CiteArk

今日值得读

聚焦大语言模型与智能体研究，不按热度排名。Agent 每天分别选出真正值得读的新工作，先生成可检查的研究计划，再持续补齐复现结果。

1大语言模型arXiv 2608.284110/4

LongPIBench: A Long-Context Benchmark for Prompt Injection

Yupei Liu, Yuqi Jia, Neil Zhenqiang Gong, Jinyuan Jia

为什么值得读

The work targets a practical security gap: an assistant can be given a long document containing untrusted text and quietly follow an embedded instruction instead of the user’s task. This benchmark can help model developers and deployers test review, hiring, email, and code workflows under more realistic document lengths, where simple defenses may look reliable on short inputs but fail in use. Its value is conditional on reproducing the task-specific data, model versions, attack criteria, and defense pipelines; the paper itself also limits the scope to static document processing rather than multi-step agents and does not identify why context length causes each degradation.

核心研究结论

  • Prevention-based defenses that perform well on short-context benchmarks retain substantial attack success on the paper's long-context synthetic benchmark, although PromptLocate and MetaSecAlign 8B are lower than several simpler defenses on some tasks.
  • The paper reports that detection-based defenses exhibit an extreme false-positive/false-negative trade-off on long-context inputs, with some methods flagging benign inputs and others missing attacks.
  • The paper reports that GCG and its universal variant achieve high attack success across all four task suites and outperform heuristic attacks on several tasks.
密码学与安全计算与语言机器学习 (cs.LG)
2智能体arXiv 2608.282640/3

Finding Where the Buck Stops: An Automated Failure Attribution-Based Reflection Framework for Multi-Agent Collaboration

Xiaoqing Wang, Keman Huang, Bin Liang, Hongyu Li, Xiaoyong Du, Wuqiong Pan

为什么值得读

The work targets a practical failure mode in multi-agent systems: feedback can be wasted on agents that did not cause the failure, potentially confusing future behavior. If the reported pattern holds, diagnosing the first consequential mistake and concentrating correction on its owner could make reflective systems more useful under limited inference or training budgets, especially for multi-step question answering, chart analysis, and web interaction. The value is conditional on the paper’s task-specific prompts, model choices, external services, failure annotations, and evaluation setup, and the fixed paper does not provide a verified implementation snapshot for direct reproduction here.

核心研究结论

  • Across five trials, DoCtOR improves initial task success rates by 22% on HotPotQA, 26% on ChartQAPro, and 27% on Mind2Web, and outperforms the reported reflection baselines.
  • The evaluation is limited to English-language task-oriented collaborations with relatively small multi-agent teams and leaves open performance on open-ended or culturally diverse tasks and larger teams.
  • ProFA achieves the strongest reported agent-level and step-level failure-attribution accuracies among the listed methods on the held-in HotPotQA, ChartQAPro, and Mind2Web datasets and on the held-out Algorithm-Generated and Hand-Crafted subsets.
人工智能计算与语言机器学习 (stat.ML)