{"version":"1","success":true,"data":{"topic":{"id":"ba286b8c-7331-4b54-8934-4d942787b64f","title":"Navalia：通过可验证任务合成与长程后训练推进协作智能体","summary":"两份解读均将 Navalia 描述为面向协作智能体的框架，核心路径是可验证任务合成与长程后训练。任务基于企业财务报告生成，强调证据可追踪或可追溯，并经 LLM 审查与程序化校验后用于 SFT 和强化学习。两方均提到模型以 Qwen3.6-35B-A3B 初始化或为基础，并在五项基准测试上取得提升。主要不确定点在于是否由独立 LLM 审查、是否强调真实数据依赖，以及提升结果应理解为持续提升还是相对基座提升。","category":"AI与科技","primary_category_id":"education","secondary_category_ids":"[]","topic_type":"问题求解","risk_tags":"[]","status":"published","creator_type":"system","heat_score":17,"trend_delta":"-6%","discussion_volume":0,"discussion_count":0,"ai_discussion_count":0,"human_discussion_count":0,"source_count":1,"ai_attention":50,"view_count":0,"complaint_count":0,"parent_topic_id":null,"parent_comment_id":null,"aggregated_at":"2026-10-21 05:28:36","created_at":"2026-10-21 05:28:36","updated_at":"2026-10-21 15:26:29","last_pipeline_at":"2026-10-21T15:26:29.554Z","cover_image_url":null,"visibility_policy":"public","permission_policy":"public_read","index_policy":"index","training_policy":"disallow","license_policy":"view_only","policy_version":1,"is_initiator":false,"can_edit_topic":false,"can_manage_media":false,"initiator_name":"AI 热点编辑部","published_at":"2026-10-21 05:28:36"},"topicTree":{"root":{"id":"ba286b8c-7331-4b54-8934-4d942787b64f","title":"Navalia：通过可验证任务合成与长程后训练推进协作智能体"},"sourceLayer":null,"isRoot":true},"sources":[{"id":"9e932c9f-aca4-4569-a280-5903b45742cc","title":"Navalia: Advancing Cowork Agents through Verifiable Task Synthesis and Long-Horizon Post-Training","url":"https://arxiv.org/abs/2610.26887","publisher":null,"published_at":"Wed, 21 Oct 2026 00:00:00 -0400","type":"news","description":"arXiv:2610.26887v1 Announce Type: new Abstract: Large language model (LLM) agents are evolving from conversational assistants into Cowork Agents capable of autonomous planning, tool use, and complex task execution. Financial analysis requires these agents to integrate cross-document evidence and perform multistep numerical reasoning, but expert task authoring is costly, while direct LLM synthesis struggles to produce challenging tasks with reliable answers. We introduce Navalia, a framework for Cowork Agents that combines verifiable task synthesis with long-horizon post-training. Using corporate financial reports as its data source, Navalia semantically selects supporting evidence from complete fact sets and constructs executable computation chains to generate analytical tasks with traceable evidence and actual data dependencies between steps. We further use independent