{"version":"2026-10-19.boundary.1","asOf":"2026-10-19","month":"2026-10","domain":{"id":"finance","title":"金融 · 资料与专业交付","summary":"本试点聚焦取数、检索、会议和跨文件交付。预测与交易的真实净收益证据另设门槛，不能由文档分数推得。","frontierId":"F20","sourceIds":["BS01"],"terms":[{"term":"财报取数 / 可执行推理","definition":"从表格和文字找到数字、财年和单位，再生成可运行公式；算对答案不等于财务解释全部正确。"},{"term":"Agent框架 / 验收项","definition":"框架提供模型可调用的工具与流程；验收项逐条检查来源、数字、解释等要求。"},{"term":"宏平均 / 多数票","definition":"先计算每道题满足验收项的比例，再给每题同等权重平均；三个模型多数票不是专业人士审定。"},{"term":"CRM / 样本外净收益","definition":"CRM保存客户工作记录；样本外用未参与开发的数据检验，净收益还需扣除交易和运营成本。"}]},"focusTask":"F01","caseStudy":{"id":"F01","explanationId":"calculation-with-tools","title":"财报计算：把取数、运算、对账分别验收","lead":"让AI找数和写公式，让程序计算；你核对财年、单位、口径和原文。计算正确仍可能引用错数据。","fit":"公开财报问答支持数值步骤的方法参考；真实报表、研究报告和投资决定需各自验收。","editorialAt":"2026-10-19","sourceIds":["BF01","BF02","BF06","BF07","R-JOBBENCH"],"steps":[{"title":"确认指标和资料","actor":"人决定","action":"锁定报告版本、财年、币种和指标定义。","output":"指标口径与允许资料","gate":"明确同比/环比及可比口径","judgment":"F01-step-1","editorialReason":"本步骤是本站为交付和接收提出的组织安排；没有本地逐步成功率。","evidenceRelations":[]},{"title":"提取数值与出处","actor":"AI草拟","action":"从给定资料列数值、单位和段落；找不到不猜。","output":"带页码/段落的取数表","gate":"每个数回到原文、财年/单位一致","judgment":"F01-step-2","editorialReason":"本步骤是本站为交付和接收提出的组织安排；没有本地逐步成功率。","evidenceRelations":[{"sourceId":"BF02","role":"condition","reason":"相近问答中资料配置影响结果；没有验证本工作逐数核验","locator":"FinanceBench指定配置实验"},{"sourceId":"R-JOBBENCH","role":"context","reason":"相邻方法启示，未验证本站案例；原有现场研究及其条件保留。","locator":"§2.1–2.3；附录E；§3.1"}]},{"title":"表达并执行公式","actor":"程序执行","action":"核对公式后由计算工具执行，保留输入与过程。","output":"公式、计算输入与结果","gate":"程序结果可复算，异常输入另列","judgment":"F01-step-3","editorialReason":"本步骤是本站为交付和接收提出的组织安排；没有本地逐步成功率。","evidenceRelations":[{"sourceId":"BF06","role":"support","reason":"直接支持限定财报问答中外部程序计算的比较；不能扩大到取数与全报表","locator":"v4 表2 FinQA Codex少样本贪心"}]},{"title":"对账后解释","actor":"人核验","action":"核对原文、公式和文字解释一致，再接收草稿。","output":"带引用解释与异常清单","gate":"没有来源或口径不一致时不交付","judgment":"F01-step-4","editorialReason":"本步骤是本站为交付和接收提出的组织安排；没有本地逐步成功率。","evidenceRelations":[{"sourceId":"BF07","role":"context","reason":"工作簿产物和rubric提示完整交付检查范围；不是本例成绩","locator":"固定revision，任务7b08cd4d…"}]}],"example":{"title":"收入、利润率与重述：短指标表","body":"【虚构数字与报告，人工构造；未执行AI任务】\n输入为两份资料A/B。报告A原列2024收入1200万元、营业利润120万元；报告B附注重述2024收入1250万元、营业利润125万元，并列2025收入1500万元、营业利润180万元。全部为同币种、同合并范围的示例假设。\n指标 | 比较输入与出处 | 公式 | 结果与检查\n收入增长 | 2024重述1250[B附注2]；2025为1500[B表1] | (1500-1250)/1250 | 20%；不能误用A的1200算成25%\n2024营业利润率 | 125/1250[B附注2] | 125/1250 | 10%\n2025营业利润率 | 180/1500[B表1] | 180/1500 | 12%\n利润率变化 | 12%与10% | 12%-10% | 增加2个百分点，不是增加2%\n每行另填：真实文件版本、页码/表/脚注、期间、币种/单位、口径是否重述、核验人。