Entropy-Gated Latent Recursion

A new training-free decoding procedure called Entropy-Gated Latent Recursion (EGLR) expands the sampling space for language-model reasoning by adding a deterministic axis to the traditional stochastic temperature sampling, according to a paper posted to arXiv on June 15, 2026 [1]. The method, proposed by Dushyant Singh Chauhan, identifies a second axis for generating diverse reasoning rollouts: the layer span L at which a frozen model's top decoder layers are recursively re-applied at high-uncertainty tokens [1]. Different choices of L produce distinct rollouts that solve different subsets of problems, with no stochasticity involved [2]. Combined with T temperature samples, EGLR turns a single-axis stochastic rollout pool into an L x T Cartesian sampling space at nearly the same per-rollout cost [1]. The paper characterizes this space across 8 instruction-tuned models and 6 math reasoning benchmarks [1]. Large language models, the class of models tested, are trained with self-supervised learning on vast amounts of text to perform natural language processing tasks such as language generation [9]. On the MATH-500 benchmark with Qwen2.5-3B-Instruct, the joint L x T oracle reached 91.6% accuracy, which is 8.2 percentage points beyond the temperature-only oracle at 83.4% and 10.4 points beyond the layer-only oracle at 81.2% [1]. The results confirm that the two axes capture genuinely complementary problems [2]. The expanded rollout pool provides richer per-prompt candidates for downstream procedures that consume rollouts, including self-consistency, best-of-N with verifiers, and group-relative RL training (GRPO) [1]. The work opens a new direction for inference-time scaling that does not rely solely on stochastic noise [2]. The paper was submitted as a 135 KB file to arXiv, an open-access repository of electronic preprints that is not peer-reviewed but is approved for posting after moderation [1][7]. arXiv, which began on August 14, 1991, passed the two-million-article milestone by the end of 2021 and receives about 24,000 submissions per month as of November 2024 [7].