Overview

Model merging studies how to combine weights or adapters from separately trained models into one model, aiming to reuse complementary skills without storing an ensemble or repeating joint training. Common approaches average checkpoints or add task vectors, while sparsification, low-rank and spectral methods, geometric alignment, and calibration-based routing aim to reduce interference. The methods are used for multi-task and continual learning, domain adaptation, and combining language, vision, speech, and specialized models; evaluations often measure task accuracy, capability retention, worst-task performance, and inference or storage cost. The central tension is that updates that help one task can erase or distort another: success depends on shared initialization, representation compatibility, update scale, and data or prompt distribution, and some task combinations may not be recoverable through weight-space fusion alone.

Latest papers

214 papers18 papers in the last four weeks, down 53% on the four weeks before. 0.1% of all new papers.

Nov 13, 2026

A Spectral Framework for Universal Proxy Regularization in Model Merging

Model merging constructs multi-task models without expensive joint training but often suffers from parameter interference. Some solutions mitigate this by constraining finetuning using data from the other tasks to be merged, violating the core independence and privacy assumption of model merging. Recent work offers preliminary empirical evidence that a universal, task-agnostic proxy distribution can substitute for such data, but leaves open when and why it suffices. To this end, we propose a Spectral Framework for Universal Proxy Regularization, which characterizes the relationship between a proxy distribution and the distribution of other tasks to be merged through spectral coverage (αα) and leakage (ββ). Based on the framework, we derive a unified parameter interference bound: smaller (α,β)(α, β) provably yields lower parameter interference through better spectral alignment. To exploit such a proxy, we augment the training objective with a proxy regularizer. Evaluating this regularizer during fine-tuning requires a concrete mechanism. We study three instantiations, including (J) an exact evaluation via Jacobian-vector products, (F) a frozen curvature approximation using the Fisher information matrix, and (O) an output-drift formulation that constrains the model's functional deviation on the proxy directly. We prove that all three coincide under the NTK linearization, where existing analyses of such regularizers stop, and go further to characterize standard non-linear fine-tuning, bounding how far each instantiation departs from its linearized reference and obtain a three-way hierarchy, zero for (J), proxy-independent for (F), and proxy-quality-scaling for (O). Empirical validation on the commonly used vision-language model CLIP confirms these theoretical predictions.
Nov 12, 2026

Joint Null-Space Projection for Data-Free Model Merging

Merging independently fine-tuned checkpoints into a single multi-task model requires no data, no labels, and no gradient steps. In practice the task vectors of different tasks interfere, and the standard remedies--averaging, magnitude pruning, sign election--operate entry by entry, ignoring that each delta induces a low-dimensional update on the hidden features it acts upon. We recast merging as a subspace interference problem. For every two-dimensional weight tensor we take the top-r right-singular subspace of each task's delta--the input directions along which that task moves the representation most--and remove from each task vector the component lying in the union of the other tasks' subspaces. The resulting Joint Null-Space Projection (JNSP) is closed-form and order-independent, needs no forward pass, and is defined entirely by the checkpoint set. On eight CLIP image-classification tasks JNSP merges ViT-B/16 experts to 89.45% average accuracy without using a single image, 11.88 points above the strongest parameter-wise training-free baseline (TIES-Merging, 77.57%), 14.22 above Task Arithmetic, and 0.98 above WUDI's data-free consensus merge (88.47%). The gain holds across three backbones on the independently trained TALL release (+10.8, +12.6, +9.1 over TIES for ViT-B/16, ViT-B/32, ViT-L/14) and replicates there at 89.20%; controls confirm the source--a single shared subspace instead of per-task ones loses 2.9 points, and a sequential variant collapses to 69.6%. On RoBERTa-large the same projection underperforms plain task arithmetic, which we report rather than hide. Code is available at https://github.com.
Nov 10, 2026

A Margin Perspective on LoRA: Robustness to Catastrophic Forgetting and Adapter Merging (MaLoRA)

