title: "Wrong by Design: Adversarial Speculative Reasoning Improves Long Horizon Agents"

Adversarial Speculative Reasoning (ASpec) is a new technique for improving the accuracy of long-horizon AI agents. It modifies speculative decoding by training the draft model to intentionally guess wrong at key decision points, such as when an agent calls a tool. This forces the main, more powerful model to step in and generate these critical tokens, reducing the risk of small, early mistakes that can compound over time. The method has been shown to improve success rates on complex software engineering benchmarks for models like GLM 5.3 and DeepSeek V4.1 Flash. ASpec aims to increase agent accuracy and decision-making reliability while also improving system throughput.