AutoTTS: Enabling AI to Automatically Discover Optimal Test-Time Scaling Strategies

AutoTTS constructs a controllable search environment to enable agents to automatically discover test-time computation allocation strategies. It discovered reasoning strategies that outperform manually designed ones at a cost of only $39.9 and 160 minutes, while achieving generalization across benchmarks and model scales. This framework marks the shift of LLM reasoning optimization from experience-driven to data-driven approaches, providing new ideas for reasoning cost optimization.

Test-Time Scaling (TTS) is an important technology for improving the inference capabilities of large language models, which trades additional computational resources during the inference phase for higher accuracy. However, current mainstream TTS strategies rely on manual design and have limitations: incomplete human understanding of optimal strategies, high cost of manual tuning across different tasks/models, and lack of systematicity making it difficult to ensure optimality.

The core of the AutoTTS framework is to shift the role of researchers from designing strategies to designing the strategy discovery environment (which needs to compress the control space and provide low-cost feedback). Specifically, it formalizes the width-depth TTS problem as a controller synthesis problem, where the controller decides operations such as branch exploration and path continuation. Evaluation does not require repeated calls to LLMs to reduce costs. Additionally, it introduces beta parameterization technology (mapping high-dimensional discrete spaces to low-dimensional continuous spaces) and fine-grained execution trajectory feedback (providing complete trajectory diagnostic information to accelerate iteration).

Experimental verification shows that the strategies discovered by AutoTTS comprehensively outperform manually designed baselines in math reasoning benchmarks—either higher accuracy at the same budget or lower cost at the same accuracy. The strategies have generalization capabilities across tasks (unseen benchmarks) and model scales. The entire discovery process only cost $39.9 and took 160 minutes, demonstrating significant cost-effectiveness.

AutoTTS marks the shift of LLM reasoning optimization from experience-driven to data-driven approaches, establishing a scalable and reproducible strategy discovery process applicable to broader scenarios such as multimodal reasoning. From an industrial perspective, it provides new ideas for cost optimization of LLM inference services, affecting the marginal cost and scalability of AI applications. Meanwhile, interpretable strategies provide materials for understanding the reasoning mechanisms of LLMs.

AutoTTS has limitations: it mainly targets math reasoning, and its effectiveness in open-domain tasks remains to be verified; environment design still requires manual input; costs may still be high in resource-constrained scenarios. Future directions include exploring more efficient search algorithms to reduce costs, expanding to multi-agent collaboration, and studying strategy composability—opening up new possibilities for the self-improvement capabilities of LLMs.