Large language models (LLMs) now have a new tool to cut costs without sacrificing accuracy. Researchers from Meta, Google, and multiple universities developed AutoTTS, a framework that automatically finds optimal ways to allocate compute during inference. This system reduced token consumption by 69.5% in tests while maintaining performance on math and reasoning benchmarks.
Test-time scaling (TTS) lets LLMs use extra compute to improve answers by exploring multiple reasoning paths. Previously, engineers had to manually set rules for how models should branch, prune, or stop reasoning. This process relied on guesswork and limited exploration of possible strategies.
Related: MiniMax M3 model boosts response speed with sparse attention
AutoTTS changes this by treating TTS strategy design as an algorithmic search problem. Instead of humans coding rules, the framework lets an AI “explorer” LLM generate and test controllers—code-defined policies that allocate compute. These controllers are refined using pre-collected reasoning trajectories from base models, avoiding the need to run expensive real-time experiments.
The explorer agent evaluates proposed controllers by analyzing how they allocate compute over time. It identifies failure modes, like over-pruning branches, and iteratively improves the code. This process uncovers complex rules that humans might never think to implement manually.
One key discovery is the Confidence Momentum Controller. Unlike traditional methods that stop reasoning when hitting a confidence threshold, this controller tracks an exponential moving average of confidence levels. It only stops when confidence is high and stable, avoiding errors from temporary spikes.
Related: Why Software is Essential for Success
AutoTTS also links width and depth control. If current reasoning paths stall, the system automatically spawns new branches. It prioritizes computation for branches that align with the leading answer, focusing resources where they matter most.
Experiments tested AutoTTS on Qwen3 models with 0.6B to 8B parameters and a distilled DeepSeek-R1 model. The framework outperformed existing methods like Self-Consistency (SC@64) and Adaptive-Consistency (ASC). In cost-conscious mode, it slashed token use by 69.5% while keeping accuracy steady across four Qwen models.
On the GPQA-Diamond benchmark, AutoTTS cut inference tokens from 510K to 151K while slightly improving accuracy. For the HMMT25 benchmark, it achieved the highest accuracy on DeepSeek models with nearly half the token cost.
Related: Top 10 Tips for Navigating Technology News in 2026
For enterprise teams, AutoTTS offers two main benefits. First, it boosts peak performance by redirecting compute toward productive reasoning branches. Second, the discovery process costs just $39.90 and takes 160 minutes, making tailored optimizations feasible without dedicated research budgets.
The AutoTTS framework and Confidence Momentum Controller are open-source on GitHub. The CMC can replace existing TTS controllers directly. Researchers say this marks a shift in how LLMs manage compute, enabling more efficient and effective reasoning at scale.
