{"schema_version":3,"paper":{"title":"JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs","kind":"benchmark","primary_source_ids":["paper","repo"],"earliest_arxiv_date":"2026-10-15","earliest_arxiv_timestamp_utc":"2026-10-15T07:27:08Z"},"artifacts":{"article":"content/_posts/091-jev-agent-harness.md","deck":"lib/decks/jev-agent-harness.html","embed_src":"/lib/decks/jev-agent-harness.html"},"sources":[{"id":"paper","kind":"paper","title":"JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs","url":"https://arxiv.org/pdf/2610.21412v1","revision":"2610.21412v1","used_for":["complete paper, appendix, cases, metrics, negative experiments"]},{"id":"arxiv-entry","kind":"metadata","title":"v1 date and reuse license","url":"https://arxiv.org/abs/2610.21412v1","revision":"v1 published 2026-10-15","used_for":["title","date","CC BY-NC-SA4.0"]},{"id":"repo","kind":"code","title":"JevAgentHarness official code and follow-up experiments","url":"https://github.com/JevAgentHarness/JevAgentHarness/tree/f2932199e1ef1d9f089b51d39430c87be0f560b5","revision":"f2932199e1ef1d9f089b51d39430c87be0f560b5","used_for":["checker","data","threshold","OPD","GRPO","bench/PROTOCOL.md and README follow-up routing"]}],"paper_reading":{"main_agent_complete_read":true,"appendices_included":true,"extracted_text_reviewed":true,"rendered_pages_reviewed":true,"page_count":44,"rendered_page_ranges":[{"start":1,"end":44}],"source_visuals_expected":54,"visual_inventory_complete":true},"note_sections":[{"role":"problem","locator":"1. Motivation：答对下一步，不等于做成整件事"},{"role":"related-work","locator":"2. Related Work：重点是执行中的选项决策，不是完整代码蒸馏"},{"role":"method","locator":"3. Method / Protocol：选项、阈值、执行与答案键"},{"role":"experiments","locator":"4. Experiments：有效的是特定状态上的监督，不是普遍可靠的接管"},{"role":"future-work","locator":"5. Discussion：做 agentic OPD，最值得借鉴的是测试设计"},{"role":"summary","locator":"6. Conclusion：一篇值得读的决策系统论文，不是 OPD 成功范例"}],"terminology":[{"term":"harness","kind":"actor","article_quote":"组织模型调用、工具、历史记录和执行流程的外层软件，本文沿用 harness 。","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"§1–2"}]},{"term":"decision model","kind":"actor","article_quote":"作者将读取当前文字状态、对有限选项打分的小模型称为 decision model 。","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Section §2"}]},{"term":"defer","kind":"operation","article_quote":"本文用“交回大模型”解释 defer ：小模型最优选项的概率低于阈值，harness 不执行该选项，改请大模型决定。","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Section §2"}]},{"term":"Cascade","kind":"operation","article_quote":"论文主要使用 Cascade 模式，即先问小模型，再按阈值决定是否交回；自由文本参数始终需要大模型生成。","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Section §2"}]},{"term":"answer key","kind":"prompt-context","article_quote":"生成程序知道标准操作计划与真实世界状态，由此生成每一步可接受选项，作者称为 answer key （答案键）。","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Appendix B.3–B.4"}]},{"term":"non-inferiority","kind":"metric","article_quote":"作者将成功率下降不超过 3 个百分点作为 DecisionGym 等任务的 non-inferiority （非劣效）界限；τ²-bench 用 5 点。","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Appendix E"}]}],"coverage":[{"id":"central-claim","claim_type":"author-claim","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"§1,7"}],"evidence":[{"artifact":"article","quote":"决策模型需要在真实执行过程里测，不能只用独立选择题排名。"}]},{"id":"task-provenance","claim_type":"author-claim","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"B.2, D.10 / jevgym.py"}],"evidence":[{"artifact":"article","quote":"作者编写生成器创建客服世界：12 个客户，每人 2–4 个订单，10 类库存"}]},{"id":"annotation-and-check-construction","claim_type":"inspected-implementation","source_ids":["paper","repo"],"source_locators":[{"source_id":"paper","locator":"B.3–B.4 / jevgym.py"},{"source_id":"repo","locator":"B.3–B.4 / jevgym.py"}],"evidence":[{"artifact":"article","quote":"没有额外的人工 check.json 标注过程。"}]},{"id":"public-vs-hidden-inputs","claim_type":"inspected-implementation","source_ids":["paper","repo"],"source_locators":[{"source_id":"paper","locator":"B.3–B.4 / oracle_acceptable_tool_sets"},{"source_id":"repo","locator":"B.3–B.4 / oracle_acceptable_tool_sets"}],"evidence":[{"artifact":"article","quote":"答案键在模型答案记录后、复制的环境里计算，不进入被测提示"}]},{"id":"evaluation-formulas","claim_type":"author-claim","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"§3 / Appendix E"}],"evidence":[{"artifact":"article","quote":"先算每题状态准确率，再对题平均"}]},{"id":"checker-code-and-cases","claim_type":"inspected-implementation","source_ids":["paper","repo"],"source_locators":[{"source_id":"paper","locator":"Table6 / jevgym.py _write_is_gold,_step,evaluate"},{"source_id":"repo","locator":"Table6 / jevgym.py _write_is_gold,_step,evaluate"}],"evidence":[{"artifact":"article","quote":"即使猜中了 ORD-4002，只要没先问清就取消，也判失败。"