---
格式版本: 2
标题: "Pretraining Reusable Inference Across Views with Synthetic Task Priors"
原文链接: "https://arxiv.org/abs/2608.19115"
发布日期: "2026-08-19"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Wed, 19 Aug 2026 17:10:03 UTC (590 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-08-20T15:30:03+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-08-20T15:26:21+08:00"
入库时间: "2026-08-20T07:30:03.379Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=performance&searchtype=all"
匹配关键词:
  - "performance"
相关厂家:
  []
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 6
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "论文主题为多视图机器学习，与超节点/AI Rack、机柜级AI基础设施、供电散热互连等完全无关，无任何相关信号。"
AI质检模型: "ali-deepseek-v4-flash"
AI质检时间: "2026-08-20T15:32:58+08:00"
AI主题相关性: 0
AI来源权威性: 5
AI新颖性: 0
AI技术细节: 0
AI商业部署信号: 0
AI完整性: 1
AI摘要: "该论文提出SIMPLE，一种基于合成任务先验的多视图上下文学习器，将多视图学习重构为可复用的任务条件推理过程，而非固定融合函数。实验显示其冻结版无需更新推理骨干即可达到竞争性能，轻量adapter校准在多数多视图与多组学基准上取得领先。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-07T03:20:03.796Z"
采集批次: "2026年8月20日14点19分32秒"
采集批次ID: "20260820-141932-079"
去重键: "https://arxiv.org/abs/2608.19115"
---

## Computer Science > Machine Learning

## Title:Pretraining Reusable Inference Across Views with Synthetic Task Priors

Authors:[Jielong Lu](https://arxiv.org/search/cs?searchtype=author&query=Lu,+J), [Zhihao Wu](https://arxiv.org/search/cs?searchtype=author&query=Wu,+Z), [Jiajun Yu](https://arxiv.org/search/cs?searchtype=author&query=Yu,+J), [Zhaoliang Chen](https://arxiv.org/search/cs?searchtype=author&query=Chen,+Z), [Haishuai Wang](https://arxiv.org/search/cs?searchtype=author&query=Wang,+H)

[View PDF](https://arxiv.org/pdf/2608.19115) [HTML (experimental)](https://arxiv.org/html/2608.19115v1)

> Abstract:Modern pretrained encoders make representations from heterogeneous views increasingly reusable, but the procedure that determines view utility and combines evidence is still relearned for each downstream task. Consequently, knowledge about view relevance, complementarity, reliability, and missingness is repeatedly discarded rather than transferred across tasks. We therefore reformulate multi-view learning as learning a reusable, task-conditioned inference procedure rather than a fixed fusion function. Based on this perspective, we propose SIMPLE, a prior-fitted multi-view in-context learner that predicts query labels by conditioning on a small labeled support set. Since existing real-world datasets cover only a limited range of view configurations and task structures, we construct a controllable synthetic task prior in embedding space. It generates diverse support-query episodes with varying class structures, shared and view-specific factors, representation geometries, cross-view dependencies, reliability levels, missingness patterns, and distribution shifts. A hierarchical inference architecture then performs reasoning within views, across views, and across support and query samples. Experiments on multi-view and multi-omics benchmarks demonstrate that the frozen variant of SIMPLE achieves competitive performance without updating the inference backbone, while lightweight adapter calibration attains leading performance on most evaluated datasets. Together, the results under frozen, one-shot, and missing-view settings support the central hypothesis that multi-view reasoning itself can be pretrained and reused, while lightweight adapter calibration provides task-specific alignment when needed.

| Subjects: | Machine Learning (cs.LG); Multimedia (cs.MM) |
| --- | --- |
| Cite as: | [arXiv:2608.19115](https://arxiv.org/abs/2608.19115) \[cs.LG\] |
|  | (or [arXiv:2608.19115v1](https://arxiv.org/abs/2608.19115v1) \[cs.LG\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2608.19115](https://doi.org/10.48550/arXiv.2608.19115) |

## Submission history

From: Jielong Lu \[[view email](https://arxiv.org/show-email/bfb56ace/2608.19115)\]  
**\[v1\]** Wed, 19 Aug 2026 17:10:03 UTC (590 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2608.19115) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))