Turn-Level Best-of-N Sampling
An inference-time scaling strategy for LLM agents that generates N candidate actions at each decision turn (rather than at the end of a full trajectory) and has a selector choose one for execution, correcting single-step errors promptly without additional training.
Timeline (last 90 days)
Turn-level Best-of-N sampling is an inference-time scaling method that improves the performance of LLM agents on long-horizon tasks through a generator-selector framework.
Experiments on long-horizon terminal tasks show that turn-level Best-of-N sampling brings accuracy improvements of 5-10%.
Experiments show that turn-level Best-of-N sampling markedly enhances the consistency of repeated attempts.
The core advantage of this method is that it is simple and effective: it requires no additional training and improves performance solely through sampling and selection.
Computational cost rises as N increases, requiring a trade-off between performance and efficiency.
All Facts (5)
Turn-level Best-of-N sampling is an inference-time scaling method that improves the performance of LLM agents on long-horizon tasks through a generator-selector framework.
50%UnverifiedExperiments on long-horizon terminal tasks show that turn-level Best-of-N sampling brings accuracy improvements of 5-10%.
50%UnverifiedExperiments show that turn-level Best-of-N sampling markedly enhances the consistency of repeated attempts.
50%UnverifiedThe core advantage of this method is that it is simple and effective: it requires no additional training and improves performance solely through sampling and selection.
50%UnverifiedComputational cost rises as N increases, requiring a trade-off between performance and efficiency.
50%