Low-rank adaptation (LoRA) is the de facto method for parameter-efficient fine-tuning of large neural networks, achieving strong task adaptation with minimal trainable parameters and low optimization cost. Beyond this efficiency, practitioners have observed two striking properties: robustness to catastrophic forgetting and the ability to merge independently trained adapters into a single model that performs competitively across multiple tasks. These properties are central to continual and multi-task learning, yet remain poorly understood. In this work, we provide a theoretical explanation through a unified margin-based perspective. We analyze LoRA in multiclass linear classification under a near-orthogonal task regime with ℓ2\ell_2-regularization, and characterize the optimal LoRA adapter across regularization regimes. Our analysis shows that, in an intermediate regime of regularization parameters, the optimal adapter aligns with the max-margin solution on the fine-tuning data. Building on this characterization, we derive two key consequences. First, we obtain closed-form expressions for the margins on pre-training and fine-tuning data, revealing a precise margin trade-off: the regularization parameter controls the balance between retention and adaptation. Second, we analyze adapter merging, proving that merged models achieve positive margin on each task and deriving optimal mixing coefficients that balance margins across tasks and maximize the margin over their union. These results lead to a simple, training-free merging rule, which we term margin-based LoRA merging (MaLoRA). Experiments on modern architectures and real datasets validate our theoretical predictions, showing that MaLoRA matches or outperforms several adapter-merging baselines across a range of vision and language classification tasks.
Nov 9, 2026

Domain-Routed Delta Optimization for Multi-Teacher On-Policy Distillation

Multi-teacher on-policy distillation can yield uneven domain capabilities. We propose Domain-Routed Delta Optimization (DRD-MOPD), composing a frozen base with expert-initialized trainable deltas. Domains update their deltas with independent optimizer states and clipping. After joint training, Post-separate exports the base plus one trained delta per domain, selected by known task domain without retraining. Using AdamW, the unweighted full student reaches 32.83% Macro across six benchmarks after 20 updates, versus 32.21% for a matched shared control. Post-separate raises Macro with Open-MOPD weighting.
Nov 9, 2026

Permutation-Based Pre-trained Model Merging via Reciprocal Orthogonality

Ito et al. (2026) showed that sufficiently wide, independently trained neural networks can exhibit linear mode connectivity (LMC) with softmax temperature calibration, even without permutation. LMC means that a low-loss linear path connects the models' parameters. This behavior is explained by layerwise exponentially weighted connectivity (LEWC), whereby each merged layer outputs an exponentially weighted sum of the original models' corresponding representations, yielding ensemble-like predictions. However, prior work observed LEWC with substantial widening, such as 8192 units per hidden layer for MLPs and 16×16\times width for VGG-11, raising the question of whether such widths are necessary. We therefore search for permutations that encourage reciprocal orthogonality (RO), which suppresses interference between each model's weights and the other model's representations. This allows both models' computations to coexist in the merged network, promoting LEWC at smaller widths. We formulate the search as independent layerwise quadratic assignment problems and approximately solve their continuous relaxations using Frank--Wolfe. Experiments show that our method facilitates LMC with less width expansion. It also merges models trained on different data distributions, where weight matching, which maximizes weight similarity through permutations, struggles. The merged predictions approximate those of a logit ensemble of the original models.
Nov 9, 2026

DeltaForge: A Post-hoc Mixture-of-Experts VLA for Multi-Domain Autonomous Driving

Fleets of driving vision-language-action (VLA) models are developed one domain at a time: each operator finetunes a shared base model on private data, producing specialized experts that are costly to consolidate into a single model. Weight merging loses accuracy as heterogeneous domains accumulate, and dense mixture-of-experts routing is prohibitively expensive to combine them. We present DeltaForge, to our knowledge the first post-hoc mixture-of-experts VLA for driving, which consolidates N independently finetuned per-domain models into a single multi-domain model from their weights and a small proxy set alone. When different operators finetune the same base model, their weight updates end up sharing a lot of structure. We exploit this by decomposing all the updates together using singular value decomposition (SVD), splitting each one into a shared part (the same across all domains) and a tiny domain-specific part. The router then only needs to work with those tiny domain-specific pieces, not the full expert weights, which is what keeps the approach efficient. Then two cooperating clip-level blind routers, one over the VLM reasoning experts and one over the flow-matching dynamics experts, are trained from soft labels on the proxy dataset. Across six diverse domains spanning urban, off-road, and highway driving on cars, trucks, and all-terrain vehicles (ATVs), DeltaForge approaches the accuracy of the per-domain experts and surpasses every merging, routing, centralized, and federated baseline at close to single-model parameter and compute cost. Adding a new domain requires only projecting its weight update onto the shared basis with no end-to-end retraining needed. The resulting model also generalizes to domain combinations it was never explicitly trained on.
Nov 8, 2026

