OpenResearchAlpha

International Conference on AI Research · 18 Nov 2026

Interaction Persistence and Adversarial Propagation in Multi-Agent Educational LLM Systems

Wenting Sun

0
Citations
no database records one yet
0
Citations in 2026
so far
–
Leading author's h-index
no author has a profile yet
0
References
works it cites, from open citation data

Citations by database

Each database indexes a different set of journals, preprints and books, so they are shown side by side rather than blended.

  • OpenCitations0
  • Crossref0
  • OpenAlex0

Citations per year

The current year is still counting.

  • 2026 so far0

Abstract

Large Language Models are increasingly deployed in educational applications such as tutoring, explanation generation, and automated assessment, yet safety research has largely examined single-turn interactions, leaving open how adversarial influence propagates through the persistent memory and feedback mechanisms of educational multi-agent systems. We investigate three research questions: how interaction persistence affects adversarial propagation, which shared mechanisms enable it, and how educational task characteristics and lightweight defences moderate it. We evaluate a unified three-phase interaction framework across pedagogical explanation generation (BloomXplain), automated essay assessment (ASAP), and dialogic tutoring (ConvoLearn), using Llama-3.1-8B and Qwen3-32B with representative prompt injection attacks and verifier-based defences. Results show that interaction persistence does not uniformly amplify adversarial influence: propagation emerges primarily through repeated reuse of evaluator-generated optimization objectives, while tasks with stronger semantic boundaries resist cross-instance propagation. Verifier-based defences interrupt recursive propagation but can suppress legitimate feedback, revealing a security-utility trade-off. These findings suggest educational AI safety evaluation should jointly consider interaction persistence, propagation mechanism, and utility preservation rather than relying solely on prompt-level robustness evaluation. Research artifact: code and data are available at https://github.com/wentinghu2025/edu-mas-attack-propagation; an interactive results explorer is available at https://wentinghu2025.github.io/edu-mas-attack-propagation/

What the data shows

  • No database records a citation yet, as is usual for a paper this new.

Graph

This paper in the middle, its authors and their most-cited papers below, the papers it cites to the left and the papers citing it to the right. The timeline lays every paper out by year.

Citing paper shares an authorAuthor without a profileSize shows citations, or h-index for authors

Select a node for its details. Drag to pan, scroll to zoom, drag a node to move it.

Start with

Authors (1)

Each author's standing in QAI's scholar index: the works their ORCID iD is attached to, and for anyone whose ORCID record names a lab OpenResearch follows, their arXiv preprints matched by name and a shared co-author. Names are as printed on the paper.

ProfileMost-cited paper

Cited by 0 papers

Who cites this paper, from open citation data, with Crossref's record of each; the 100 newest when there are more.

0 shown
No citing papers for this filter.

References (0)

What it cites, from open citation data, with Crossref's record of each.

No references in the open citation data yet.