Search papers, labs, and topics across Lattice.
The paper introduces Evo-L2S, a framework for Long-to-Short (L2S) reasoning that uses multi-objective evolutionary model merging to optimize the trade-off between accuracy and output length. To improve tractability, they use an entropy-based subset sampling technique to reduce the cost of fitness estimation during the evolutionary search. Experiments on mathematical reasoning benchmarks show Evo-L2S can reduce reasoning trace length by over 50% while maintaining or improving accuracy across 1.5B, 7B, and 14B parameter models.
Achieve 50% shorter reasoning chains without sacrificing accuracy by merging models using multi-objective evolutionary optimization.
Reasoning models have demonstrated remarkable capabilities in solving complex problems by leveraging long chains of thought. However, this more deliberate reasoning comes with substantial computational overhead at inference time. The Long-to-Short (L2S) reasoning problem seeks to maintain high accuracy using fewer tokens, but current training-free model merging approaches rely on scalarized, fixed-hyperparameter arithmetic methods that are highly brittle and force suboptimal compromises. To address this gap, we introduce Evo-L2S, a novel framework that formulates L2S reasoning as a multi-objective optimization challenge. By leveraging evolutionary model merging, Evo-L2S explicitly optimizes the trade-off between accuracy and output length to produce a robust Pareto front of merged models. To make this search computationally tractable for large language models, we propose an entropy-based subset sampling technique that drastically reduces the overhead of fitness estimation. Comprehensive experiments across 1.5B, 7B, and 14B parameter scales on six mathematical reasoning benchmarks demonstrate that Evo-L2S can reduce the length of generated reasoning traces by over 50% while preserving, or even improving, the problem-solving accuracy of the original reasoning models.
