<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Multi-Turn Dialogue | Xinjie Shen</title><link>https://xinjie-shen.com/tag/multi-turn-dialogue/</link><atom:link href="https://xinjie-shen.com/tag/multi-turn-dialogue/index.xml" rel="self" type="application/rss+xml"/><description>Multi-Turn Dialogue</description><generator>Wowchemy (https://wowchemy.com)</generator><language>en-us</language><lastBuildDate>Wed, 19 Aug 2026 20:00:00 +0800</lastBuildDate><image><url>https://xinjie-shen.com/media/icon_hu646f7301b7fde7528ecdae8cec89fc29_9606_512x512_fill_lanczos_center_3.png</url><title>Multi-Turn Dialogue</title><link>https://xinjie-shen.com/tag/multi-turn-dialogue/</link></image><item><title>Distributed Hidden Malice: Attacks and Defenses</title><link>https://xinjie-shen.com/talk/distributed-hidden-malice-attacks-and-defenses/</link><pubDate>Wed, 19 Aug 2026 20:00:00 +0800</pubDate><guid>https://xinjie-shen.com/talk/distributed-hidden-malice-attacks-and-defenses/</guid><description>&lt;p>The talk connects two sides of multi-turn agent safety: searching for harmful capability assembled from locally benign exchanges, and intervening at the earliest turn where that capability becomes actionable.&lt;/p>
&lt;p>It covers &lt;a href="https://arxiv.org/abs/2512.01353" target="_blank" rel="noopener">The Trojan Knowledge&lt;/a>, &lt;a href="https://github.com/Graph-COM/CKA-Agent" target="_blank" rel="noopener">CKA-Agent&lt;/a>, and &lt;a href="https://arxiv.org/abs/2605.05630" target="_blank" rel="noopener">TurnGate&lt;/a>.&lt;/p></description></item><item><title>One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue</title><link>https://xinjie-shen.com/publication/conference-paper/turngate/</link><pubDate>Thu, 07 May 2026 00:00:00 +0000</pubDate><guid>https://xinjie-shen.com/publication/conference-paper/turngate/</guid><description>&lt;p>Hidden malicious intent can be distributed across individually benign-looking turns. TurnGate uses response-aware, turn-level supervision to identify when the conversation crosses the harm-enabling boundary.&lt;/p></description></item></channel></rss>