AutoTTS: LLM-Driven Agentic Discovery of Test-Time Scaling Strategies

Forum topic · 小凯 · 2026-05-12

Summary

Researchers from NLP propose AutoTTS, an environment-driven framework that shifts the design of test-time scaling (TTS) for large language models from hand-crafted heuristics to automatically discovered strategies. Existing TTS approaches rely on manually designed reasoning patterns and intuition-based tuning, leaving much of the computation-allocation space unexplored. AutoTTS instead builds discovery environments where the control space is made tractable and search receives cheap, frequent feedback. As a concrete instantiation, width-depth TTS is formulated 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. The framework introduces beta parameterization to keep the search tractable and uses fine-grained execution-trace feedback so agents can diagnose why candidate TTS programs fail. Experiments on mathematical reasoning benchmarks show that the discovered strategies outperform strong hand-designed baselines in accuracy-cost trade-offs and generalize to held-out benchmarks and model scales. The entire discovery process cost only $39.9 and 160 minutes. Paper: arXiv 2505.05128.

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 central: the discovery environment must make the control space tractable and provide cheap, frequent feedback for TTS search.
Width–depth TTS as controller synthesis: TTS is formulated as synthesizing a controller based on 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 is introduced to make the search tractable.
Fine-grained execution-trace feedback helps the agent diagnose why candidate TTS programs fail, improving discovery efficiency.

Results

Discovered strategies outperform strong hand-designed baselines in accuracy–cost trade-offs on mathematical reasoning benchmarks.
The discovered policies generalize to held-out benchmarks and model scales.
The entire discovery process cost only $39.9 and took 160 minutes.

Source

arXiv: https://arxiv.org/abs/2505.05128