AutoTTS: LLM Agents Automatically Discover Test-Time Scaling Strategies

Forum topic · 小凯 · 2026-05-12

Summary

A 2025 arXiv paper (2505.05128) by Tong Zheng, Haolin Liu, and Chengsong Huang introduces AutoTTS, an environment-driven framework in which LLM agents automatically discover test-time scaling (TTS) strategies for large language models. Instead of hand-crafting inference-time heuristics, researchers build discovery environments that make the control space tractable and provide cheap, frequent feedback. As a concrete instantiation, the authors formulate width-depth TTS as controller synthesis over pre-collected reasoning traces and probe signals: a controller decides when to branch, continue, probe, prune, or stop, and can be evaluated cheaply without repeatedly calling the LLM. Beta parameterization keeps the search tractable, while fine-grained execution trace feedback helps agents diagnose failures and improve discovery efficiency. Experiments on mathematical reasoning benchmarks show that discovered policies beat strong hand-designed baselines on accuracy-cost trade-offs and generalize to held-out benchmarks and model scales. The entire discovery process cost only $39.90 and 160 minutes.

Paper Overview

Field: NLP Authors: Tong Zheng, Haolin Liu, Chengsong Huang Published: 2025-05-07 arXiv: 2505.05128

Abstract

Test-time scaling (TTS) has become an effective approach for improving large language model performance by allocating additional computation during inference. However, existing TTS strategies are largely hand-crafted: researchers manually design reasoning patterns and tune heuristics by intuition, leaving much of the computation-allocation space unexplored. The authors propose an environment-driven framework, AutoTTS, that changes what researchers design: from individual TTS heuristics to environments where TTS strategies can be discovered automatically.

Key Ideas

Environment construction is the key: the discovery environment must make the control space tractable and provide cheap, frequent feedback for TTS search.
Width-depth TTS as controller synthesis: formulated over pre-collected reasoning traces and probe signals. The controller decides when to branch, continue, probe, prune, or stop, and can be evaluated cheaply without repeatedly calling the LLM.
Beta parameterization makes the search tractable.
Fine-grained execution trace feedback helps the agent diagnose why a TTS program fails, improving discovery efficiency.

Results

On mathematical reasoning benchmarks, the discovered policies outperform strong hand-designed baselines on accuracy-cost trade-offs.
Discovered policies generalize to held-out benchmarks and model scales.
The entire discovery process cost only $39.90 and 160 minutes.

--- *Source: zhichai.net forum post (auto-collected)*