32B 小模型逼近 GPT-5.2:S1-DeepResearch 如何用数据而非参数训练长程研究 Agent
A 32B Model Closing in on GPT-5.2: How S1-DeepResearch Trains Long-Horizon Research Agents with Data, Not Scale
BLUF:ScienceOne 团队开源了 S1-DeepResearch(论文 arXiv:2606.15367,模型 32B + 8B-Preview,Apache 2.0)。它最反直觉的一点是:一个 32B 的小模型,只靠监督微调(SFT)、完全不用强化学习,就在覆盖 5 大能力、20 个基准的评测里逼近 GPT-5.2、Claude-4.6、GLM-5 这些闭源旗舰,甚至在文本推理上追平了体量大 71 倍的 Qwen3.5-397B。它的核心不是模型架构,而是一套**「知识图谱出题 → Agent 跑轨迹 → 五维验证」的数据构造范式**,以及随之开源的 15K 高质量 Agent 轨迹数据集。这篇文章拆解它怎么做到的,并梳理业界对「深度研究 Agent」这条路线的最新反思。
一、深度研究 Agent 卡在哪?
过去一年,「Deep Research」从 OpenAI、Google 的产品名,变成了一整类 Agent 能力的统称:给一个复杂问题,让模型自己上网搜索、读文件、跑代码、反复推理,最后产出一份带引用的报告。这件事听起来简单,工程上却极难——因为它要求模型在**长程(long-horizon)、多跳(multi-hop)**的任务里持续保持状态、不跑偏、不幻觉。
而真正卡住开源社区的,不是模型不够大,而是训练数据的形态错了。S1 团队的判断很直接:现有的 Agent 训练数据过度集中在「搜索」这一个动作上(search-centric)。大量数据集本质是「多跳问答」——把答案藏深一点,逼模型多搜几次。但真实的研究工作远不止搜索:
- 你要读懂一个 PDF、一张表格、一段网页,而不只是检索关键词;
- 你要遵守复杂指令——“只用 2024 年后的一手资料""分三部分""每段不超过 200 字”;
- 你要写出有据可查的报告,每个论点都能追溯到引用;
- 你要调用技能——画图、跑数据分析、生成文件,而不是把代码塞在聊天框里。
只在「搜索」上训练出来的 Agent,benchmark 分数可能不低,但一碰真实任务就露馅。S1-DeepResearch 想解决的,正是这个**「能力维度太窄」**的根本问题。
二、S1-DeepResearch 是什么?
一句话:一个端到端的长程深度研究模型系统,把上面四类被忽视的能力和「搜索」一起,统一进了训练目标。
关键规格:
| 维度 | 参数 |
|---|---|
| 模型 | S1-DeepResearch-32B(旗舰)+ 8B-Preview(轻量探索版) |
| 基座 | Qwen3 |
| 上下文 | 128K tokens |
| 连续工具调用 | 150+ 轮 |
| 内置工具 | 9 类(网页搜索、学术搜索、网页浏览、图像/视频分析、文件解析、代码执行、命令行等) |
| 许可证 | Apache 2.0(模型 + 15K 轨迹数据集全开源) |
它把深度研究拆成五个能力维度,并在每个维度上都构造了训练数据:
- 长链复杂推理 —— 多文档检索、证据聚合、跨多轮维持状态记忆;
- 研究指令遵循 —— 解析带多重约束的研究指令;
- 研究报告写作 —— 产出有引用、可溯源的结构化报告;
- 文件理解与生成 —— 处理 PDF、表格、网页并生成结构化产物;
- 技能调用 —— 文献检索、数据分析、可视化等模块化能力。
“150+ 连续工具调用”是个值得停下来体会的数字:它意味着模型要在一条任务里做上百次「观察环境→决策→调工具→读反馈」的循环而不崩溃。这正是「长程」二字的含金量所在。
三、核心方法:用数据,而不是参数
S1 最大的贡献不在模型,而在怎么造出那 15K 条高质量轨迹。整套范式分三步。
第一步:知识图谱出题(Graph-Grounded Task Formulation)
不是让 LLM 凭空编题,而是让知识图谱来保证题目的复杂度可控、可验证。
- 种子实体从 Wikipedia 选取,经过拓扑结构、热度、信息密度、可搜索性、安全性五重过滤;
- 围绕每个实体,用 Wikidata 关系 + 网页搜索结果展开成一张有向无环图(DAG),并纳入多模态信息(文本实体、视觉概念);
- 出题前先注入九个约束维度(信息来源、论证方式、推理路径、目标、假设、输出格式、规模、执行、上下文),界定探索边界;
- 对闭式问答,用「实体语义改写 + 条件泛化」反复重写题目,削弱问题和答案之间的直接关联,逼模型真去推理而非背答案;
- 最关键的过滤:先不给工具让模型裸答一遍(过滤掉靠参数记忆就能答的题),再用信息流、反馈依赖、宽度、深度四个指标度量推理图复杂度,只保留最难的前 30%。
这一步解决的是「题目质量」——垃圾进、垃圾出,数据范式的成败首先取决于题出得够不够硬。
第二步:Agent 跑轨迹(Agentic Trajectory Rollout)
有了题,让 Agent 在一个真实的 9 类工具环境里实际去解,记录下完整轨迹:
τ = (x, a₁, o₁, a₂, o₂, …, aₜ, oₜ, y)
