Tag
The paper presents REST, a novel training objective for latent recursive LLM systems that enhances accuracy by up to 7.5 percentage points across benchmarks by incorporating properties like causality and minimality into differentiable losses.
A new method, FLORA (Frequency-corrected Learning of Ordered Rank Alignment), improves LLMs by aligning generator and validator modes using a principled frequency correction. Experiments show substantial gains in G-V consistency and generator performance on benchmarks like IFEval and HumanEval.