With the growing availability of open-source task-specific models, model merging has become a practical way to combine multiple models into one. However, existing studies mainly focus on fully fine-tuned models and overlook widely used LoRA modules. We show that full fine-tuning merging methods perform poorly on LoRA due to its larger cross-task magnitude disparities, which distort merged parameters and degrade performance. To address this, we propose DO-Merging (Decouple and Orthogonalize), a data-free framework for LoRA merging. Specifically, we decompose model weights into magnitude and direction components and merge them independently, preventing magnitude imbalances from distorting the merged direction. We further introduce a layer-wise gradient descent method with orthogonal constraints that reduces inter-task interference among direction components without requiring any training data. We provide theoretical analysis supporting both components. Extensive experiments across vision, language, and multi-modal domains show that DO-Merging consistently outperforms existing methods, achieving up to 5% gains on vision tasks, with negligible additional cost. Both parts can be integrated into existing merging methods as plug-and-play enhancements.

A model can know when two events occurred yet still answer which came first incorrectly. Multi-hop QA accuracy conflates such composition errors with missing facts. We introduce a double-gate protocol for temporal factual chains and evaluate a case for composition only if the model (i) answers every constituent fact correctly across paraphrased probes and (ii) answers every aligned sub-question correctly in isolation. If the original multi-hop answer remains incorrect, the case is a residual composition failure; we call the systematic occurrence of such failures composition collapse. On D4v2, a benchmark of 390 temporal chains, two post-trained models differ by only 0.3 percentage points in atomic stability yet by 47.4 points in depth-2 (two-fact) residual failure. In a controlled same-base comparison, outcome-verified reinforcement learning reduces depth-4 residual failure by 54 points relative to reasoning-trace SFT under matched data and LoRA budget. On 80 examples that pass both gates for every GRPO variant, failure decreases monotonically, showing that model-specific gate subsets are unlikely to drive the trend. Finally, chain-of-thought decoding recovers 70–75% of gate-passing failures, consistent with an important role for inference-time computation. These results motivate reporting atomic stability and gate-controlled composition separately when evaluating multi-hop reasoning.

Recently, there has been a growing availability of pre-trained text models on various model repositories. These models greatly reduce the cost of training new models from scratch as they can be fine-tuned for specific tasks or trained on large datasets. However, these datasets may not be publicly accessible due to the privacy, security, or intellectual property issues. In this paper, we aim to develop a lightweight student network that can learn from multiple teacher models without accessing their original training data. Hence, we investigate Data-Free Knowledge Amalgamation (DFKA), a knowledge-transfer task that combines insights from multiple pre-trained teacher models and transfers them effectively to a compact student network. To accomplish this, we propose STRATANET, a modeling framework comprising: (a) a steerable data generator that produces text data tailored to each teacher and (b) an amalgamation module that implements a self-regulative strategy using confidence estimates from the teachers’ different layers to selectively integrate their knowledge and train a versatile student. We evaluate our method on three benchmark text classification datasets with varying labels or domains. Empirically, we demonstrate that the student model learned using our STRATANET outperforms several baselines significantly under data-driven and data-free constraints.