How Logically Inconsistent Prompts Can Turn Reasoning Models Into a DoS Problem

Short summary

Researchers from Zhejiang University and Alibaba demonstrated that logically inconsistent prompts can force reasoning models into unproductive overthinking loops, creating a denial-of-service vector for commercial AI systems. Using an evolutionary algorithm to mutate math problems, they found output lengths up to 26.1x longer than normal on DeepSeek-R1. The attack works externally without model internals access, affecting DeepSeek-R1, Qwen3-Thinking, GPT-o3, and Gemini 2.5 Flash across math, coding, and dialogue tasks.

• •Logically inconsistent prompts can trigger reasoning models into excessive token generation, acting as a DoS vector
• •Evolutionary algorithm mutated math problems to maximize output length, achieving 26.1x increase on DeepSeek-R1
• •Attack works on closed-source models via external queries, impacting cost, latency, and throughput for API providers

Generated with AI, which can make mistakes.