When Are Nearby Experts Composable? Cross-Task Model Merging in Neural Thickets

Random perturbations around a shared checkpoint can produce experts for distinct tasks, but shared origin does not ensure that their gains survive composition. We formulate expert-pair selection for joint base-relative retention, measured by strict joint success and worst-task gain with unconditional abstention accounting. An exact finite-difference decomposition and an arbitrarily small orthogonal counterexample show why proximity and orthogonality alone are insufficient. We instantiate this formulation with quality-constrained spectral selection, a fixed-budget framework combining cross-task screening, layerwise pair geometry, separate validation, and Base fallback. Across six checkpoints, 21 task pairs, and five independent perturbation pools, the procedure improves strict joint success over Top-1 averaging throughout and leads the four-method model-scale comparison on four of six checkpoints. For the five models with at least 1.5B parameters, matched quality-only, global-cosine, and spectral ranking attain 49.1%, 50.5%, and 53.1% joint success. Under this matched protocol, the spectral selector also yields the highest worst-task gain and Weak Pareto rate and the lowest regression rate. Mean worst-task gain reaches +0.61 percentage points, while 17.1% regression and 13.0% abstention show that composability remains conditional. Project page: https://anonymous.4open.science/r/Thickets-Merge/.
Nov 7, 2026

Robust and Efficient Continual Model Merging via Global Singular Subspace Separation and Restoration

Continual model merging incrementally integrates task-specific models without revisiting past models or data. However, repeated merging entangles past task components and distorts their singular geometry, making balanced task-wise allocation increasingly difficult. We propose Global Singular Subspace Separation and Restoration (GS3R), which augments the merged update with a compact restoration state. At each merge, GS3R restores the retained singular geometry, separates task-wise subspaces, and globally reallocates the available rank across tasks fairly. To reduce merging-time memory overhead, we further introduce Pre-merge Restoration and Flushing (PRF). Experiments on vision, NLP, and multimodal benchmarks show that GS3R provides a strong accuracy-efficiency trade-off.
Nov 7, 2026

Enhancing Mathematical Reasoning in Turkish Large Language Models with Search-Based Inference and Model Merging

We improve Turkish mathematical reasoning in Large Language Models by combining LoRA fine-tuning, weight-space model merging, and search-based inference. Using Turkish-Llama 8B-DPO with Turkish translations of MetaMath, NuminaMath, and TigerLab, we find data quality beats scale (Mix-30k best among fine-tuned; large MetaMath underperforms). Linear merging delivers the largest gains (GSM8K up to 0.63), while Ties is modest and Breadcrumbs/Slerp degrade. Search with MCTS and Tree of Thoughts improves mean LLM-judge scores for reasoning quality on a 0-100 scale. Limitations include machine translation, limited scale, and search stochasticity.
Nov 4, 2026

FedFiOT: Fisher-Weighted Optimal Transport for Model Merging in One-Shot and Few-Round Federated Learning