},{"artifact":"article","quote":"实际代码也检查这些行为。判分规则推得的失败例不冒充实测日志。"}]},{"id":"model-results","claim_type":"author-claim","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Tables2,27–29"}],"evidence":[{"artifact":"article","quote":"共享 GPU 上却从 6.83 s 慢到 10.21 s。"}]},{"id":"related-work-comparison","claim_type":"author-claim","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"§8 / Appendix H"}],"evidence":[{"artifact":"article","quote":"论文没有加入训练后的 RouteLLM 作为完整路由基线"}]},{"id":"limitations","claim_type":"author-claim","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Table19 / Appendix D.6, E"}],"evidence":[{"artifact":"article","quote":"新规则会导致高置信度失败；影子校准无法覆盖接管后的全部状态"}]},{"id":"opd-negative","claim_type":"author-claim","source_ids":["paper"],"source_locators":[{"source_id":"paper","locator":"Table16 / D.2–D.4"}],"evidence":[{"artifact":"article","quote":"Qwen 路线的开发准确率在 SFT 后为 99.7%，OPD 后为 98.9%"}]}],"source_visuals":[{"id":"figure-1","kind":"figure","label":"Figure 1","source_id":"paper","source_locator":"p.1, Figure 1","status":"omitted","rationale":"开篇直觉由 Figure3 的完整控制权图替代，避免重复画相同大/小模型分工。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"figure-2","kind":"figure","label":"Figure 2","source_id":"paper","source_locator":"p.2, Figure 2","status":"reused","rationale":"Figure 2保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 2 · 工作总览。Stage 1 分别测试每种接管方式，Stage 2 冻结筛选出的组合并测新题。图画出了 SFT 与 GRPO，但不代表两阶段都贡献了收益。Zhu et al., v1"},{"artifact":"deck","locator":"分别测试，再冻结组合到新任务"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 2, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-2.png"]},{"id":"figure-3","kind":"figure","label":"Figure 3","source_id":"paper","source_locator":"p.3, Figure 3","status":"reused","rationale":"Figure 3保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 3 · 左边是接管位置，右边是控制权分配。Zhu et al., v1, p.3，CC BY-NC-SA 4.0。"},{"artifact":"deck","locator":"小模型能接管哪些决定，什么时候该交回？"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 3, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-3.png"]},{"id":"figure-4","kind":"figure","label":"Figure 4","source_id":"paper","source_locator":"p.6, Figure 4","status":"reused","rationale":"Figure 4保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 4 · 上图的 L1 成本—准确率与下图 BFCL 工具题列不是同一指标。下图 L3 在 shifted tasks 上让模型选所有工具，不使用置信阈值。Zhu et al., v1, "}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 4, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-4.png"]},{"id":"figure-5","kind":"figure","label":"Figure 5","source_id":"paper","source_locator":"p.7, Figure 5","status":"reused","rationale":"Figure 5保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 5 · v1 p.7，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 5, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-5.png"]},{"id":"figure-6","kind":"figure","label":"Figure 6","source_id":"paper","source_locator":"p.17, Figure 6","status":"omitted","rationale":"早期单轨迹由附录Figures11–12的完整实测替代，后者保留调用、参数、成本和成败。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"figure-7","kind":"figure","label":"Figure 7","source_id":"paper","source_locator":"p.19, Figure 7","status":"reused","rationale":"Figure 7保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 7 · 有限选项的问题才可直接打分，自由文本另交给大模型。Zhu et al., v1, p.19，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 7, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-7.png"]},{"id":"figure-8","kind":"figure","label":"Figure 8","source_id":"paper","source_locator":"p.25, Figure 8","status":"omitted","rationale":"工程模块结构不另展示；正文沿真实请求与官方代码说明接口，不把模块图当执行日志。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"figure-9","kind":"figure","label":"Figure 9","source_id":"paper","source_locator":"p.27, Figure 9","status":"omitted","rationale":"训练链已由Table16和三阶段正文重构；避免链图让读者误以为OPD保留或GRPO有收益。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"figure-10","kind":"figure","label":"Figure 10","source_id":"paper","source_locator":"p.34, Figure 10","status":"omitted","rationale":"完整八乘五模式结果矩阵不全部展示；正文解释五模式，保留有效Cascade与失败组合对照。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"figure-11","kind":"figure","label":"Figure 11","source_id":"paper","source_locator":"p.43, Figure 11","status":"reused","rationale":"Figure 11保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 11 · 实测日志。蓝色是 AgentDecision-4B 的选择，橙色是 Gemini 3.5 Flash 的文本或参数输出。