Reproduce “Resource Consumption Threats in Large Language Models” on OpenResearch

About this paper

Yuanhe Zhang, Xinyue Wang, Zhican Chen, Weiliu Wang, Zilu Zhang, Zhengshuo Gong, Zhenhong Zhou, Kun Wang and 3 more

A systematic review unifies the fragmented understanding of resource consumption threats in large language models, categorizing them into "Overthinking" and "Unbounded Drift" regimes. It details attack surfaces, mechanisms, and mitigation strategies across various LLM architectures, also proposing a standardized evaluation framework.

Key insights

• Resource consumption has evolved from a performance issue to a critical systemic safety risk for LLMs, directly impacting service reliability, availability, and economic sustainability.
• The distinction between "Overthinking" (excessive but task-aligned generation) and "Unbounded Drift" (runaway, non-convergent generation) provides a crucial lens for understanding and addressing diverse resource abuse phenomena.
• Complex LLM ecosystems, including Reasoning LLMs, Multimodal LLMs, and agentic systems, significantly expand the attack surface for resource consumption, necessitating a broader analytical approach beyond standard LLMs.

Summary by alphaXiv.