实际材料没有对应数字时写未找到，不能沿用此例。","note":"A/B报告、数字和页码全是构造。20%、25%、2个百分点仅为示例复算，不是模型成绩、真实公司或未来预测。"},"failures":[{"trigger":"把季度收入与全年收入相除","why":"运算能执行，但比较口径不同","action":"停止交付，核对期间后重新取数"},{"trigger":"给出正确公式，却引用不存在的页码","why":"计算可靠性不能证明资料真实性","action":"逐个回到允许原文；无对应内容则写未找到"},{"trigger":"忽略附注重述，仍用旧报告的分母","why":"跨报告运算无错，却比较了不同口径","action":"保留原值与重述值及出处；按目标口径重算并解释差异"},{"trigger":"把利润率增加2个百分点写成增长2%","why":"比例的差值与相对变化是不同指标","action":"写明公式与单位，核对文字解释再交付"},{"trigger":"材料期间或单位冲突，却直接合成结论","why":"局部结果不足以证明完整交付合格","action":"保留完整人时及原始依据，停止扩大委派并由人复核"}],"trial":{"scope":"两份允许使用的财报或同一报告的两期数据；锁定重述/合并范围，准备一张短指标表及出处/异常清单","baseline":"记录人独立取数/计算/引用核验所需时间和相同质量要求","acceptance":["逐个数值/期间/单位能回到真实原文","公式与程序输入、结果可复算","引用、解释与异常列表一致"],"timeFields":["活动人时：资料准备","活动人时：人工取数核验","活动人时：公式检查","活动人时：引用对账","活动人时：返工","活动人时：提示、交互与监督","活动人时：接管与适用的维护","另记：机器运行、等待及端到端历时；重叠活动不重复累计；若等待占用本人，注明计量约定"],"stop":"来源无法核对、口径变更、异常被遗漏时，暂停解释与交付"},"comparisons":[{"title":"PoT同模型系统比较","sourceId":"BF06","relationship":"部分覆盖","task":"FinQA财报问答1,147测试题","people":"非真实组织完整报表","setup":"Codex少样本贪心；CoT vs Python执行","outcome":"40.4% → 64.5%","metric":"答案指标","limitation":"不是财报全流程通过率；CoT相对容差0.001；未匹配全部成本"},{"title":"资料处理方式比较","sourceId":"BF02","relationship":"相近问答","task":"150题财务问答","people":"公开问题集","setup":"同GPT4-Turbo：全360份文件共享检索库 vs 预先给定相关财报的长上下文；超限截断","outcome":"同论文实验窗口配置29/150 vs 118/150","metric":"正确答案计数","limitation":"配置差异，不是两个年份能力进步；取数仍须核验"},{"title":"专业推理与委派偏好","sourceId":"R-JOBBENCH","relationship":"adjacent：方法相邻，不是本任务完成率","task":"多文件专业产物","people":"作者实验；本站没有参与者或任务复跑","setup":"偏好用于选题；固定协议、材料和judge后评产物","outcome":"测专业推理能力，没有测员工实际省时或岗位替代","metric":"本页不摘录分数；不与现有现场研究合并","limitation":"相邻方法启示，未验证本站案例；原有现场研究及其条件保留。"}],"receiptFields":["任务/版本/资料与环境","已授权范围与工具配置","实际输入/产物路径","全部尝试与逐项验收","最终接收/拒绝与接管","基线/准备/核验/返工人时","未知与未验证项","提示/交互/监督/接管/适用维护的活动人时；机器运行/等待/端到端历时及重叠计量约定"],"exampleStatus":"synthetic_not_ai_executed","actualCoverage":null,"evidenceStatus":"active","sourceRelations":[{"sourceId":"BF01","role":"support","reason":"支持研究本身所述范围；对本工作组织和收益仍需独立验证。"},{"sourceId":"BF02","role":"support","reason":"支持研究本身所述范围；对本工作组织和收益仍需独立验证。"},{"sourceId":"BF06","role":"support","reason":"支持研究本身所述范围；对本工作组织和收益仍需独立验证。"},{"sourceId":"BF07","role":"context","reason":"支持研究本身所述范围；对本工作组织和收益仍需独立验证。"},{"sourceId":"R-JOBBENCH","role":"context","reason":"相邻方法启示，未验证本站案例；原有现场研究及其条件保留。"}]},"caseOptions":[{"id":"F01","title":"财报计算：把取数、运算、对账分别验收","lead":"让AI找数和写公式，让程序计算；你核对财年、单位、口径和原文。计算正确仍可能引用错数据。"