Crowded in B-Space: Calibrating Shared Directions for LoRA Merging

Hugging Face Daily Papers Papers

Summary

This paper introduces Pico, a data-free method that improves LoRA adapter merging by separately calibrating the output-side matrix B to reduce interference from shared directions while preserving task-specific information. Pico achieves 3.4–8.3 point accuracy improvements over existing merging methods across math, coding, finance, and medical benchmarks.

Merging separately trained LoRA adapters is a practical alternative to joint multi-task training, but it often hurts performance. Existing methods usually treat the LoRA update ΔW = BA as a single object and do not distinguish the two LoRA matrices. We show that the main source of LoRA merge interference comes from the output-side matrix B. Across tasks, B repeatedly uses a small set of shared directions, while A remains much more task-specific. As a result, the merged adapter overemphasizes these shared directions, and task-specific information is lost. We propose Pico (Pre-merge interference calibration in output-space), a data-free method that calibrates B before merge by downscaling over-shared directions and then rescaling the merged update. Pico plugs directly into existing merging methods such as Task Arithmetic, TIES, and TSV-M. Across eight different benchmarks from math, coding, finance, and medical domains, Pico improves average accuracy by 3.4-8.3 points over the corresponding base method and achieves the best overall average performance. Pico also enables merged adapters to outperform the LoRA trained with all task data. These results show that LoRA merging works better when the two LoRA matrices are treated separately.
Original Article Export to Word Export to PDF
View Cached Full Text

Cached at: 04/21/26, 07:21 AM

Paper page - Crowded in B-Space: Calibrating Shared Directions for LoRA Merging

Source: https://huggingface.co/papers/2604.16826 Published on Apr 18

·

Submitted byhttps://huggingface.co/yixuantt

yixuanon Apr 21

Abstract

LoRA adapter merging performance can be improved by separately calibrating the output-side matrix B to reduce interference from shared directions while preserving task-specific information.

Merging separately trainedLoRA adaptersis a practical alternative to joint multi-task training, but it often hurts performance. Existing methods usually treat theLoRA updateΔW = BA as a single object and do not distinguish the two LoRA matrices. We show that the main source of LoRA merge interference comes from the output-side matrix B. Across tasks, B repeatedly uses a small set ofshared directions, while A remains much more task-specific. As a result, the merged adapter overemphasizes theseshared directions, andtask-specific informationis lost. We propose Pico (Pre-merge interference calibrationinoutput-space), a data-free method that calibrates B before merge by downscaling over-shared directionsand then rescaling the merged update. Pico plugs directly into existing merging methods such asTask Arithmetic,TIES, andTSV-M. Across eight different benchmarks from math, coding, finance, and medical domains, Pico improves average accuracy by 3.4-8.3 points over the corresponding base method and achieves the best overall average performance. Pico also enablesmerged adaptersto outperform the LoRA trained with all task data. These results show that LoRA merging works better when the two LoRA matrices are treated separately.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2604\.16826

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2604.16826 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2604.16826 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2604.16826 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

MiniCPM-o 4.5: Towards Real-Time Full-Duplex Omni-Modal Interaction

Hugging Face Daily Papers

MiniCPM-o 4.5 is a 9B parameter multimodal model featuring Omni-Flow, a framework enabling real-time full-duplex interaction where the model can simultaneously perceive and respond proactively. It achieves state-of-the-art open-source performance comparable to Gemini 2.5 Flash and runs on edge devices with less than 12GB RAM.

JumpLoRA: Sparse Adapters for Continual Learning in Large Language Models

arXiv cs.CL

JumpLoRA introduces a novel sparse adapter framework for continual learning in LLMs using JumpReLU gating to dynamically isolate task parameters and prevent catastrophic forgetting. The method enhances LoRA-based approaches and outperforms state-of-the-art continual learning methods like ELLA.