CXBench: a visual introduction
Original paper figures with animated annotations. English narration and synchronized English / Chinese subtitles.
01 / 00:00:00
The destination is only half the story
Imagine two AI agents helping a customer dispute a transaction.
设想两个 AI 智能体，正在帮助客户对一笔交易提出争议。
Both file the correct dispute. Their final database records are identical.
它们都正确提交了争议，最终的数据库记录完全相同。
Does that mean both agents succeeded? The answer depends on the path they took.
这是否意味着它们都成功了？答案取决于它们采取的过程。
02 / 00:00:14
Same outcome. Different process.
In this example, the bank requires written confirmation of an oral error notice.
在这个例子中，银行要求客户书面确认口头提出的错误通知。
One agent explains the requirement and gives the submission address. The other omits that notice.
一个智能体解释了这一要求，并提供提交地址；另一个却省略了这项告知。
Both obtain draft confirmation. Both reach the same outcome. But only one satisfies the required disclosure.
它们都获得草稿确认，也得到相同结果，但只有一个履行了必要的披露义务。
03 / 00:00:34
Separate outcome from process
CXBench separates two questions. Did the agent achieve the customer goal? And did it follow the required process?
CXBench 区分两个问题：智能体是否实现客户目标？是否遵循必要流程？
Overall success requires both. A correct outcome with a failed process is a false-pass.
整体成功要求两者同时通过。结果正确、流程失败，就是一次 false-pass，即假通过。
It is precisely the kind of failure that outcome-only evaluation would accept.
这种失败恰好会被只检查最终结果的评估方式判为成功。
04 / 00:00:53
Make every requirement inspectable
Rule cards connect source provisions to applicable behavior and observable trace evidence.
规则卡将来源条款、适用行为和交互轨迹中的可观察证据连接起来。
Deterministic rules check ordering and authority. An LLM judge checks semantic obligations, such as disclosures and unsupported promises.
确定性规则检查时序和权限；大模型裁判检查披露、无依据承诺等语义要求。
The process passes only when both components pass. Each finding can be inspected against the recorded interaction.
两个评分组件都通过，流程才通过。每项判断都可以对照交互记录检查。
05 / 00:01:17
Separate preparation from authority
In the default runtime, the agent prepares a draft. The customer confirms the required fields.
默认运行模式下，智能体准备草稿，客户确认必要字段。
The runtime submits only when the draft still matches that confirmation. This is a benchmark control, distinct from the disclosure obligation.
只有草稿仍与确认内容一致，运行时才会提交。这是基准的控制机制，与披露义务不同。
Controlling submission helps enforce selected constraints. It does not replace evaluation of the whole process.
控制提交权限有助于落实特定约束，但不能替代对整个流程的评估。
06 / 00:01:40
The gap becomes visible
The study covers one hundred and twenty tasks, six regulated domains, and seven models: eight hundred and forty retained executions.
研究覆盖 120 个任务、6 个受监管领域、7 个模型，共保留 840 次执行记录。
Five hundred and eighty-two pass the outcome check. Only one hundred and sixteen pass both outcome and process.
其中 582 次通过结果检查，只有 116 次同时通过结果和流程检查。
The remaining four hundred and sixty-six are false-passes. That is about eighty percent of the runs that outcome-only scoring accepts.
其余 466 次是假通过，约占结果评估判为成功的执行记录的百分之八十。
07 / 00:02:04
A different criterion changes the picture
The evaluation criterion also changes the observed model ordering.
评估标准也会改变观察到的模型排序。
DeepSeek and GLM tie on outcome success, yet their both-pass rates differ. GPT five point four has a lower outcome rate, but a higher both-pass rate than DeepSeek.
DeepSeek 和 GLM 的结果通过率相同，但双通过率不同。GPT 5.4 的结果通过率较低，双通过率却高于 DeepSeek。
These are descriptive comparisons. The study does not establish a complete statistical ranking.
这些是描述性比较；研究并未确立完整且具有统计显著性的排序。
08 / 00:02:26
Evaluate the outcome. Examine the process.
Among the false-passes, the semantic judge alone catches three hundred and twenty-nine. Hard rules alone catch thirty-eight. Both catch ninety-nine.
在假通过中，语义裁判单独识别 329 次，确定性规则单独识别 38 次，两者共同识别 99 次。
These complementary checks measure the requirements encoded in the benchmark, not legal certification. Judge validation and repeated-run uncertainty remain important limits.
这些互补检查衡量基准中编码的要求，而不提供法律认证。裁判验证和重复运行的不确定性仍是重要限制。
CXBench asks us to evaluate the outcome, and examine the process. Explore the paper at cxbench dot github dot io.
CXBench 提醒我们：评估结果，也审视过程。欢迎访问 cxbench.github.io 阅读论文。
Figures 1–6 and Table 3 are from the supplied October 2026 manuscript. The video uses cropped views, zooms, and animated annotations; full original figures are available on the project website. Narration is synthesized with the Andrew neural voice.