LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling

The subtler claim here is not just that AutoTTS automates strategy discovery, but that it treats the space of test-time computation choices (how many parallel samples to run, how deep to chain reasoning steps) as a search problem that an agent can navigate, which means the framework could generalize across model families rather than being tuned to one architecture.

This is largely disconnected from recent activity in our archive, as we have no prior coverage to anchor it to. It belongs to a cluster of research exploring how to extract more from existing model weights without additional training, a direction that has been gaining traction as pretraining scaling costs rise and labs look for cheaper performance gains at inference time. The AutoTTS work sits at the intersection of that compute-efficiency thread and the broader meta-learning question of whether AI systems can meaningfully improve the procedures used to run AI systems.

Watch whether AutoTTS-discovered strategies transfer to models the framework was not trained on. If the efficiency gains hold on a held-out model family within the next few months of follow-up work, the generalization claim is credible; if results are model-specific, this is a useful but narrow tool.