Authors: Zeyan Li, Panqi Yang, Qirong Guo, Shengda Zhuo, SIyuan Qiu, Hu Xu, Chun Li, Jianfeng Xu
Organizations: Shanghai Jiao Tong University · Xi’an Jiaotong University · The Hong Kong University of Science and Technology (Guangzhou) · Jinan University
Low-rank adapters (LoRA) make it cheap to fine-tune a large language model once per task, but combining several independently trained adapters into one model remains difficult: merging the updates in weight space causes interference, retraining on all task data is expensive, and routing between separate adapters gives up the goal of a single combined model. We trace the difficulty to two choices that every composition method makes implicitly. A LoRA update admits infinitely many equivalent factorizations; the choice among them is invisible while an adapter serves alone, but it determines what a learned interaction between adapters can see. A coupling between an old skill and a new one can likewise point in either direction, and the direction decides whether the old skills keep computing what they computed before. We introduce READ (Read-only Expansion of Adapter Deltas), which fixes both choices: each adapter is rewritten into a balanced canonical form that preserves its update exactly, and the coupling grows in one direction only, so a new skill can read the input subspaces of old skills but cannot write into their output subspaces. The only trainable object at each append is the new skill's row of the coupling matrix, and the composed update folds into the base weights with no inference cost, routing, or task-specific rules. We evaluate READ across four benchmark suites and two model families, adding skills one at a time. Across several families, READ improves every suite average over the strongest published baselines built from the same adapters---by more than twenty points on SuperGLUE and more than seven points on the domain suite---and nearly all complete addition sequences end above every direct baseline. Factor coordinates and coupling direction, which a lone adapter never exposes, are what decide whether composed skills survive.
Figures & tables
Figure 1: Overview. Naive composition of independently trained LoRA skills leaves two freedoms unmanaged, arbitrary factor coordinates and unrestricted write directions. READ canonicalizes the factors, trains only the new skill’s read row in G , blocks writes into old skills, and folds BstackGAstack into one update.
Figure 2: The READ operator at one append. (a) Every source adapter is rewritten into balanced canonical factors with BicAic=BiAi , so skills meet in coordinates determined by their updates. (b) The stacked factors are joined by a coupling matrix. The old block G(k) stays frozen, the write column G≤k,k+1 is fixed to zero, and only the new skill’s row is trained. (c) The product folds into the base weights W0 , giving one merged projection, no router, and no extra inference path.
Figure 3: Append phase space by suite. Blue marks READ and red marks merge baselines; dashed lines mark pass thresholds.
Figure 5: Append dynamics by suite. Top row: incoming score at each append depth. Bottom row: mean old-bank change. Bands are one standard deviation across lineages; the dashed line marks the −2 -point retention threshold.
Single-document LoRA adapters have been proposed as units of knowledge to be merged and retrieved; yet open-book scoring, with the document in the prompt, can credit a document's own adapter (its specialist) for reading that document as readily as for holding it. Training one adapter for each of 150 Wikipedia articles on Llama-3.1-8B-Instruct, we find that open-book scores mostly measure a reading behaviour every adapter shares: a random adapter trained on another document recovers 82 % of the specialist's open-book advantage over the base model but only 2 % of its advantage on a cloze probe that asks for the same document's facts without providing the source text, a share we cannot distinguish from zero. Content does reach the weights, since the specialists beat the base model in cloze recall by 0.094, and the geometry of their updates follows the documents they read; yet, in a test of 30 queries, adapters chosen by weight-space proximity or text embedding are not distinguishable from random ones on either probe, although that test cannot exclude an advantage of about four fifths of what a query's own adapter gains over a random one. At 16 documents or more, correctly scaled merges recover 87 % to 89 % of the specialists' open-book advantage but only 12 % to 16 % of their cloze advantage, with intervals that run from below zero to at most 32 %, against 94 % for joint training. We also report four silent errors, each of which produced a clean, wrong result.
Martino M. L. Pulici, Cuong Xuan Chu, Evgeny Kharlamov +1
Bosch Center for Artificial Intelligence, Germany · LMU Munich, Germany · Munich Center for Machine Learning, Germany
Low-Rank Adaptation (LoRA) has become a widely adopted technique in text-to-image diffusion models, enabling the personalisation of visual concepts such as characters, styles, and objects. However, existing approaches struggle to effectively compose multiple LoRA adapters, particularly in open-ended settings where the number and nature of required skills are not known in advance. In this work, we present LoRAtorio, a novel train-free framework for multi-LoRA composition that leverages intrinsic model behaviour. Our method is motivated by two key observations: (1) LoRA adapters trained on narrow domains produce denoised outputs that diverge from the base model, and (2) when operating out-of-distribution, LoRA outputs show behaviour closer to the base model than when conditioned in distribution. The balance between these two observations allows for exceptional performance in the single LoRA scenario, which nevertheless deteriorates when multiple LoRAs are loaded. Our method operates in the latent space by dividing it into spatial patches and computing cosine similarity between each patch's predicted noise and that of the base model. These similarities are used to construct a spatially-aware weight matrix, which guides a weighted aggregation of LoRA outputs. To address domain drift, we further propose a modification to classifier-free guidance that incorporates the base model's unconditional score into the composition. We extend this formulation to a dynamic module selection setting, enabling inference-time selection of relevant LoRA adapters from a large pool. LoRAtorio achieves state-of-the-art performance, showing up to a 1.3% improvement in ClipScore and a 72.43% win rate in GPT-4V pairwise evaluations, and generalises effectively to multiple latent diffusion models. Project page: https://nickyfot.github.io/loratorio-skill-composer/
Niki Foteinopoulou, Ignas Budvytis, Stephan Liwicki
Cambridge Research Laboratory, Toshiba Europe · Independent Researcher
Composing independently trained LoRA adapters into a single large language model is useful for multi-domain adaptation, especially when the original training data cannot be shared. A common approach is to use MoE-style routing over LoRA experts, but for frozen pretrained adapters, soft weighted combinations can change the unit-scale additive update under which each LoRA module was originally trained. We propose Hard-Routed MoR-LoRA, a two-stage framework for composing frozen reasoning LoRA experts through unit-scale hard selection. First, domain-specific LoRA adapters are trained independently using reinforcement learning from verifiable feedback to obtain reasoning experts. Then, all experts are frozen, reasoning traces are distilled from them, and only a lightweight shared router together with a small attention LoRA is trained for integration. The router selects exactly one expert per token using hard top-1 routing, while a straight-through estimator enables gradient-based training. Experiments across five benchmarks, multiple model scales, and additional model families show that Hard-Routed MoR-LoRA preserves expert behavior while requiring substantially fewer trainable parameters than soft-routing mixture baselines. Our analysis further shows that normalized soft mixtures often concentrate most routing mass on a single expert, suggesting that hard unit-scale routing provides a simple and efficient abstraction for frozen LoRA expert composition.
Seyed Alireza Molavi, Zhan Su, Yan Hu +3
1Halmstad University, Halmstad, Sweden · 2The Chinese University of Hong Kong, Shenzhen, China