ScholarPulse 日报 2026-07-19

2026-07-19 学术简报：2 篇。SYNAPSE是一个采用神经多样性优先设计的多LLM协调AI辅导平台，用于安全软件开发教育。

今日速览

序号	标题	来源	日期	主题	推荐等级
1	SYNAPSE: A Multi-LLM Orchestrated AI Tutor for Secure Software Development Education with Neurodivergent-First Design	arXiv	2026-07-16	MCP	高
2	AdaTurn: Budget-Aware Test-Time Scaling for Active Visual Perception Agents	arXiv	2026-07-16	AI-Agent	高

重点论文与技术动态

1. SYNAPSE: A Multi-LLM Orchestrated AI Tutor for Secure Software Development Education with Neurodivergent-First Design

来源：arXiv
日期：2026-07-16
作者/机构：Giusy Ferrara, Ashkan Sami
主题标签：MCP, arXiv
推荐等级：高
分类：cs.SE

一句话结论

SYNAPSE是一个采用神经多样性优先设计的多LLM协调AI辅导平台，用于安全软件开发教育。

一段话：该平台协调Claude、GPT-4o和Gemini通过Model Context Protocol实现基于教学意图的交互路由，集成18个始终可见的无障碍功能支持ADHD等执行功能差异学习者。基于ShopSecure应用（映射OWASP Top 10 (2021)漏洞类别）实践检测-理解-修复循环，可行性试点19名参与者显示系统可用性量表76.4、参与度4.2/5，认知负荷水平在神经多样性和典型群体间相似。

核心内容

协调Claude、GPT-4o和Gemini通过Model Context Protocol，实施三阶段苏格拉底提示策略。
集成18个始终可见的无障碍功能，针对性支持ADHD等执行功能差异学习者。
基于ShopSecure应用实践OWASP Top 10 (2021)漏洞修复，锚定软件维护的检测-理解-修复循环。

方法与数据

使用Model Context Protocol协调多LLM，并在ShopSecure上实践漏洞修复。
可用性试点数据：19名参与者，系统可用性量表76.4，参与度4.2/5。

价值判断

值得关注：神经多样性优先设计有效解决安全教育中的无障碍问题，提升ADHD学习者参与度。
可复用点：多LLM协调框架和ShopSecure应用可推广至其他安全教育平台。
局限/待核查：试点规模小（19人），需更大样本验证认知负荷相似性。

摘要

Developers who maintain real systems must continually recognise and remediate vulnerabilities in existing code, yet this skill is rarely trained directly: secure software development is commonly taught only after programming fluency is acquired, and accessibility support is treated as a secondary concern, disadvantaging learners with ADHD and related executive-function differences. This paper presents SYNAPSE, a publicly deployed adaptive tutoring platform for Java, Python for cybersecurity, and secure software development. SYNAPSE coordinates Claude, GPT-4o, and Gemini through the Model Context Protocol, routing interactions by pedagogical intent under a three-stage Socratic hint policy. It exposes eighteen always-visible accessibility features and anchors practice in ShopSecure, a deliberately vulnerable web application mapped to six OWASP Top 10 (2021) categories, on which learners practise the detect-understand-remediate loop characteristic of software maintenance. A feasibility pilot with nineteen participants across neurodivergent and neurotypical cohorts returned a System Usability Scale score of 76.4 and engagement of 4.2/5, with comparable cognitive-load levels across cohorts. SYNAPSE is available at https://synapse-course.com; a screencast is available at https://youtu.be/9R17KC47qQI.

2. AdaTurn: Budget-Aware Test-Time Scaling for Active Visual Perception Agents

来源：arXiv
日期：2026-07-16
作者/机构：Susan Liang, Chao Huang, Filippos Bellos, Jing Bi, Jason J Corso, Chenliang Xu
主题标签：AI-Agent, arXiv
推荐等级：高
分类：cs.CV

一句话结论

AdaTurn通过预算感知框架有效解决主动视觉感知代理在低预算条件下的catastrophic truncation问题，显著提升任务准确性。

核心内容

现有方法在预算不足时导致交互在有效答案前被截断（catastrophic truncation），因未考虑实际预算限制。
AdaTurn条件化于允许的回合数，显式训练预算边界行为，避免交互中断。
关键组件FA-DAPO将超预算事件转化为可训练的最终决策步骤，使模型在无法进一步操作时合成部分证据。

方法与数据

随机化训练和推理中的滚动预算，并引入负载均衡调度器提升操作效率。
摘要未明确数据集。

价值判断

值得关注：在4回合条件下，VisualProbe-Medium准确率从36.7%提升至47.6%，同时保持大预算下的强扩展性。
可复用点：框架可迁移至多种骨干网络及通用多模态基准，有效提升任务性能。
局限/待核查：摘要未明确提及具体局限。

摘要

Active visual agents solve fine-grained image tasks by interleaving reasoning with image-grounding actions across multiple turns. However, deployment-time rollout budgets are rarely fixed: some requests permit long rollouts, while others require the agent to act under a tight turn limit. Existing methods train the policy as if the rollout budget were hidden, so when the available budget is smaller than the trajectory the agent prefers, the interaction is often truncated before any valid answer is produced; we term this failure \emph{catastrophic truncation}. To overcome this challenge, we present AdaTurn, a budget-aware framework that conditions the agent on the allowed number of turns and explicitly trains the boundary behavior induced by the budget. Our key component, Forced-Answer DAPO (FA-DAPO), converts the over-budget event from a masked or penalized failure into a trainable final-decision step, teaching the model to synthesize partial evidence when further tool use is no longer possible. We further randomize rollout budgets during both training and inference and introduce a load-balanced scheduler that makes such operations practical. AdaTurn substantially improves low-budget accuracy, for example raising VisualProbe-Medium from 36.7% to 47.6% at four turns, while preserving strong scaling at larger budgets and transferring effectively to multiple backbones and general multimodal benchmarks.