其中 x 是任务,aₜ 是模型的动作,oₜ 是环境观测(网页、解析后的文件、代码执行结果),y 是最终回答。模型反复「观察→决策→调工具→收反馈」,自然产生长程多步行为。
然后做场景化精修,把轨迹改造成更真实的形态:报告类强化证据连贯性;文件类把文本输出转成可执行产物;多模态类插入原生视觉输入;技能类把内联代码转成文件附件。这一步让训练数据贴近真实研究产物,而不是干巴巴的问答对。
第三步:五维验证过滤(Multi-Dimensional Verification)
跑出来的轨迹大部分是噪声,必须严格过滤。S1 按五条能力轨道分别验证:
- 复杂推理:LLM-as-judge 对照参考答案校验逻辑正确性和中间推导;
- 报告生成:引用校验器检查正文引用是否真实对应参考、是否提供实质支撑;
- 指令遵循:九维约束检查器逐项核对;
- 文件任务:验证轨迹与可执行代码的一致性、与任务的语义对齐;
- 技能调用:追踪技能是否正确激活、是否有效、资源利用率如何。
三步走完,沉淀出 15,000 条横跨五大维度的高质量轨迹。
一个反直觉的设计:只用 SFT,不用 RL
值得专门点出:S1 的训练只有监督微调,没有强化学习。目标函数就是让模型同时模仿中间动作和最终回答:
ℒ = −Σ log p(aₜ | x, a<t, o<t) − log p(y | x, a≤T, o≤T)
在「人人都在卷 Agentic RL」的 2026 年,这是一个有点逆潮流的选择。它传递的信号是:当数据质量足够高、能力维度足够全时,SFT 本身就能把一个 32B 模型推到逼近旗舰的水平——参数效率来自数据,而非堆规模或堆 RL。这对算力有限的团队是个重要的方法论参考。
四、跑分:32B 追平百亿级旗舰
S1 在覆盖 5 大维度的 20 个 Agentic 基准上做了评测(含 GAIA、BrowseComp、BrowseComp-ZH、XBench-DeepSearch、HLE、DeepResearch Bench、ComplexBench、FileSys、GTA、SkillsUse 等)。论文报告的几组关键数字:
相对基座 Qwen3-32B 的提升(同样 32B,差距全来自训练数据):
| 基准 | Qwen3-32B | S1-DeepResearch-32B |
|---|---|---|
| GAIA(文本) | 30.2 | 72.8 |
| DeepResearch Bench | 36.0 | 46.5 |
| DeepResearchIF(查询准确率) | 4.1 | 25.2 |
| FileSys(文件系统任务) | 44.7 | 69.3 |
GAIA 文本从 30 分翻到 72 分、指令遵循从 4 分跳到 25 分——这种量级的提升,纯粹来自数据,没有动模型架构。
相对闭源旗舰: 论文称整体性能接近 GPT-5.2、Claude-4.6、GLM-5——报告生成逼近 GPT-5.2(48.7 vs 51.4),文件任务达到 GLM-5 水平(尽管参数差 23 倍)。更具冲击力的是:在文本推理上,32B 的 S1 追平了体量大 71 倍的 Qwen3.5-397B。
⚠️ 需要客观看待:这 20 个基准是团队自己选定的能力维度,“逼近旗舰”是在这套评测口径下的结论;闭源模型在更开放的真实任务上仍可能有优势。但即便打个折扣,“32B + 纯 SFT 逼近百亿级旗舰”这个结论本身,已经足够说明数据范式的威力。
五、业界在思考什么?深度研究 Agent 的三条暗线
S1 不是孤例。把它放进 2026 年的大背景里看,会发现整个领域正沿三条暗线推进。
暗线一:竞争焦点从「模型」转向「轨迹数据怎么造」
S1 之外,同期还有一批工作在解同一个问题:
- Tongyi DeepResearch(通义,技术报告 arXiv:2510.24701)系统性地论证了合成轨迹数据对深度研究能力的决定性作用;
- OpenResearcher(arXiv:2603.20278)干脆开源了一整套「长程深度研究轨迹合成」的全流程管线;
- Marco DeepResearch(arXiv:2603.28376)则主打「以验证为中心」的设计来提升 Agent 效率。
共识正在形成:深度研究能力的瓶颈不在基座模型,而在能不能规模化地造出高质量、多维度、可验证的轨迹数据。S1 的三步范式,是这条路线上一个相当完整的样本。
暗线二:评测本身成了战场
当大家都宣称「逼近 GPT-5.2」,怎么评测就变得至关重要。2026 年涌现了一批专门修正评测口径的工作:
- BrowseComp-Plus(ACL 2026)用固定语料库把「检索器」和「Agent 推理」解耦,让对比公平、可复现——因为开放网络评测里,分数高可能只是因为它搜到了更好的页面,而非推理更强;
- DeepResearch Bench(arXiv:2506.11763)、DeepResearchEval(arXiv:2601.09688)则在构建更贴近真实研究任务的自动化评测框架。
这背后是一个尖锐的质疑:刷榜分数 ≠ 真实有用。
暗线三:对「自主性」的祛魅
最值得创业者和研究者警惕的是 arXiv:2512.01948 这篇《How Far Are We from Genuinely Useful Deep Research Agents?》提出的反思。它指出当前深度研究 Agent 普遍存在一个**「性能—效用鸿沟」(performance-utility gap)**:
