Research
Evolva: A Multi-turn Contextual Attack for Long-Reasoning LLMs | Knowledge Science, Engineering and Management
Publication date not yet evaluated · Added acm.org
Page published
Coverage timeline
Single-source research — one report is available.
Why it matters
Evolva demonstrates that long-range reasoning and multi-turn interaction meaningfully expand the LLM attack surface and degrade safety robustness, showing defenders that single-turn prompt filters are insufficient against gradual context poisoning.
Evolva is a research jailbreak framework that formalizes multi-turn prompt attacks as a multi-stage structured process, using auxiliary LLMs to build hierarchical, progressively complex prompt sequences that inject adversarial context across dialogue rounds to bypass prompt-level safety controls on long-reasoning LLMs. The authors introduce an Attack_Robust metric and evaluate several state-of-the-art models on REDTask, a red-teaming dataset for multi-turn adversarial prompting, reporting that Evolva outperforms prior approaches.