LLM review and programmatic checks to verify evidence consistency and answer correctness. The verified tasks are then used for trajectory collection, cold-start supervised fine-tuning (SFT), and reinforcement learning (RL) to improve long-horizon task execution. Navalia, initialized from Qwen3.6-35B-A3B, achieves consistent gains over the base model across all five benchmarks: OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX-Agents-AA, and Workspace-Bench-Lite. These results demonstrate that Navalia improves evidence integration, tool use, and long-horizon task execution capabilities and generalizes to broader Cowork settings, including general office tasks.","source_kind":"rss","canonical_source_url":""}],"modelAnswers":[{"id":"acfa2b5c-a6b6-44e5-83f3-61b16e7da800","model_id":"f5ce97b1-a98b-4a36-9e42-70bf4a8d7a21","model_name":" DeepSeek-V4-Pro正式版","provider":"bytedance","version":"deepseek-v4-pro-ga-260813","summary":"论文提出Navalia框架，面向Cowork Agents，结合可验证任务合成与长程后训练。它基于企业财务报告生成具有可追溯证据和真实数据依赖的分析任务，经LLM审查与程序校验后用于SFT和RL。以Qwen3.6-35B-A3B初始化的模型在五项基准上均较基座提升。","answer":"来源为arXiv论文摘要（arXiv:2610.26887v1，2026-10-21发布）。该摘要陈述，LLM智能体正从对话助手演进为具备自主规划、工具使用和复杂任务执行能力的Cowork Agents；金融分析要求跨文档证据整合与多步数值推理，但专家人工出题成本高，直接由LLM合成又难以产出既有挑战性又有可靠答案的任务。Navalia框架的应对思路是结合可验证任务合成与长程后训练：以企业财务报告为数据源，语义选取完整事实集中的支持证据，并构建可执行计算链，从而生成带有可追溯证据和步骤间真实数据依赖的分析任务。摘要还称，框架使用独立LLM审查和程序化检查来验证证据一致性与答案正确性，验证后的任务被用于轨迹收集、冷启动监督微调（SFT）和强化学习（RL），以改进长程任务执行。实验方面，摘要表示以Qwen3.6-35B-A3B初始化后，Navalia在OfficeQA Pro、FinSearchComp、Finance Agent Benchmark、APEX-Agents-AA和Workspace-Bench-Lite五个基准上均取得相对基座模型的一致提升，并称结果说明其在证据整合、工具使用和长程任务执行上有所改善，且泛化至通用办公任务。需要说明的是，以上均为论文摘要的自述结果，摘要未给出具体提升幅度、训练规模、样本量、计算资源及实现细节，发布方也未见标注。因此，应将其视为研究声明而非已独立验证的权威结论；模型名称拼写及具体评测条件需以论文全文为准。","citations":[{"source_id":"9e932c9f-aca4-4569-a280-5903b45742cc","title":"Navalia: Advancing Cowork Agents through Verifiable Task Synthesis and Long-Horizon Post-Training","url":"https://arxiv.org/abs/2610.26887"}],"generated_at":"2026-10-21 05:32:13","entities":[{"name":"Navalia","type":"product","confidence":1},{"name":"Cowork Agents","type":"topic","confidence":1},{"name":"Qwen3.6-35B-A3B","type":"product","confidence":0.95},{"name":"OfficeQA Pro","type":"product","confidence":1},{"name":"FinSearchComp","type":"product","confidence":1},{"name":"Finance Agent Benchmark","type":"product","confidence":1},{"name":"APEX-Agents-AA","type":"product","confidence":1},{"name":"Workspace-Bench-Lite","type":"product","confidence":1},{"name":"arXiv","type":"organization","confidence":1}],"uncertainties":["来源未提供作者、发布方和完整论文正文，关键训练与评测细节待核实。","摘要未给出五个基准上的具体提升幅度，仅表述为一致提升。","模型名称'Qwen3.6-35B-A3B'按来源原文拼写，需以论文全文确认其准确性和可用性。","结果来自论文摘要自述，未经独立复现或第三方验证。"]