Researchers propose Fork-think with confidence, a method that improves large language model reasoning by identifying forking points using model confidence within a single seeding path before sampling multiple continuations. This decide-first-then-think paradigm contrasts with existing methods that sample paths first and then prune them.
- Reduces token consumption by up to 30% and run-time by up to 57% across three models and three reasoning benchmarks.
- Performs comparable to or better than parallel thinking without requiring re-training, warm-up, or offline training.
- Demonstrates that sampling at later positions leads to substantially better generations.
- Combines with existing mechanisms like early stopping and weighted voting to match state-of-the-art performance.
The approach establishes pre-determined forking as a promising direction for efficient LLM reasoning by avoiding the overgeneration inherent in think-first-then-decide methods.