作者按 seed 顺序选出“两边均成功且至少一次交回”的"},{"artifact":"deck","locator":"五次工具选择接管，仍有两次大模型调用"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 11, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-11.png"]},{"id":"figure-12","kind":"figure","label":"Figure 12","source_id":"paper","source_locator":"p.43, Figure 12","status":"reused","rationale":"Figure 12保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 12 · 两边均选所有工具、不使用阈值，大模型填参数。Jev 的首个不可撤销错误发生在必须追问的状态；右边 4B 先问。作者按 seed 选择的诊断案例。Zhu et al., v1, "},{"artifact":"deck","locator":"0.98 直接退款，仍是不可撤销错误"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 12, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-12.png"]},{"id":"figure-13","kind":"figure","label":"Figure 13","source_id":"paper","source_locator":"p.44, Figure 13","status":"reused","rationale":"Figure 13保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Figure 13 · 另做 ALFWorld SFT 的模型，不是客服模型直接转用。Zhu et al., v1, p.44，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 13, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/figure-13.png"]},{"id":"table-1","kind":"table","label":"Table 1","source_id":"paper","source_locator":"p.7, Table 1","status":"reused","rationale":"Table 1保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 1 · 主文的扩展任务结果，不同列分别是成功率、节省成本和时间。Zhu et al., v1, p.7，CC BY-NC-SA 4.0。"},{"artifact":"deck","locator":"省钱、提速和可靠性要分别核查"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 1, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-1.png"]},{"id":"table-2","kind":"table","label":"Table 2","source_id":"paper","source_locator":"p.7, Table 2","status":"reused","rationale":"Table 2保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 2 · v1 p.7，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 2, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-2.png"]},{"id":"table-3","kind":"table","label":"Table 3","source_id":"paper","source_locator":"p.8, Table 3","status":"omitted","rationale":"精简主表数值与Appendix Table39重叠，后者提供全部变化任务模型与执行条件。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-4","kind":"table","label":"Table 4","source_id":"paper","source_locator":"p.16, Table 4","status":"omitted","rationale":"40个实现组合由正文八位置五模式概括；不把实现覆盖误当全部完成效果验证。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-5","kind":"table","label":"Table 5","source_id":"paper","source_locator":"p.17, Table 5","status":"omitted","rationale":"所有套件size×attempts目录不逐项复刻；正文分别列所用L1/L2/L3、ALF及coding分母。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-6","kind":"table","label":"Table 6","source_id":"paper","source_locator":"p.19, Table 6","status":"reused","rationale":"Table 6保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 6 · 两道生成器开发题及答案键。它说明任务与标准步骤，不是模型执行日志。下面沿第一题解释判分。Zhu et al., v1, p.19，CC BY-NC-SA 4.0。"},{"artifact":"deck","locator":"生成器写世界和答案键，不是人工逐题标注"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 6, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-6.png"]},{"id":"table-7","kind":"table","label":"Table 7","source_id":"paper","source_locator":"p.25, Table 7","status":"omitted","rationale":"Guard细分误报表不展开；正文保留Guard不构成安全边界，主要案例聚焦实际不可撤销写错误。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-8","kind":"table","label":"Table 8","source_id":"paper","source_locator":"p.25, Table 8","status":"omitted","rationale":"state type分项由核心ask案例与Table39代表；未展示全部guard/stop状态矩阵。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-9","kind":"table","label":"Table 9","source_id":"paper","source_locator":"p.25, Table 9","status":"reused","rationale":"Table 9保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 9 · v1 p.25，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 9, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-9.png"]},{"id":"table-10","kind":"table","label":"Table 10","source_id":"paper","source_locator":"p.26, Table 10","status":"omitted","rationale":"单个最终200题组合在正文解释，主图选Table1与成本反例，避免重复成功点估计。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-11","kind":"table","label":"Table 11","source_id":"paper","source_locator":"p.26, Table 11","status":"omitted","rationale":"miniSWE两组完整routing分项不放图片；正文保留300题77.3→76.3与100题70→75边界。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-12","kind":"table","label":"Table 12","source_id":"paper","source_locator":"p.26, Table 12","status":"omitted","rationale":"SWEagent和pi路由完整表不另展示；Table1与正文保留任务数、退步和pi变慢。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-13","kind":"table","label":"Table 13","source_id":"paper","source_locator":"p.28, Table 13","status":"omitted","rationale":"角色状态清单不全部展示；正文明确coding hooks仅实现、路由实测、规划及完整代码蒸馏未覆盖。