Entropy-Gated Latent Recursion
Researchers have introduced Entropy-Gated Latent Recursion (EGLR), a novel decoding procedure designed to enhance language model reasoning by expanding the sampling space beyond traditional token-level stochasticity. EGLR introduces a deterministic axis by recursively re-applying a model's top decoder layers at high-uncertainty tokens, creating a complementary dimension to temperature sampling. This combined approach, tested on instruction-tuned models and math reasoning benchmarks, significantly improves performance, demonstrating that the layer-span axis captures distinct problem-solving capabilities. AI
IMPACT Introduces a novel decoding strategy that could enhance LLM performance on complex reasoning tasks by expanding the sampling space.