Paper

Entropy-Guided Loop: Achieving Reasoning through Uncertainty-Aware Generation

Authors: Andrew G. A. Correa, Ana C. H de Matos

A test-time refinement technique that leverages token-level uncertainty to enhance model reasoning. The approach extracts probability information, computes Shannon entropy across top alternatives, and applies a trigger mechanism based on perplexity and confidence metrics — feeding uncertainty data back to the model for targeted corrections.

Smaller models equipped with this loop achieve ~95% of larger reasoning models' performance at ~1/3 the computational cost. Selective refinement activates on ~31% of responses, yielding a 16 percentage point accuracy improvement over standard single-pass inference.

View on arXiv

Full Paper

---

Architecture

---

Thread