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-14","kind":"table","label":"Table 14","source_id":"paper","source_locator":"p.28, Table 14","status":"omitted","rationale":"小样本开发诊断不当主证据；正文利用Table16正式训练对照与实测退款失败说明状态差异。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-15","kind":"table","label":"Table 15","source_id":"paper","source_locator":"p.29, Table 15","status":"omitted","rationale":"获得模型的成本汇总不重复展示；正文保留SFT、OPD、GRPO与公共数据的成本和执行反转。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-16","kind":"table","label":"Table 16","source_id":"paper","source_locator":"p.30, Table 16","status":"reused","rationale":"Table 16保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 16 · 每个阶段的信号、成本与结果。recov. / off-plan 指之前已偏离标准计划的状态；Dev 的两条路线不是同一题集，不能横向对分数。Zhu et al., v1, p.3"},{"artifact":"deck","locator":"SFT 有效，OPD 没被保留"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 16, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-16.png"]},{"id":"table-17","kind":"table","label":"Table 17","source_id":"paper","source_locator":"p.30, Table 17","status":"reused","rationale":"Table 17保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 17 · v1 p.30，CC BY-NC-SA 4.0；每个配置一个 seed。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 17, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-17.png"]},{"id":"table-18","kind":"table","label":"Table 18","source_id":"paper","source_locator":"p.30, Table 18","status":"omitted","rationale":"0.8B评分头完整消融不展开；这不是正文主线，源码阅读路径仍给出评分实现。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-19","kind":"table","label":"Table 19","source_id":"paper","source_locator":"p.30, Table 19","status":"reused","rationale":"Table 19保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 19 · v1 p.30，CC BY-NC-SA 4.0；各段题数与设置不同，只在段内比较。"},{"artifact":"deck","locator":"新规则下，置信度没有及时交回控制权"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 19, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-19.png"]},{"id":"table-20","kind":"table","label":"Table 20","source_id":"paper","source_locator":"p.31, Table 20","status":"omitted","rationale":"跨ALF和tau状态准确率全矩阵不展示；正文限定ALF另训，而非客服模型直接迁移。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-21","kind":"table","label":"Table 21","source_id":"paper","source_locator":"p.31, Table 21","status":"omitted","rationale":"tau单独训练失败完整表未展示；正文保留tau的五点界限，未将其说成通用成功。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-22","kind":"table","label":"Table 22","source_id":"paper","source_locator":"p.32, Table 22","status":"omitted","rationale":"模板基线只保留97.5%与近满分答案键对照，凸显生成器套路，不复刻全部特征行。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-23","kind":"table","label":"Table 23","source_id":"paper","source_locator":"p.34, Table 23","status":"omitted","rationale":"外部工具许可目录不复刻；正文将本论文CC BY-NC-SA与本仓库MIT明确分开。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-24","kind":"table","label":"Table 24","source_id":"paper","source_locator":"p.35, Table 24","status":"omitted","rationale":"注册检验完整清单不复制；主要组合的检验失败和事先披露的探索边界在正文可见。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-25","kind":"table","label":"Table 25","source_id":"paper","source_locator":"p.35, Table 25","status":"omitted","rationale":"决策模型各自注册主对照不全展示；正文集中同任务成本与成功是否同时保留。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-26","kind":"table","label":"Table 26","source_id":"paper","source_locator":"p.36, Table 26","status":"omitted","rationale":"L1全部准确率校准延迟表不复制；正文明确3991决定与2091参数槽位分母不同。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-27","kind":"table","label":"Table 27","source_id":"paper","source_locator":"p.36, Table 27","status":"omitted","rationale":"Stage2每个零假设和p值不重复展示；正文明确97.2%组合非劣效失败和4B共享GPU变慢。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-28","kind":"table","label":"Table 28","source_id":"paper","source_locator":"p.37, Table 28","status":"omitted","rationale":"次要注册/描述检验表未展示；避免把描述趋势写成统计保证，正文给出检验界限。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-29","kind":"table","label":"Table 29","source_id":"paper","source_locator":"p.37, Table 29","status":"reused","rationale":"Table 29保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 29 · v1 p.37，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 29, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-29.png"]},{"id":"table-30","kind":"table","label":"Table 30","source_id":"paper","source_locator":"p.37, Table 30","status":"omitted","rationale":"跨复杂任务配对wins/losses表不另展示；正文保留ALF未见房间p=.175与重训不确定。