential data contamination in five model-dataset pairs involving AsiaM, reinforcing the conclusion that LLMs should serve as externally validated soft causal priors rather than direct evidence of causal structure

Why It Matters

This work directly addresses a growing practice in the AI community where LLMs are used to supply prior causal knowledge for structural causal discovery, yet their reliability as direct edge classifiers remains unproven. For practitioners building causal models or integrating LLMs into scientific workflows, the findings serve as a critical caution against trusting raw LLM causal judgments and their confidence scores without external validation. The results reshape how researchers should deploy LLMs in causal inference pipelines, shifting expectations from direct structural evidence to calibrated soft priors.

Technical Details

  • Evaluation scope: 12 instruction-tuned open-weight LLMs tested across 6 benchmark causal graphs, 5 prompting strategies, and 4 confidence estimation methods (verbalized, logit-based, cross-prompt agreement, cross-model agreement)
  • Protocol: Language-only pairwise causal-edge classification under a structured evaluation framework comparing LLM predictions against published reference graphs
  • Key metrics: Precision-recall analysis, calibration assessment, and error classification distinguishing indirect edges, reversed edges, and other non-edges
  • Confidence analysis: Logit-based confidence was found to collapse near 1.0 irrespective of prediction correctness; verbalized confidence showed 80.8% and 84.6% of false positives receiving scores above 80% for indirect and reversed non-edges respectively
  • Data contamination audit: Identified potential benchmark familiarity in 5 model-dataset pairs, all involving the AsiaM dataset, raising concerns about evaluation validity

Industry Insight

  • Researchers and engineers should treat LLM-generated causal edges as hypothesis-generating priors requiring independent empirical validation rather than as standalone evidence; this applies to any workflow using LLMs for scientific causal discovery
  • Confidence calibration remains a critical unresolved challenge—logit-based and verbalized confidence scores from current LLMs are unreliable indicators of prediction correctness, suggesting the need for dedicated calibration methods before deploying LLMs in high-stakes causal reasoning tasks
  • The diminishing returns of model scale on larger graphs and the lack of statistical significance in agreement-based calibration improvements indicate that simply scaling up models will not solve the fundamental overconfidence and overprediction problems in LLM-based causal edge classification

TL;DR

  • 系统评估12个开源LLM在6个因果图基准上的直接边分类能力,发现模型倾向于过度预测边(高召回、低精确)
  • LLM能捕捉因果相关性但无法可靠识别直接性和方向性,40%间接边和36%反向非边被误判为直接边
  • 传统置信度估计不可靠:logit置信度常坍缩至1.0,80.8%的误判获得≥80%口头置信度
  • 跨提示/跨模型一致性比传统置信度更可靠,但统计显著性经Holm校正后不显著
  • 5个模型-数据集对存在基准熟悉度问题(均涉及AsiaM),建议将LLM视为需外部验证的软因果先验而非直接证据

为什么值得看

本文为LLM在因果发现领域的应用提供了系统性评估,揭示了当前模型在因果结构识别上的根本性局限。对AI从业者而言,研究结果警示了盲目信任LLM因果判断的风险,为构建更可靠的因果推理系统提供了实证依据。

技术解析

  • 评估框架:采用纯语言配对协议,在6个基准因果图上测试12个指令微调开源模型,覆盖5种提示策略和4种置信度来源(口头化、logit、跨提示一致性、跨模型一致性)
  • 核心发现一:LLM因果判断呈强召回主导特征,预测图过于密集且假阳性边众多;提示策略仅改变精确-召回权衡而非解决过度预测;模型规模增益在最大图上递减且无法消除校准偏差
  • 核心发现二:模型误分类率差异显著——40.0%间接边和36.0%反向非边被误判为直接边,远高于其他非边的28.2%;80.8%和84.6%的此类误判获得≥80%口头置信度,显示结构性错误预测存在严重过度自信
  • 核心发现三:logit置信度在正确性无关时频繁坍缩至1.0;跨提示和跨模型一致性实现更好的平均校准和区分能力,但经Holm校正后优势无统计显著性;基准熟悉度审计发现5个模型-数据集对存在潜在熟悉度(均涉及AsiaM)

行业启示

  • 将LLM用于因果发现时应定位为"需外部验证的软先验源"而非因果结构直接证据,建议结合传统因果发现算法进行交叉验证
  • 开发因果AI系统时需重视置信度校准问题,跨模型/跨提示一致性可作为更可靠的置信度代理指标,但需进一步研究提升统计效力
  • 基准数据集设计应避免熟悉度偏差,AsiaM等数据集的潜在泄露问题提示社区需建立更严格的评估协议和去熟悉化机制

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

LLM 大模型 Evaluation 评测 Benchmark 基准测试 Research 科学研究