}],"tasks":[{"id":"F01","domainId":"finance","title":"从财报取数并计算指标","input":"财报表格/文字、财年与问题","output":"数值、公式和页码依据","human":"核对单位、财年、重述与会计口径","acceptance":["每个输入能定位原文","公式可复算","财年/币种一致"],"relation":"FinQA、FinanceBench的限定公开问题；企业私有报表效果未知","workTaskId":null,"actualCoverage":null}],"events":[{"id":"F01-2021","taskId":"F01","availableAt":"2021-09-01","eventDate":null,"dateBasis":"按该条支持材料的公开日期进入时间轴；不是实验统一发生日","title":"数字回答开始附可执行推理程序","kind":"benchmark","mode":"bounded","result":"FinQA共8281个专家标注问题，其中1147题为测试集。FinQANet（RoBERTa-large）检索文字与表格并生成可执行计算程序，测试执行准确率65.05%。","aiRole":"检索财报数字并生成可执行计算程序；指定1147题测试集执行准确率65.05%。","config":"固定首稿的1147题测试集、规则表格布局；BERT-base检索器取排名前3条事实，RoBERTa-large程序生成编码器。","metric":"1147题测试集执行答案准确率；不是8281题全量测量或通用财报分析通过率","meaning":"取数、公式和计算可分别核对；这条路径早于通用生成式Agent。","limit":"不是65%的财务工作自动化；专家参照与模型不是同预算增长序列。","sourceIds":["BF01"],"comparison":"任务/配置或指标不同，不计算统一增长率","supersedes":null,"actualCoverage":null,"measurementWindow":null},{"id":"F01-tools-2023","taskId":"F01","availableAt":"2023-10-23","eventDate":null,"dateBasis":"按所引用固定版本的公开日进入；不是首次发现或运行日","title":"把财报问答的计算交给Python","kind":"benchmark","mode":"assist","result":"FinQA测试1,147题；表2同一Codex底层模型的少样本贪心结果为CoT 40.4%、PoT 64.5%。","aiRole":"模型表达变量/计算程序，Python执行；人仍需核对真实财报的取数与口径。","config":"code-davinci-002；Python3.8/SymPy；按数据集使用4–8示例；比较不同提示/执行系统","metric":"财报问答答案指标；CoT使用相对容差0.001，非完整报告验收","meaning":"可计算步骤有外部执行接口，但正确运算仍可能基于错误取数。","limit":"历史模型和限定问答；不是现场报表流程、审查人时或岗位覆盖。","sourceIds":["BF06"],"comparison":"只解释这一来源/任务/配置；不与其他题集连成增长曲线","supersedes":null,"actualCoverage":null,"measurementWindow":null},{"id":"F01-2023","taskId":"F01","availableAt":"2023-11-20","eventDate":null,"dateBasis":"按该条支持材料的公开日期进入时间轴；不是实验统一发生日","title":"检索方式决定财报问答可靠性","kind":"benchmark","mode":"bounded","result":"FinanceBench公开150题中，GPT-4-Turbo共享向量库29/150答对；整份财报长上下文118/150答对。","aiRole":"模型从财报回答问题；同150题的正确数因资料流程不同为29或118。","config":"同一论文不同配置；长文档超限时截断至约95k–100k tokens","metric":"人工核对的正确回答数；包含逻辑与计算依据检查","meaning":"这组实验中，同一模型因资料处理方式不同表现悬殊；实际使用时仍要核对材料、页码与口径。","limit":"这是同论文、同实验窗口中的配置比较，不是两次时间节点的增长。","sourceIds":["BF02"],"comparison":"同一150题的配置差异可比较；不能当历史增长","supersedes":null,"actualCoverage":null,"measurementWindow":{"start":"2023-11","end":"2023-11","precision":"month","label":"2023年11月测试；Appendix B仅给月份，具体日期未知；不是整月连续测量"}}],"sources":[{"id":"BF01","title":"FinQA固定首稿","url":"https://arxiv.org/html/2109.00122v1","publishedAt":"2021-09-01","kind":"研究基准","locator":"§3任务；§4.4数据切分；§6设置（BERT-base/top3）与§6.1表2（RoBERTa-large）。","limit":"65.05%为FinQANet执行准确率；91.16%专家参照不作同样本同预算受控增长比较","family":"BF01","readAt":"2026-10-14","readingScope":"本轮直接阅读所列原文位置；未复跑实验","dateBasis":"页面标示发布日期或固定预印本v1页首日期","status":"active"},{"id":"BF02","title":"FinanceBench固定首稿","url":"https://arxiv.org/html/2311.11944v1","publishedAt":"2023-11-20","kind":"研究基准","locator":"§4配置及§4.1标注；§5表2（Shared Vector Store与Long Context）。 