| 1 |
引用产生方式分两个范式:G-Cite(生成时)与 P-Cite(事后归因);P-Cite 覆盖率高、正确性有竞争力,高风险场景推荐 P-Cite 优先 |
https://arxiv.org/html/2509.21557 |
“P-Cite methods achieve high coverage with competitive correctness and moderate latency, whereas G-Cite methods prioritize precision at the cost of coverage and speed. We recommend a retrieval-centric, P-Cite-first approach for high-stakes applications” |
单源(2025 论文,有代码+人评) |
| 2 |
两个范式下引用质量的主驱动都是检索,不是生成模型的引用能力 |
https://arxiv.org/html/2509.21557 |
“a consistent trade-off between coverage and citation correctness, with retrieval as the main driver of attribution quality in both paradigms” |
单源 |
| 3 |
小模型存在 scaffold collapse:复杂格式要求下模型直接放弃遵守结构协议;格式强制必须由 harness 承担,与内容生成能力无关 |
https://arxiv.org/pdf/2605.12129 |
“under complex format requirements, LLaMA 3.2(3B) abandons JSON structure without harness support (TSR=0.429, 7 violations), showing harness format enforcement functions independently of content generation ability” |
单源(2026 预印本,2-3B 模型实验) |
| 4 |
非单调现象:轻量协议包装(minimal-shell)比裸模型更差;要么不加协议,要么上完整 plan→execute→verify→recover 流水线(恢复机制是主要贡献者) |
https://arxiv.org/pdf/2605.12129 |
“A non-monotonic phenomenon—minimal-shell TSR < model-only TSR—is observed in two models… Ablation shows planning and recovery each contribute ~24.7% of total gain” |
单源 |
| 5 |
即使前沿模型,DR 报告链接有效率>94%、相关性>80%,但事实准确率仅 39-77%;开源模型不足一半能 one-shot 产出带引用报告 |
https://doi.org/10.48550/arxiv.2605.06635 |
“even the strongest frontier models maintain link validity above 94% and relevance above 80%, yet achieve only 39-77% factual accuracy, while fewer than half of open-source models successfully generate cited reports in a one-shot setting” |
单源(14 模型 benchmark) |
| 6 |
检索越多引用越不准:工具调用从 2 扩到 150,Fact Check 准确率平均掉 ~42% |
https://doi.org/10.48550/arxiv.2605.06635 |
“Fact Check accuracy drops by approximately 42% on average across two frontier models as tool calls scale from 2 to 150, demonstrating that more retrieval does not produce more accurate citations” |
单源 |
| 7 |
ContextCite(MIT/NeurIPS 2024):事后把 LLM 生成的陈述归因回上下文来源,开源可用 |
https://github.com/MadryLab/context-cite |
“Attribute (or cite) statements generated by LLMs back to in-context information” |
一手(开源仓库,NeurIPS 论文) |
| 8 |
Cohere 走”结构化引用元数据”路线:引用不是模型在正文里写的标记,而是 API 返回的 citation 对象(start/end 索引+text span+sources),Command 系模型开箱自带 |
https://docs.cohere.com/docs/rag-citations |
“Each citation object contains: start and end: the start and end indices of the text that cites a source(s); text: its corresponding span of text; sources: the source(s) that it references” |
一手官方文档 |
| 9 |
Gemini grounding 同样是平台层引用绑定:回答文本附 inline annotations(url_citation 带 start_index/end_index),模型不需在散文里遵守引用协议 |
https://ai.google.dev/gemini-api/docs/interactions/google-search |
“the model’s text output includes inline annotations directly on the text content block… Each url_citation annotation links a text segment” |
一手官方文档 |
| 10 |
训练侧路线(THUDM CaRR):把引用忠实性做进 RL 奖励——问题拆成可验证单跳 rubric,要求显式实体识别+正确引用+完整证据链;比纯结果奖励减少捷径与幻觉 |
https://arxiv.org/abs/2601.06021 |
“CaRR decomposes complex questions into verifiable single-hop rubrics and requires agents to satisfy these rubrics by explicitly identifying hidden entities, supporting them with correct citations, and constructing complete evidence chains” |
单源(清华,代码开源) |
| 11 |
Anthropic Research 采用 orchestrator-worker 拓扑,lead agent 协调并行 subagent |
https://www.anthropic.com/engineering/multi-agent-research-system |
“a multi-agent architecture with an orchestrator-worker pattern, where a lead agent coordinates the process while delegating to specialized subagents that operate in parallel” |
一手官方 |
| 12 |
Anthropic 引用不由研究 agent 产生:研究循环结束后交专门 CitationAgent 基于文档+报告做事后定位 |
https://www.anthropic.com/engineering/multi-agent-research-system |