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-31","kind":"table","label":"Table 31","source_id":"paper","source_locator":"p.38, Table 31","status":"omitted","rationale":"补充三层表与Figure4重叠，正文保留图并逐列解释其条件。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-32","kind":"table","label":"Table 32","source_id":"paper","source_locator":"p.38, Table 32","status":"omitted","rationale":"更多候选模型排名不全复制；Figure4和Table39已表达独立题/任务排名变化。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-33","kind":"table","label":"Table 33","source_id":"paper","source_locator":"p.39, Table 33","status":"omitted","rationale":"vLLM部署全部价格与GPU参数不展示；正文区别训练GPU-hours与推理价格，保留独占/共享GPU结果。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-34","kind":"table","label":"Table 34","source_id":"paper","source_locator":"p.39, Table 34","status":"omitted","rationale":"所有不可撤销错误类型频率不展示；Figure12完整退款错误与Table19批量失败为明确证据。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-35","kind":"table","label":"Table 35","source_id":"paper","source_locator":"p.39, Table 35","status":"omitted","rationale":"重复投票和精排消融不作为正文方法；不宣称二次询问普遍有效或替代训练。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-36","kind":"table","label":"Table 36","source_id":"paper","source_locator":"p.39, Table 36","status":"omitted","rationale":"另两个训练后开放模型全结果未展示；主线明确Qwen与StartLux是不同训练路线。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-37","kind":"table","label":"Table 37","source_id":"paper","source_locator":"p.40, Table 37","status":"omitted","rationale":"同骨干评分头与token读出对照不全展示；正文提供一次前向选项读出和52选项上限。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-38","kind":"table","label":"Table 38","source_id":"paper","source_locator":"p.40, Table 38","status":"omitted","rationale":"学习defer二级头不作为主方法；正文明确主要阈值及其变化任务失败，未宣称校准已解决。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]},{"id":"table-39","kind":"table","label":"Table 39","source_id":"paper","source_locator":"p.40, Table 39","status":"reused","rationale":"Table 39保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 39 · v1 p.40，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 39, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-39.png"]},{"id":"table-40","kind":"table","label":"Table 40","source_id":"paper","source_locator":"p.41, Table 40","status":"reused","rationale":"Table 40保留原始英文、条件与数值，置于对应机制、真实案例或实验说明旁；图片裁去外部正文，不重画数据。","placements":[{"artifact":"article","locator":"Table 40 · v1 p.41，CC BY-NC-SA 4.0。"}],"permission_status":"verified-allowed","permission_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 40, CC BY-NC-SA4.0","asset_paths":["/lib/papers/jev-agent-harness/table-40.png"]},{"id":"table-41","kind":"table","label":"Table 41","source_id":"paper","source_locator":"p.42, Table 41","status":"omitted","rationale":"更难coding探索表未展示；正文主coding结果及后来独立代理测试已分开，未推广为所有任务效果。","placements":[],"permission_status":"not-applicable","permission_basis":"","credit":"","asset_paths":[]}],"figures":[{"path":"/lib/papers/jev-agent-harness/figure-2.png","origin":"paper-original","source_visual_id":"figure-2","source_id":"paper","source_locator":"p.2, Figure 2","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 2, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/figure-3.png","origin":"paper-original","source_visual_id":"figure-3","source_id":"paper","source_locator":"p.3, Figure 3","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 3, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/figure-4.png","origin":"paper-original","source_visual_id":"figure-4","source_id":"paper","source_locator":"p.6, Figure 4","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 4, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/figure-5.png","origin":"paper-original","source_visual_id":"figure-5","source_id":"paper","source_locator":"p.7, Figure 5","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 5, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/figure-7.png","origin":"paper-original","source_visual_id":"figure-7","source_id":"paper","source_locator":"p.19, Figure 7","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 7, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/figure-11.png","origin":"paper-original","source_visual_id":"figure-11","source_id":"paper","source_locator":"p.43, Figure 11","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 11, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/figure-12.png","origin":"paper-original","source_visual_id":"figure-12","source_id":"paper","source_locator":"p.43, Figure 