Self-Supervised Reasoning via Recursive Latent Tree Search in Large Language Models

Abstract

While chain-of-thought prompting improves complex task solving in large language models, linear generation trajectories frequently succumb to error propagation. We introduce Recursive Latent Tree Search (RLTS), a framework that executes non-linear exploratory reasoning directly in continuous latent activation space without text generation overhead. Across MATH and Codeforces benchmarks, RLTS outperforms test-time compute baselines by 14.8%.

Keywords

Download Complete Research Article

The full manuscript includes introduction, methodology, results, discussion, conclusion, figures, tables, and complete bibliography with all 1 references.