“passes all findings to a CitationAgent, which processes the documents and research report to identify specific locations for citations. This ensures all claims are properly attributed” |
一手官方 |
| 13 |
Anthropic:多 agent 系统 token ≈ chat 的 15 倍,只适合高价值任务;prompt 工程是改进 agent 行为的首要杠杆 |
https://www.anthropic.com/engineering/multi-agent-research-system |
“multi-agent systems use about 15× more tokens than chats… Since each agent is steered by a prompt, prompt engineering was our primary lever” |
一手官方 |
| 14 |
Anthropic:模型能力是 token 效率乘数,升级基座 > 加倍 token 预算 |
https://www.anthropic.com/engineering/multi-agent-research-system |
“upgrading to Claude Sonnet 4 is a larger performance gain than doubling the token budget on Claude Sonnet 3.7” |
一手官方 |
| 15 |
Kimi 放弃 prompt workflow 的官方理由:绑定特定 LLM 版本、需随模型/环境频繁人工维护 |
https://moonshotai.github.io/Kimi-Researcher/ |
“they are tied to specific LLM versions and need frequent manual updates as models or environments change, reducing scalability and flexibility” |
一手官方 |
| 16 |
Kimi 能力几乎全部来自端到端 RL(HLE 8.6%→26.9%);工具调用格式遵守靠 format reward 训练保证,不靠 prompt 约束 |
https://moonshotai.github.io/Kimi-Researcher/ |
“reached 26.9% almost entirely through end-to-end RL training… Format Reward: The model is penalized for trajectories that include invalid tool calls” |
一手官方 |
| 17 |
Kimi 用上下文管理机制(保留关键/丢弃无用)把单轨迹扩展到 50+ 迭代 |
https://moonshotai.github.io/Kimi-Researcher/ |
“a context-management mechanism that allows the model to retain important information while discarding unnecessary documents, thereby extending a single rollout trajectory to over 50 iterations” |
一手官方 |
| 18 |
OpenAI Deep Research 是在浏览任务上端到端 RL 训练的模型(非外部编排);训练用 CoT 模型按 ground truth/rubric 打分 |
https://cdn.openai.com/deep-research-system-card.pdf |
“reinforcement learning training on these browsing tasks… the model responses are graded against the ground truth answers or rubrics using a chain-of-thought model as a grader” |
一手官方 |
| 19 |
Gemini DR 计划先行且用户可在执行前审改(collaborative planning);Max 版用扩展 test-time compute 迭代 reason-search-refine |
https://ai.google.dev/gemini-api/docs/deep-research |
“the agent returns a proposed research plan instead of executing immediately. You can then review, modify, or approve the plan” |
一手官方 |
| 20 |
Gemini 官方把引用核验责任部分外置给用户(建议复核 citations 字段以防恶意网页) |
https://ai.google.dev/gemini-api/docs/deep-research |
“We recommend reviewing the citations provided in the response to verify the sources” |
一手官方 |
| 21 |
Perplexity(社区拆解):引用在生成前绑定进 pipeline——检索/排序层给每个 chunk 预分配 provenance record(citation_id/url/title/date/snippet),模型只输出标号 |
https://gist.github.com/Co-Messi/bfcfb39eede5c6bc2fadd2c04139a136 |
“Citations in Perplexity are not added by the LLM as an afterthought. They are wired into the pipeline before generation begins… The ranking layer assigns each selected chunk a provenance record” |
二手拆解/推测(单源) |
| 22 |
该拆解断言:引用质量是检索+prompt 结构问题而非模型能力问题;弱模型+结构化引用钩子 > 强模型+泛泛 cite your sources |
https://gist.github.com/Co-Messi/bfcfb39eede5c6bc2fadd2c04139a136 |
“A weaker model with a well-structured citation prompt will outperform a stronger model with a vague ‘cite your sources’ instruction” |
二手拆解/推测(单源) |
| 23 |
Perplexity 官方:DR 是迭代 search-read-reason 并随认知更新计划;源材料评估完后再统一成文 |
https://www.perplexity.ai/hub/blog/introducing-perplexity-deep-research |
“iteratively searches, reads documents, and reasons about what to do next, refining its research plan… Once the source materials have been fully evaluated, the agent then synthesizes all the research into a clear and comprehensive report” |
一手官方 |
| 24 |
LangChain ODR 三阶段:Scope→Research→Write;研究并行、写作在全部研究完成后进行 |
https://www.langchain.com/blog/open-deep-research |
“Open deep research uses an agent to conduct research as part of a three step process: Scope – clarify research scope; Research – perform research; Write – produce the final report” |
一手官方 |
| 25 |
LangChain 亲测失败教训:早期并行分节写作导致报告割裂,改为多 agent 只做研究、写作单独后置 |