- 任务复杂度:真实研究是细腻、多面的问题,远超简单检索或事实核查;
- 信息整合:Agent 擅长检索,但很难把分散来源综合成推进认知的连贯叙事;
- 推理深度:缺乏迭代式假设检验和实质性分析推理的机制;
- 用户对齐:系统倾向于优化「任务完成」,而非满足研究者真正的认知目标。
它的结论很冷静:当前 Agent 的「自主」宣称可能被高估了,人类的监督和迭代修正仍然不可或缺。——这是给所有「全自动研究 Agent」叙事的一盆冷水,也是 S1 这类工作下一步真正要跨越的鸿沟。
六、这对我们意味着什么?
把三条暗线收拢,S1-DeepResearch 的价值有三层:
- 对开发者:它和 15K 数据集都是 Apache 2.0 开源的。如果你在做垂直领域的研究 Agent,这套「图谱出题→轨迹→验证」的范式比直接拿模型微调更值得借鉴——先把数据范式立住,再谈训练。
- 对算力有限的团队:“32B + 纯 SFT 逼近旗舰”证明了一条不依赖海量算力和 RL 的可行路径。数据工程的杠杆,可能比想象中大得多。
- 对整个行业:深度研究 Agent 正在从「能不能搜到」走向「能不能真正帮人推进认知」。前者已被基本解决,后者——按 arXiv:2512.01948 的判断——我们还在半路上。
S1 是这条半路上一块扎实的路标:它没有解决「自主性」的终极问题,但它把「怎么造数据」这件最脏最累、却最关键的活,做出了一个可复现的开源范本。
FAQ
Q:S1-DeepResearch 和 OpenAI Deep Research、通义 DeepResearch 有什么不同? A:OpenAI Deep Research 是闭源产品;通义 DeepResearch 同样强调合成轨迹,但 S1 的差异化在于显式地把能力拆成五维(推理/指令/报告/文件/技能)并各自构造数据,且完全只用 SFT、不用 RL,模型和 15K 数据集全部 Apache 2.0 开源。
Q:为什么是 32B 而不是更大? A:论文的核心论点恰恰是「参数效率来自高质量多维度轨迹数据,而非模型规模」。32B 在文本推理上追平 397B、在文件任务上达到 GLM-5 水平,就是为了证明这一点。另有 8B-Preview 供轻量探索。
Q:普通人现在能用吗? A:模型在 HuggingFace(ScienceOne-AI/S1-DeepResearch-32B)可下载,代码在 GitHub(ScienceOne-AI/S1-DeepResearch),需自行部署带 9 类工具的沙箱环境,对工程能力有一定要求,暂非开箱即用的产品。
📌 原文与资源(请直接复制访问): 论文 arXiv:2606.15367 —— https://arxiv.org/abs/2606.15367 GitHub —— https://github.com/ScienceOne-AI/S1-DeepResearch 模型 —— https://huggingface.co/ScienceOne-AI/S1-DeepResearch-32B 数据集 —— https://huggingface.co/datasets/ScienceOne-AI/S1-DeepResearch-15k
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用,须注明作者姓名及原文链接,不得去除署名后以原创发布。
BLUF: The ScienceOne team open-sourced S1-DeepResearch (paper arXiv:2606.15367; 32B + 8B-Preview, Apache 2.0). Its most counterintuitive claim: a 32B model, trained with supervised fine-tuning only and no reinforcement learning, approaches closed-source flagships (GPT-5.2, Claude-4.6, GLM-5) across 20 benchmarks spanning five capability dimensions — even matching the 71× larger Qwen3.5-397B on textual reasoning. The secret isn’t architecture; it’s a graph-grounded trajectory construction paradigm — “knowledge graph poses tasks → agent rolls out trajectories → five-track verification filters” — plus the 15K high-quality agent trajectory dataset released with it. This piece breaks down how it works and where the industry’s thinking on deep research agents is heading.