One-shot and few-round federated learning (FL) reduce communication burden by constructing a global model within one or a few communication rounds, but often achieve lower performance than conventional multi-round FL. Without frequent synchronization, client models may drift apart, causing neurons that learn similar features to occupy different positions across clients. At the same time, local data heterogeneity can cause corresponding parameters to differ in their importance to each client's predictions. Together, these challenges cause direct averaging to combine mismatched neurons and assign equal weight to parameters despite differences in their importance. We propose FedFiOT, a Fisher-weighted optimal transport framework for one-shot and few-round FL that leverages Fisher information to quantify parameter importance and guide neuron alignment and model aggregation. FedFiOT jointly optimizes neuron correspondences and global model parameters through alternating minimization of a shared Fisher-weighted objective. Theoretically, we show that FedFiOT's optimization procedure monotonically decreases the layer-wise merging objective and establish a stationarity bound that quantifies the effects of client heterogeneity and nonuniform Fisher weighting.Experiments across multiple datasets demonstrate consistent improvements over competing baselines in both one-shot and few-round federated learning settings.
Nov 4, 2026

Model Merging Does Not Require Uniform LoRA Ranks

Model merging combines separately fine-tuned experts into a single model without retraining. Standard practice trains every expert at the same LoRA rank, which couples the rank choice across experts even though each domain has its own best rank. We show this coupling is a property of the merge method, not of merging. We train 36 experts on three domains (mathematics, instruction following, multilingual) at four ranks and three model scales, then merge with six methods at uniform ranks and at each domain's own best rank. Giving experts different ranks costs averaging and utility-based selection nothing, and gains them up to 1.8 pp of multi-domain accuracy, while the conflict-resolution methods are inconsistent and can lose up to 4 pp. The split has a simple cause. Merging keeps only a fixed share of the pooled spectral components, and methods that choose them by size discard many that would have reduced loss. They also tend to discard the same expert: the instruction adapter has the smallest singular values, so it is dropped almost entirely and IFEval falls by 16 to 25 pp. Choosing components by their effect on loss instead keeps all three experts, and leaves accuracy nearly flat across an 8×\times rank range (0.7 pp, against 6.7 pp for TIES at 4B scale). Rank can therefore be chosen per expert whenever the merge operator does not select by magnitude.
Nov 3, 2026

Modulating Merging Strengths via Joint Loss Estimation for LoRA-based Continual Learning

Continual learning (CL) aims to adapt models to new tasks while preserving previously learned knowledge. Recently, Low-Rank Adaptation (LoRA), a representative Parameter-Efficient Fine-Tuning (PEFT) method, has gained increasing attention in CL due to its efficiency and scalability. Several LoRA-based merging methods maintain a single inference model by scaling the newly learned LoRA update with one coefficient and merging it into the previous weights after each task. However, they usually control the whole update with a shared scalar factor, although different update elements may affect previous and current tasks in different ways. To address this issue, we propose Merging Strength Modulation via Joint Loss Estimation (M2LE). It consists of Hessian Information-guided Adaptive Merging (HIM) and Task Simulation Perturbation for Merging (TSP). Instead of searching for a global merge coefficient, HIM directly solves for the target merged update under a joint objective over previous and current tasks. Solving this objective gives an analytic solution that naturally takes the form of element-wise modulation over the current task LoRA update. To improve the stability of the learned update under subsequent merging, we introduce TSP which simulates different tasks interference. Experiments on multiple benchmarks show that M2LE consistently outperforms existing PEFT-based continual learning methods. Code is available at https://github.com/bbdeeer/M2LE_NeurIPS2026.
Nov 3, 2026

Separating Common and Unique Directions for Model Merging

Model merging combines multiple fine-tuned models without additional training, and arithmetic methods have emerged as a simple yet effective way to merge models without data. However, existing arithmetic methods treat task vectors in parameter space and apply a single element-wise rule to directions that can differ in mergeability. In a shared directional basis, we find a consistent two-regime structure: a low-dispersion common bulk shared across tasks and a high-dispersion task-dominated tail that carries disproportionate energy. These findings motivate Common--Unique Subspace Decomposition (CUSD), a data-free pre-merging method that separates task vectors into common and unique directional components, merges the common components with a downstream arithmetic operator, and retains unique residuals separately. Experiments across 1111 base models, four model families, and four arithmetic-based merging operators show that CUSD improves 42 of 44 model--operator combinations and achieves up to 22.5% improvement over the corresponding baselines. Our code is publicly available at https://github.com/stwang666/CUSD
Nov 2, 2026