https://www.langchain.com/blog/open-deep-research |
“Earlier versions of our research agent wrote sections of the final report in parallel with sub-agents. It was fast, but… the reports were disjoint because the section-writing agents were not well coordinated. We resolved this by using multi-agent for only the research task itself” |
一手官方 |
| 26 |
LangChain 最终报告单次 LLM 调用一次性成文(输入=brief+全部研究发现);引用产生于子 agent 收尾的清洗压缩调用 |
https://www.langchain.com/blog/open-deep-research |
“This final LLM call produces an output in one-shot, steered by the brief… it makes a final LLM call to write a detailed answer to the subquestion posed… citing helpful sources” |
一手官方 |
| 27 |
HF smolagents:同一系统从 CodeAgent 换成 JSON tool-calling,GAIA 验证集 55.15%→33%(-22pp);Code 动作比 JSON 平均少 30% 步数 |
https://huggingface.co/blog/open-deep-research |
“when switching to a standard agent that writes actions in JSON instead of code, performance of the same setup is instantly degraded to 33% average on the validation set” |
一手官方 |
| 28 |
GPT-Researcher:planner-executor 架构;引用在采集侧生成(逐资源 summarize and source-track),非成文侧补挂 |
https://github.com/assafelovic/gpt-researcher |
“The planner generates research questions, while the execution agents gather relevant information… Summarize and source-track each resource. Filter and aggregate summaries into a final research report” |
一手官方 |
| 29 |
GPT-Researcher 对结构化输出配四级兜底链:json_repair→限 token 重试→降级 SMART_LLM→默认 persona |
https://deepwiki.com/assafelovic/gpt-researcher/9.1-query-planning-and-decomposition |
“If parsing fails, the system uses json_repair or falls back to a default ‘AI critical thinker’ persona… Token Limit Retry… falls back to the SMART_LLM” |
二手拆解(DeepWiki,标注源码文件) |
| 30 |
GPT-Researcher × Deep Agents:验证有效引用 18.6→35.2/篇(+89%),官方归因”拉全文+返回预引用综合(pre-cited synthesis)而非搜索摘要” |
https://github.com/assafelovic/gpt-researcher/blob/main/deep_agents/README.md |
“Effective (verified) citations per report: 18.6 → 35.2 (+89%)… scrapes and reads full pages, and returns pre-cited synthesis (instead of search snippets), the same agent grounds its reports in roughly 2x the verified evidence” |
一手官方 |
| 31 |
Deep Agents 版是分节写作的反例实现:researcher 子 agent 各写一节带引用草稿落盘,chief editor 审阅-退回修订-统一组装-参考文献去重 |
https://github.com/assafelovic/gpt-researcher/blob/main/deep_agents/README.md |
“writes a cited draft to sections/*.md, returning only a short summary… Chief Editor - Reads and reviews each draft… Assembles and writes the final report including… a deduplicated references section” |
一手官方 |
| 32 |
STORM:预写作(视角发现+模拟对话提问+大纲)/写作两阶段;成文显式分节,每节用节标题对参考库语义检索取证 |
https://arxiv.org/abs/2402.14207 |
“simulating conversations where writers carrying different perspectives pose questions to a topic expert… the full-length article can be composed section by section… use the section title… to retrieve relevant documents from R” |
一手官方 |
| 33 |
STORM 专家评估的主要失效模式是”红鲱鱼”(牵强关联/塞无关内容),而非事实幻觉 |
https://storm-project.stanford.edu/research/storm/ |
“we find the major challenge stems from red herrings (establishing shaky links or incorporating irrelevant content) rather than widely discussed factual hallucination issue” |
一手官方 |
| 34 |
deer-flow 1.x:Coordinator→Planner→Research Team→Reporter 拓扑;计划节点 human-in-the-loop 可自然语言改计划后执行;Reporter 末端单节点聚合成文 |
https://github.com/bytedance/deer-flow/tree/main-1.x |
“allows you to review, edit, and approve research plans before they are executed… Reporter: Aggregates findings from the research team… Generates comprehensive research reports” |
一手官方 |
| 35 |
deer-flow 2.0 放弃硬编码研究图,重写为通用 super-agent harness,官方推荐开源模型(Doubao-Seed-2.0-Code/DeepSeek v3.2/Kimi 2.5)运行 |
https://github.com/bytedance/deer-flow |
“DeerFlow 2.0 is a ground-up rewrite. It shares no code with v1… We strongly recommend using Doubao-Seed-2.0-Code, DeepSeek v3.2 and Kimi 2.5 to run DeerFlow” |
一手官方 |