Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence

ID: a3d9c22a
原文链接: https://arxiv.org/abs/2609.05385
PDF: https://arxiv.org/pdf/2609.05385v2
作者: Urja Pawar, Rajitha Ramanayake, Nabeel Kemal, Ashwin Kandath, Owen O'Neill, Guillaume Bourgeon, Houssem Chatbri, Christopher Martin, Vadim Pertsovskiy
发布日期: 2026-09-04
更新日期: 2026-09-07
arXiv 分类: cs.AI
条目分类: learning
来源类型: paper
标签: llm-eval, agent-oversight, explainability, behavioral-test
质量评分: 4/5
简评作者: openclaw
抓取时间: 2026-09-09 (UTC+8)

中文导读

零位黑箱限制下用控制变量评估 LLM 输出以外的紧要性与充足性；在贵建计伐估项与伤害性判断两个任务上对 Claude/GPT/Gemini 八个模型的 top-3 引用因子进行测量，与行为分跳相关仅约 0.35～0.58作为安全可靠性检查间隔，适合报 AI 类趋势/代理路线上的嵌入使用

为什么值得关注

零位黑箱限制下用控制变量评估 LLM 输出以外的紧要性与充足性；在贵建计伐估项与伤害性判断两个任务上对 Claude/GPT/Gemini 八个模型的 top-3 引用因子进行测量，与行为分跳相关仅约 0.

要点摘录：

来源：arXiv 论文页面元数据 + 摘要
通过黑箱受控干预衡量 LLM 解释因素的真实影响，与行为观测对比
价值在于为代理流程中的解释提供可复现的可靠性检查

关键信息

论文标题：Necessary or Sufficient? Evaluating LLM Explanations With Behavioural Evidence
作者：Urja Pawar, Rajitha Ramanayake, Nabeel Kemal, Ashwin Kandath, Owen O'Neill, Guillaume Bourgeon, Houssem Chatbri, Christopher Martin, Vadim Pertsovskiy
arXiv：https://arxiv.org/abs/2609.05385
PDF：https://arxiv.org/pdf/2609.05385v2
发布时间：2026-09-04
更新：2026-09-07
arXiv 分类：cs.AI
关联标签：llm-eval, agent-oversight, explainability, behavioral-test

English Abstract

LLM decision components that can operate within agent workflows often produce action-relevant recommendations or judgements together with explanations. Operators may use the named factors to monitor a system, diagnose errors, or decide when to escalate an output. Such use assumes that the explanations agree with the component's observable decision behaviour. We test two interpretations of the named factors: necessity, meaning that changing a factor would change the output, and sufficiency, meaning that retaining it while removing other changeable information would preserve the output. We evaluate these interpretations in two synthetic use cases: recommending advisors to clients and judging prompts for harmfulness or risk. Models return an output and the top three factors that most influenced it. Controlled black-box interventions estimate a necessity score for each factor by measuring how often changing it changes the output, and a sufficiency score by measuring how often retaining it preserves the output. Across eight models from the Claude, GPT, and Gemini families, the mean Spearman correlations between the cited ranking and the necessity and sufficiency scores are 0.349 and 0.354 for advisor recommendation, and 0.431 and 0.580 for prompt monitoring. Furthermore, an uncited factor scores above the lowest-scoring cited factor in 57.6% of advisor responses under necessity and 58.1% under sufficiency; the corresponding prompt-monitoring rates are 25.8% and 8.9%. The cited top three contain useful information but do not reliably identify the three factors with the strongest measured influence under necessity or sufficiency. The framework provides a black-box reliability check for explanations used in agent oversight while remaining scoped to individual LLM decisions.

English Summary

Black-box behavioural evaluation of LLM explanations in agent workflows. Defines necessity and sufficiency scores via controlled interventions over cited top-3 factors across advisor recommendation and prompt monitoring tasks. Across eight Claude/GPT/Gemini models, mean Spearman correlations between cited ranking and necessity/sufficiency scores are 0.349/0.354 (advisor) and 0.431/0.580 (prompt monitoring). In 57.6% of advisor responses an uncited factor scores above the lowest-scoring cited factor under necessity, showing cited factors do not reliably identify the most influential inputs. Useful as a reliability check for explanations used in agent oversight while staying scoped to individual decisions.

Obsidian Notes

内容由 opencli arxiv paper 拉取 arXiv 元数据与摘要生成。
中文导读与价值判断锚定在条目已有摘要、论文摘要、作者、日期与分类信息；未补充论文摘要之外的实验细节。