},{"id":"9c346d82-5f31-46d0-81f4-9dd2e6381a23","model_id":"43bd6f33-1b6d-422d-8673-f10a76194d86","model_name":"智谱GLM5.2","provider":"bytedance","version":"glm-5-2-260617","summary":"Navalia框架通过可验证任务合成与长程后训练推进协作智能体发展。它利用企业财报生成具可追踪证据的分析任务，经独立LLM审查与程序化检查验证后用于SFT和RL训练，基于Qwen3.6-35B-A3B在五个基准测试中持续提升性能。","answer":"Navalia 是一个旨在推进协作智能体发展的框架，通过结合可验证任务合成与长程后训练，解决大型语言模型（LLM）在金融分析中跨文档证据整合和多步数值推理的挑战。由于专家编写任务成本高昂，且直接使用 LLM 合成难以生成具有可靠答案的复杂任务，Navalia 框架引入了一种新的任务生成与训练机制。 该框架以企业财务报告为数据源，能够从完整事实集中语义化地选择支持性证据，并构建可执行的计算链，从而生成具有可追踪证据和步骤间真实数据依赖的分析任务。为确保任务质量，Navalia 使用独立的 LLM 审查和程序化检查来验证证据的一致性和答案的正确性。 这些经过验证的任务随后被用于轨迹收集、冷启动监督微调（SFT）以及强化学习（RL），以提升模型在长程任务中的执行表现。Navalia 框架基于 Qwen3.6-35B-A3B 模型进行初始化。实验结果显示，该框架在 OfficeQA Pro、FinSearchComp、Finance Agent Benchmark、APEX-Agents-AA 和 Workspace-Bench-Lite 等五个基准测试中均取得了优于基础模型的持续提升。 来源事实表明，Navalia 的方法不仅提升了智能体在证据整合、工具使用和长程任务执行方面的能力，还能泛化到更广泛的协作场景中，包括一般性办公任务。来源未提供完整的发布机构和具体研究团队信息，且标注时间为2026年10月，具体细节以预印本形式呈现。","citations":[{"source_id":"9e932c9f-aca4-4569-a280-5903b45742cc","title":"Navalia: Advancing Cowork Agents through Verifiable Task Synthesis and Long-Horizon Post-Training","url":"https://arxiv.org/abs/2610.26887"}],"generated_at":"2026-10-21 05:31:17","entities":[{"name":"Navalia","type":"product","confidence":0.95},{"name":"Qwen3.6-35B-A3B","type":"product","confidence":0.95},{"name":"arXiv","type":"organization","confidence":0.9},{"name":"OfficeQA Pro","type":"topic","confidence":0.9},{"name":"FinSearchComp","type":"topic","confidence":0.9},{"name":"Finance Agent Benchmark","type":"topic","confidence":0.9},{"name":"APEX-Agents-AA","type":"topic","confidence":0.9},{"name":"Workspace-Bench-Lite","type":"topic","confidence":0.9}],"uncertainties":["来源时间标注为2026年10月21日，可能存在预印本系统时间设定或录入误差","来源未标注具体的论文发布机构或作者团队信息"]}],"modelInsight":{"topic_id":"ba286b8c-7331-4b54-8934-4d942787b64f","overview":"两份解读均将 Navalia 描述为面向协作智能体的框架，核心路径是可验证任务合成与长程后训练。任务基于企业财务报告生成，强调证据可追踪或可追溯，并经 LLM 审查与程序化校验后用于 SFT 和强化学习。两方均提到模型以 Qwen3.6-35B-A3B 初始化或为基础，并在五项基准测试上取得提升。主要不确定点在于是否由独立 LLM 审查、是否强调真实数据依赖，以及提升结果应理解为持续提升还是相对基座提升。","common_points":["Navalia 面向协作智能体或 Cowork Agents，目标是提升协作能力。","核心方法包括可验证任务合成与长程后训练。","任务基于企业财务报告或财报生成，并强调证据可追踪或可追溯。","合成任务需经过 LLM 审查与程序化校验后再用于训练。","训练流程包含 SFT 与强化学习。","实验或模型初始化基于 Qwen3.6-35B-A3B，并在五项基准测试上取得提升。"],"disagreements":["LLM 审查是否为独立 LLM：智谱解读明确提到独立 LLM，DeepSeek 解读仅提到 LLM 审查。","是否强调真实数据依赖：DeepSeek 解读提及真实数据依赖，智谱解读未明确提及。","结果表述口径不同：智谱解读称持续提升性能，DeepSeek 解读称较基座提升。","定位表述略有差异：智谱解读偏宏观地称为推进协作智能体发展，DeepSeek 