CoSMo Framework Improves Reasoning Efficiency in Large Language Models Through Split-Merge Optimization
Researchers have developed CoSMo, a framework that optimizes how large reasoning models generate solutions by eliminating redundant reasoning steps rather than simply limiting output length. The approach uses a split-merge algorithm combined with reinforcement learning to refine reasoning chains while maintaining coherence. This work addresses a key challenge in deploying reasoning models: reducing computational overhead and latency without sacrificing accuracy.
CoSMo (Consistency-Guided Split-Merge Optimization) is a new framework designed to make large reasoning models more efficient by targeting structural redundancy in their reasoning chains. Rather than indiscriminately restricting the number of tokens generated, the method dynamically refines reasoning by merging redundant segments and splitting logical gaps to ensure the model's reasoning remains coherent. The framework employs structure-aligned reinforcement learning with segment-level budgeting to train models to maintain efficient reasoning structures. Experiments across multiple benchmarks show CoSMo achieves a 3.3-point accuracy improvement while reducing segment usage by 28.7% on average compared to existing reasoning efficiency baselines. The work addresses a practical problem in deploying large reasoning models: their verbose generation creates significant latency and computational costs.
What different sources said
- arXiv cs.CLCenter
Short Chains, Deep Thoughts: Balancing Reasoning Efficiency and Intra-Segment Capability via Split-Merge Optimization
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.