1. Where Do Deep Research Agents Get Stuck?
Over the past year, “Deep Research” went from an OpenAI/Google product name to shorthand for an entire class of agent capability: hand the model a hard question, let it search the web, read files, run code, reason iteratively, and produce a cited report. Simple to state, brutally hard to build — it demands the model hold state across long-horizon, multi-hop tasks without drifting or hallucinating.
The S1 team’s diagnosis: the open-source bottleneck isn’t model size, it’s the shape of the training data. Existing agent datasets are overwhelmingly search-centric — essentially multi-hop QA that just buries the answer deeper. But real research is far more than search: reading a PDF or table, following multi-constraint instructions, writing traceable reports, invoking skills (plotting, data analysis, file generation). Agents trained only on “search” score fine on benchmarks but collapse on real tasks. S1-DeepResearch targets exactly this capability-dimension narrowness.
2. What Is S1-DeepResearch?
An end-to-end long-horizon deep research model system that folds those four neglected capabilities — alongside search — into a unified training objective.
| Spec | Value |
|---|---|
| Model | S1-DeepResearch-32B (flagship) + 8B-Preview |
| Base | Qwen3 |
| Context | 128K tokens |
| Consecutive tool calls | 150+ rounds |
| Built-in tools | 9 categories (web/academic search, browsing, image/video analysis, file parsing, code execution, bash) |
| License | Apache 2.0 (model + 15K trajectory dataset) |
Five capability dimensions: long-chain reasoning, research instruction following, report writing, file understanding/generation, and skill usage. The “150+ consecutive tool calls” figure is worth pausing on — it means the model runs hundreds of “observe → decide → call → read feedback” loops in a single task without breaking. That’s what “long-horizon” really costs.
3. The Core Method: Data, Not Scale
S1’s main contribution is how it builds those 15K trajectories — a three-stage paradigm.
Stage 1 — Graph-grounded task formulation. Rather than having an LLM invent questions, a knowledge graph controls difficulty and verifiability. Seed entities are filtered from Wikipedia (topology, popularity, info density, searchability, safety), expanded into a DAG via Wikidata relations and web search (with multimodal info), then nine constraint dimensions are injected before generation. Closed-form questions are rewritten to weaken the answer’s direct association, forcing reasoning over recall. Crucially, tasks are first answered without tools to filter out anything solvable from parametric memory — then only the top 30% most complex (by information flow, feedback dependency, width, depth) survive.
Stage 2 — Agentic trajectory rollout. The agent solves tasks in a real 9-tool environment, recording full trajectories τ = (x, a₁, o₁, …, aₜ, oₜ, y). Scenario-specific refinement reshapes outputs into realistic artifacts (coherent cited reports, executable files, native multimodal inputs, file attachments instead of inline code).
Stage 3 — Multi-dimensional verification. Five tracks filter noise: LLM-as-judge for reasoning, a citation verifier for reports, a nine-dimensional constraint checker for instruction following, code-consistency checks for file tasks, and skill-activation tracking. The result: 15,000 high-quality trajectories across five dimensions.
The counterintuitive choice: SFT only, no RL. The objective simply imitates both intermediate actions and final answers: ℒ = −Σ log p(aₜ | x, a<t, o<t) − log p(y | x, a≤T, o≤T). In a year where everyone is racing on agentic RL, the message is pointed: with high enough data quality across enough capability dimensions, SFT alone pushes a 32B model to flagship level — parameter efficiency comes from data, not scale or RL.