12","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 12, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/figure-13.png","origin":"paper-original","source_visual_id":"figure-13","source_id":"paper","source_locator":"p.44, Figure 13","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Figure 13, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/table-1.png","origin":"paper-original","source_visual_id":"table-1","source_id":"paper","source_locator":"p.7, Table 1","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 1, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/table-2.png","origin":"paper-original","source_visual_id":"table-2","source_id":"paper","source_locator":"p.7, Table 2","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 2, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/table-6.png","origin":"paper-original","source_visual_id":"table-6","source_id":"paper","source_locator":"p.19, Table 6","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 6, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/table-9.png","origin":"paper-original","source_visual_id":"table-9","source_id":"paper","source_locator":"p.25, Table 9","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 9, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/table-16.png","origin":"paper-original","source_visual_id":"table-16","source_id":"paper","source_locator":"p.30, Table 16","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 16, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/table-17.png","origin":"paper-original","source_visual_id":"table-17","source_id":"paper","source_locator":"p.30, Table 17","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 17, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/table-19.png","origin":"paper-original","source_visual_id":"table-19","source_id":"paper","source_locator":"p.30, Table 19","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 19, CC BY-NC-SA4.0","used_in":["article","deck"]},{"path":"/lib/papers/jev-agent-harness/table-29.png","origin":"paper-original","source_visual_id":"table-29","source_id":"paper","source_locator":"p.37, Table 29","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 29, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/table-39.png","origin":"paper-original","source_visual_id":"table-39","source_id":"paper","source_locator":"p.40, Table 39","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 39, CC BY-NC-SA4.0","used_in":["article"]},{"path":"/lib/papers/jev-agent-harness/table-40.png","origin":"paper-original","source_visual_id":"table-40","source_id":"paper","source_locator":"p.41, Table 40","license_status":"verified-allowed","license_basis":"arXiv v1 entry CC BY-NC-SA4.0; noncommercial blog; attribution and same-license crops","credit":"Zhu et al., JevAgentHarness v1, Table 40, CC BY-NC-SA4.0","used_in":["article"]}],"qa":{"reading_layout":{"deck_first":true,"toc_h2_only":true,"surface":"neutral-light","body_font":"serif","body_font_px":16,"line_height":1.85,"figures_centered":true,"article_visuals_interleaved":true,"internal_scroll_for_tables_and_code":true,"page_horizontal_overflow":false},"viewports":[{"width":1600,"height":1000,"passed":true},{"width":1512,"height":982,"passed":true},{"width":390,"height":844,"passed":true}],"tested_surfaces":{"article_body":true,"embedded_deck":true,"standalone_deck":true},"language_review":{"technical_terms_contextualized":true,"actors_disambiguated":true,"read_aloud_complete":true,"facts_rechecked_after_rewrite":true},"interactions":{"buttons":true,"keyboard":true,"hash_navigation":true,"fullscreen":true,"escape":true,"standalone_return_to_article":true,"immersive_mode_exit":true,"usable_state_after_exit":true},"inspected_slides":["opening","densest","largest-table-or-code","final"],"current_review":{"date":"2026-10-18","browser":"isolated HeadlessChrome 154","native_window_tested":false,"scope":"Complete article and embedded/standalone deck checked.","reports":["tmp/jev-layout-qa.js","tmp/jev-interactions.cjs"],"external_requests":"No page errors in final layout checks. One local twikoo script request aborted during navigation in Jev interaction test; external live comments verification remains a separate release check.","source_sha256":{"article":"ae891667297db763274d16a3f73f1baa727eaaaffda86701949b38682296cdee","deck":"e8cb533a652970d8a8934fbdf6088f7d8a6894273bb66cc549698c9facbaafa0"}}}}