LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

Tong Zheng, Haolin Liu, Chengsong Huang, Huiwen Bao, Sheng Zhang et al.|May 8, 2026arXiv

Key Takeaway

You can automatically discover better inference strategies for LLMs by treating it as a search problem over execution traces, rather than manually designing heuristics—and it's cheap to do at scale.

Summary

This paper presents AutoTTS, a framework that automatically discovers test-time scaling strategies for LLMs instead of relying on hand-crafted heuristics.

reasoning

Key Terms

test-time-scaling reasoning-trajectory controller-synthesis execution-trace-feedback