4. The Numbers: 32B Matching Giants
Across 20 agentic benchmarks (GAIA, BrowseComp/-ZH, XBench-DeepSearch, HLE, DeepResearch Bench, ComplexBench, FileSys, GTA, SkillsUse, etc.):
vs. base Qwen3-32B (same size — gap is all data):
| Benchmark | Qwen3-32B | S1-DeepResearch-32B |
|---|---|---|
| GAIA (text) | 30.2 | 72.8 |
| DeepResearch Bench | 36.0 | 46.5 |
| DeepResearchIF (query acc.) | 4.1 | 25.2 |
| FileSys | 44.7 | 69.3 |
vs. closed-source flagships: overall performance close to GPT-5.2, Claude-4.6, GLM-5 — report generation near GPT-5.2 (48.7 vs 51.4), file tasks at GLM-5 level (despite a 23× parameter gap), and textual reasoning matching the 71× larger Qwen3.5-397B.
⚠️ Read with care: these 20 benchmarks are the team’s own chosen dimensions; “approaching flagships” holds under that evaluation lens, and closed models may still lead on more open real-world tasks. Even discounted, “32B + pure SFT approaching giants” demonstrates the power of the data paradigm.
5. What Is the Industry Thinking? Three Undercurrents
Undercurrent 1 — competition shifts from “the model” to “how you build trajectories.” Tongyi DeepResearch (arXiv:2510.24701) argues synthetic trajectory data is decisive; OpenResearcher (arXiv:2603.20278) open-sourced a full long-horizon synthesis pipeline; Marco DeepResearch (arXiv:2603.28376) pushes verification-centric design. Emerging consensus: the bottleneck is scalable, high-quality, multi-dimensional, verifiable trajectory data, not the base model.
Undercurrent 2 — evaluation becomes the battleground. When everyone claims “approaching GPT-5.2,” how you measure matters. BrowseComp-Plus (ACL 2026) decouples retriever from agent reasoning via a fixed corpus for fair, reproducible comparison; DeepResearch Bench and DeepResearchEval build more realistic automated frameworks. The sharp question underneath: leaderboard scores ≠ genuine usefulness.
Undercurrent 3 — disenchantment with “autonomy.” The most sobering read is arXiv:2512.01948, How Far Are We from Genuinely Useful Deep Research Agents?, which names a performance-utility gap: agents handle retrieval but struggle to synthesize sources into knowledge-advancing narratives; they lack iterative hypothesis-testing; they optimize task completion over researchers’ real epistemic goals. Its verdict: current autonomy claims are likely overstated; human oversight remains essential. That’s the gap S1-class work must still cross.
6. What This Means
- For developers: both model and 15K dataset are Apache 2.0. For vertical research agents, the “graph → trajectory → verification” paradigm is more worth borrowing than the weights themselves — get the data paradigm right first.
- For compute-constrained teams: “32B + pure SFT approaching flagships” proves a path that doesn’t depend on massive compute or RL. The leverage of data engineering may be larger than assumed.
- For the field: deep research is moving from “can it find it” (largely solved) to “can it genuinely advance human understanding” (per arXiv:2512.01948, still mid-journey).
S1 is a solid signpost on that journey: it doesn’t solve autonomy, but it turns the dirtiest, most essential work — how to build the data — into a reproducible open-source template.
FAQ
Q: How does it differ from OpenAI Deep Research or Tongyi DeepResearch? A: OpenAI’s is closed; Tongyi also stresses synthetic trajectories, but S1 explicitly decomposes capability into five dimensions with per-dimension data, uses SFT only (no RL), and open-sources everything under Apache 2.0.
Q: Why 32B, not bigger? A: The thesis is that parameter efficiency comes from high-quality multi-dimensional trajectory data, not scale. Matching 397B on reasoning and GLM-5 on file tasks is the proof. An 8B-Preview exists for lightweight exploration.
Q: Can I use it now? A: Model on HuggingFace (ScienceOne-AI/S1-DeepResearch-32B), code on GitHub (ScienceOne-AI/S1-DeepResearch). You must deploy a sandbox with the 9 tools yourself — it’s research-grade, not yet a turnkey product.
📌 Source & resources (copy to visit): Paper arXiv:2606.15367 — https://arxiv.org/abs/2606.15367 GitHub — https://github.com/ScienceOne-AI/S1-DeepResearch Model — https://huggingface.co/ScienceOne-AI/S1-DeepResearch-32B Dataset — https://huggingface.co/datasets/ScienceOne-AI/S1-DeepResearch-15k
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
💬 评论与讨论
使用 GitHub 账号登录后发表评论