Appendix B（测试月份）。","limit":"150题公开子集；29/150与118/150是同论文实验窗口中的不同流程，不能画成时间进步","family":"financebench","readAt":"2026-10-15","readingScope":"2026-10-15直接复核§4/§5及Appendix B；未复跑实验","dateBasis":"页面标示发布日期或固定预印本v1页首日期","status":"active"},{"id":"BS01","title":"金融NLP系统综述","url":"https://arxiv.org/html/2504.07274v1","publishedAt":"2025-04-09","kind":"综述","locator":"§2筛选；§3任务；§7.1指标与§7.2危机期间","limit":"374候选/221金融任务论文；论文数不是能力分/真实交易净收益","family":"BS01","readAt":"2026-10-14","readingScope":"本轮直接阅读所列原文位置；未复跑实验","dateBasis":"页面标示发布日期或固定预印本v1页首日期","status":"active"},{"id":"BF06","title":"Program of Thoughts：外部计算接口（v4）","url":"https://arxiv.org/html/2211.12588v4","publishedAt":"2023-10-23","kind":"数值任务研究预印本","locator":"§2.2；§3.1数据/实现/指标；§3.2表2","limit":"同底层模型改变提示与执行流程；未匹配完整延迟/预算；不是端到端财务工作通过率","family":"program-of-thoughts","readAt":"2026-10-15","readingScope":"本轮直接核对所列原文位置/版本；未复跑外部实验","dateBasis":"固定v4页首：2023-10-23；不把v4表格倒填v1","status":"active"},{"id":"BF07","title":"GDPval v2公开材料：虚构巡演损益表","url":"https://huggingface.co/datasets/openai/gdpval/blob/11e7900cdcac61bc4daf59e65feb238acda98fbf/README.md","publishedAt":"2026-02-10","kind":"固定数据集工作材料","locator":"task_id 7b08cd4d-df60-41ae-9102-8aaa49306ba2；参考工作簿与rubric","limit":"会计产物与F01相邻，非真实上市公司财报；许可未明确，不复制材料或rubric；本站未执行","family":"gdpval-v2-material","readAt":"2026-10-16","readingScope":"本轮直接阅读所列原文和固定版本；未复跑外部实验","dateBasis":"2026-02-10是固定revision发布日期，单项任务首次公开时间未知","status":"active"},{"id":"R-JOBBENCH","title":"JobBench：专业证据推理与委派偏好","url":"https://arxiv.org/html/2605.26329v1","publishedAt":"2026-05-25","kind":"论文","locator":"§2.1–2.3；附录E；§3.1","limit":"偏好用于选题；基准没有测量员工采用、满意度或净省时。","family":"JobBench","readAt":"2026-10-19","readingScope":"本轮实际阅读指定方法段落，未复跑","dateBasis":"arXiv固定v1公开日期","status":"active"}],"scope":"全模块16个代表任务切片；当前聚焦1个F01；不是完整职业清单或任务覆盖比例。","actualCoverage":null,"knowledgeRule":"主轴按原始材料公开日期；实验/上线时点另列。历史空白仅表示本次未收录证据；后来的质疑不偷偷改写早期结果。","method":{"coverage":"覆盖的含义：在指定输入、环境和验收下，材料支持AI参与哪些动作；未知比例保持null。","speed":"速度的含义：证据节点之间的日历间隔、流程交接变化；指标/样本/预算不一致时，不计算能力增长率或外推一年三年。","sourcePolicy":"综述用于发现任务和原始文献；每条公开主张引用原始论文/官方记录。转载及同一文献家族不累计成独立验证。","selection":"目的性选择三个领域的代表任务，不是穷尽检索或无偏职业采样；基础时间轴于2026-10-14核对，本次精选补充与证据截止2026-10-15，固定论文优先，不把动态榜排名倒填历史。"}}