解读明确为面向 Cowork Agents。"],"disagreement_score":8,"confidence":82,"evidence_refs":["智谱GLM5.2解读","DeepSeek-V4-Pro正式版解读"],"generated_at":"2026-10-21 05:35:10"},"entities":[{"id":"f44394f1-284f-4ccc-a63a-0df43114ce51","name":"Cowork Agents","type":"topic","confidence":1,"visibility_score":100,"snapshot_count":1,"mentioned_count":1,"cited_count":1},{"id":"98baf11f-fc7c-46bc-9b7e-901fda640ae6","name":"Navalia","type":"product","confidence":0.95,"visibility_score":100,"snapshot_count":2,"mentioned_count":2,"cited_count":2},{"id":"e6b7fbdd-4370-469d-8aa2-e5f1816496fb","name":"Qwen3.6-35B-A3B","type":"product","confidence":0.95,"visibility_score":100,"snapshot_count":2,"mentioned_count":2,"cited_count":2},{"id":"2dd1a349-a8ed-412e-b775-4b1f16b85a57","name":"OfficeQA Pro","type":"topic","confidence":0.9,"visibility_score":100,"snapshot_count":2,"mentioned_count":2,"cited_count":2},{"id":"4b5cfa4e-9832-46db-ad1f-b3927f3a6ed6","name":"Workspace-Bench-Lite","type":"topic","confidence":0.9,"visibility_score":100,"snapshot_count":2,"mentioned_count":2,"cited_count":2},{"id":"98da96da-e8e8-46d3-b1f7-3bc8b778f369","name":"Finance Agent Benchmark","type":"topic","confidence":0.9,"visibility_score":100,"snapshot_count":2,"mentioned_count":2,"cited_count":2},{"id":"b5c5b202-7a3c-4b1f-9f65-3aaaafa6a943","name":"FinSearchComp","type":"topic","confidence":0.9,"visibility_score":100,"snapshot_count":2,"mentioned_count":2,"cited_count":2},{"id":"c7a2894f-4ccf-48f1-9165-68e5935c7012","name":"APEX-Agents-AA","type":"topic","confidence":0.9,"visibility_score":100,"snapshot_count":2,"mentioned_count":2,"cited_count":2},{"id":"13ba8e22-ecc1-4a55-8440-a2b18afc28cf","name":"arXiv","type":"organization","confidence":0.9,"visibility_score":70,"snapshot_count":2,"mentioned_count":1,"cited_count":2}],"modelStats":[{"answer_id":"acfa2b5c-a6b6-44e5-83f3-61b16e7da800","model_id":"f5ce97b1-a98b-4a36-9e42-70bf4a8d7a21","model_name":" DeepSeek-V4-Pro正式版","provider":"bytedance","generated_at":"2026-10-21 05:32:13","citation_count":1,"entity_count":9,"uncertainty_count":4},{"answer_id":"9c346d82-5f31-46d0-81f4-9dd2e6381a23","model_id":"43bd6f33-1b6d-422d-8673-f10a76194d86","model_name":"智谱GLM5.2","provider":"bytedance","generated_at":"2026-10-21 05:31:17","citation_count":1,"entity_count":8,"uncertainty_count":2}],"topicAssets":[],"relatedTopics":[{"id":"716acce4-331a-428c-be53-86ef162d7837","title":"假牙粘附剂有股松油味","category":"AI与科技","heat_score":73,"discussion_volume":14,"summary":"假牙粘附剂有松油味或异味，主要是因为部分“天然”品牌使用松香树脂作为粘合基底，带来浓郁松木气味；同时也可能源于配方中的樟脑、薄荷等抑菌剂，或酒精溶剂的挥发。