Re-FORC Predicts Reasoning Quality in Real Time to Cut AI Compute Costs

Researchers have introduced Re-FORC, a technique that forecasts the value of extra reasoning steps before an AI model completes them. By predicting expected rewards based on thinking length, the method helps systems decide when to stop reasoning chains early and which model sizes to deploy for each task.

How It Works#

Re-FORC trains a lightweight adapter on top of reasoning models to estimate future performance gains as a function of additional thinking tokens. The approach learns to map from current model states to expected improvement, allowing dynamic decisions about reasoning budgets without running full forward passes.

Key Results#

The method achieves three main benefits:

• Early stopping: Halts unpromising reasoning chains mid-computation, cutting compute consumption by up to 26% while preserving accuracy
• Model selection: Combines smaller models with longer reasoning to outperform a single large model, gaining 1.7 percentage points in peak accuracy while using up to 12% less compute at comparable performance levels
• Adaptive scaling: Improves accuracy by 9.9 percentage points on average at maximum compute, substantially exceeding confidence-based alternatives

Re-FORC also enables cost-aware reasoning by allowing systems to set per-token budgets and predict computation time upfront, turning variable-length thinking into a manageable resource tradeoff.

Why It Matters#

As reasoning-capable models become more computationally expensive, the ability to predict which reasoning investments will pay off—and stop investing when they won't—addresses a core efficiency bottleneck in large-scale AI deployment.