Locus cuts language model output length without hurting preference quality
LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation
Computation and LanguageArtificial IntelligenceMachine Learning
Summary
Large language models often produce longer answers than needed, which increases costs and time to get results. The authors show that by carefully adjusting just a small part of the model using a special low-rank method, they can make the model give shorter replies while keeping the answers just as good. This technique works by changing how the model generates words without changing its original goals. Their method, called LOCUS, was tested on two medium-sized language models and successfully reduced the length of responses by up to 40% without lowering quality.
What this means in practice
- •For ai service providers: Reduce inference costs by trimming model outputs via low-rank post-training adaptations without retraining entire models.
- •For dialogue system developers: Deploy models that generate more concise conversational responses while preserving alignment with user preferences.
Authors
Dongfang Zhao
Abstract
Large language model serving costs scale directly with output sequence length, yet standard preference alignment often inflates response verbosity without improving utility. We study whether the parameterization of post-training updates affects generation length: low-rank subspaces alter sequence length without modifying the alignment loss. We present LOCUS, a method that selects a task-aware low-rank adaptation subspace to minimize output-token cost subject to a utility constraint. Within this subspace, post-training retains the native preference objective with a frozen backbone. Across Anthropic HH-RLHF dialogue preferences, we evaluate two $\sim$3B decoder backbones, Pythia-2.8B and Qwen2.5-3B, against protocol-matched full-parameter DPO and DrDPO branches and the released SamPO checkpoint. LOCUS reduces continuation length by up to 39.84\% on Pythia-2.8B and by 14.87--17.58\% on Qwen2.5-3B while updating only 0.24--0.28\% of model parameters, with no material change in the internal preference diagnostic.