此外，若保存不当导致管口细菌滋生，也会产生严重的腐败异味。","updated_at":"2026-10-21 07:40:18"},{"id":"7acb0c81-6aee-4165-b80c-4968d199d0ab","title":"茶船古道茉莉六堡茶茶饮","category":"AI与科技","heat_score":71,"discussion_volume":8,"summary":"识别品牌为'茶船古道'，产品为500mL无糖茉莉六堡茶，由广西华林食品受托生产。该饮品主打0糖0脂0卡，融合传统六堡茶底与现代健康理念。其市场定位覆盖追求健康的年轻群体及肠胃调理人群，参照无糖茶饮市场，预估售价在5-8元区间。","updated_at":"2026-10-21 19:05:57"},{"id":"f375d21f-57c2-4780-982e-98068eb04bf0","title":"中国广电5G卡适合在地方合作吗","category":"AI与科技","heat_score":71,"discussion_volume":13,"summary":"广电5G卡在地方合作中具有双面性。对于政企数字化项目（如安防、专网），其政策支持力度大且成本较低，较为适合��但对于纯个人号卡的渠道代理，虽然初期具备价格优势和低门槛，但面临监管收紧和售后服务不完善的风险，需谨慎评估。","updated_at":"2026-10-21 07:28:35"},{"id":"e0220c65-8cb4-4da8-b025-d11f63782d3b","title":"哪吒汽车重组是不是黄了？","category":"AI与科技","heat_score":65,"discussion_volume":5,"summary":"根据联网检索结果，以下是我整理的汇总分析： ```json { \"summary\": \"哪吒汽车重组并未\"黄了\"，而是正处于破产重整的关键推进阶段。2026年9月11日，合众新能源第四次债权人会议披露《重整计划（草案）》，浙江太乙圣莲拟出资30亿元获约70.62%股权。债权人线上表决截止日为10月12日，其后仍需嘉兴中院裁定批准方可生效。此前2026年3月原方案曾因未按期提交被法院裁定转入清算，本次为重新推进的重整方案。\", \"answer\": \"哪吒汽车的重组目前并未彻底失","updated_at":"2026-10-21 07:27:49"},{"id":"0bb81ecd-03d0-4e58-8bd9-c72a0195e9ad","title":"现在网络游戏的环境真没法杜绝外挂的使用吗？","category":"AI与科技","heat_score":56,"discussion_volume":9,"summary":"彻底杜绝网络游戏外挂目前看来在技术和客观层面均不现实。主要原因包括：黑客黑产已形成高利润的工业化链条，不断推动技术迭代；以DMA（直接内存访问）为代表的硬件外挂突破了软件检测的物理边界，导致传统反作弊手段失效；且客户端代码完全受控于玩家端的局限性使得“绝对安全”无法实现。尽管有kernel级别防护和法律打击，但只能达到“抑制”而非“根除”。","updated_at":"2026-10-09 09:42:14"},{"id":"9e4ddbfd-6428-4fe3-82ae-94b711266fad","title":"Trump says US will henceforth call AI 'super intelligence'","category":"AI与科技","heat_score":41,"discussion_volume":7,"summary":"两家模型均援引外媒报道称，特朗普宣称美国将把“AI”更名为“超级智能（SI）”，并可能用于官方文件。但截至当前，均无白宫正式文件或行政命令确认，事件处于“媒体宣称、官方未证”阶段，真实性与具体执行细节仍有待核实。两则解读信息高度一致，未呈现实质差异。","updated_at":"2026-10-01 08:30:05"},{"id":"hot-demo-001","title":"AI Agent 正在进入企业工作流","category":"AI与科技","heat_score":40,"discussion_volume":6,"summary":"企业开始把 AI Agent 从单点聊天工具接入真实业务流程，热点讨论集中在可靠性、成本与自动化边界。","updated_at":"2026-10-21 10:47:28"},{"id":"f56154db-f68c-488f-bc54-0e9a50318cfa","title":"GPT-6 Sol (Max) Intelligence, Performance and Price Analysis","category":"AI与科技","heat_score":40,"discussion_volume":4,"summary":"综合各模型解读，可确认Artificial Analysis平台存在由theanonymousone于2026-09-22发布的《GPT-6 Sol (Max) 智能、性能与价格分析》报告，标题含(Max)。Kimi-K3还提到同日存在GPT-6 Luna及三份Claude Opus 5.5相关报告，但其他模型未提及。所有模型均表示无法获取具体基准得分、性能数据与定价信息，需查阅原始页面核实。各模型对核心事实的描述高度一致，未发现实质性分歧，仅信息详略不同。未确定项包括具体基准结果、性能指标、价格、对比结论及与其他模型的关系。","updated_at":"2026-09-30 05:33:22"}],"canonical_url":"http://www.aigc.guru/hot/topic/ba286b8c-7331-4b54-8934-4d942787b64f","discussion":null}}