P2^2-MOPD: Accelerating Multi-Teacher On-Policy Distillation via Parameter Priors

Multi-teacher on-policy distillation consolidates expert capabilities by supervising student-generated trajectories, whereas task arithmetic offers a training-free alternative through linear parameter combinations. We view the linear combination of student and expert parameters as a single differentiable model and propose P2^2-MOPD, which jointly optimizes the mixing coefficients and the student weights under the same on-policy distillation objective. Unlike task arithmetic, which stays within the span of the frozen student and expert parameters for any choice of merging coefficients, P2^2-MOPD keeps the student trainable, so the experts act as parameter priors without confining it to this span. Across four domains at 1.7B and 4B scales, P2^2-MOPD achieves faster and more balanced recovery than MOPD from the same initialization, surpassing its final average recovery in half the training steps. At the end of training, our 1.7B model reaches 94.7% average recovery, versus 80.9% for MOPD.
Nov 2, 2026

Low-Rank Hierarchical Merging for Efficient Long-to-Short Reasoning

Long-to-Short (L2S) model merging seeks to combine the strong reasoning capability of long chain-of-thought models with the concise style of base models, but existing methods that rely on linear interpolation and first-order activation statistics fail to capture how the loss landscape responds to parameter displacement, leading to degraded accuracy on hard tasks. We propose Low-Rank Hierarchical Merging (LHM), a training-free framework that replaces heuristic statistics with lightweight second-order curvature information. LHM consists of three components: (1) a low-rank Hessian approximation that efficiently estimates per-layer Hessian traces via Hutchinson's estimator on random inputs, eliminating the need for task-specific data; (2) a Layer Second-Order Sensitivity (LSOS) index that combines Hessian trace with task-vector norm to quantify per-layer fusion difficulty; and (3) a Dual-Constraint Coefficient Mapping (DCCM) that converts LSOS scores into layer-adaptive merge coefficients while jointly satisfying a fusion-error bound and a global length-reduction target. Experiments across six mathematical benchmarks on 1.5B and 7B models show that LHM outperforms state-of-the-art methods such as ACM-TIES and ORCA, achieving over 4% accuracy gain and up to 93.1% reduction in output length relative to the long-CoT model.
Oct 30, 2026

ResMerge: Residual Policy Learning and Merging for Continual Adaptation of Pre-trained Robot Policies

Continual adaptation to new tasks is essential for robots deployed in the real world, but adapting pre-trained policies remains challenging due to parameter growth and the need for continual demonstration collection. Parameter-Efficient Continual Fine-Tuning (PECFT) isolates task-specific updates, but adapter-per-task designs require demonstrations and grow with the task set. We propose ResMerge, a two-stage framework that learns mergeable residual policies atop a frozen pre-trained policy: residual reinforcement learning (RL) acquires task-specific corrections without additional expert demonstrations during adaptation, and parameter-space merging consolidates compatible residuals to reduce stored parameters. We introduce Residual-Field Consistency (RFC), a simple, bounded, and interpretable metric that assesses compatibility in residual action-field space by comparing correction direction and magnitude. Experiments on LIBERO across two backbones and on a real-world Franka robot show improved continual learning performance over strong PECFT baselines, while selective merging reduces the residual-parameter footprint. Project website: https://resmerge.github.io/.
Oct 27, 2026

Dynamic Merging Mutually-Distilled Multi-View Networks for RGB-Event-based Representation Learning

Multi-modal fusion algorithms combining conventional frame-based cameras and event cameras have recently attracted substantial research attention. Typically, existing approaches convert event streams into frames, voxels, or tensors, and then learn the required spatiotemporal information from these transformed event representations for fusion with conventional frame-camera data. Obviously, such paradigms only exploit the limited representational capacity of event data. Moreover, standard training pipelines fail to fully leverage diverse model parameters obtained throughout the learning process. In this paper, we propose a novel mutual distillation method leveraging multi-view teacher network weights to better integrate multiple event representations and achieve model-parameter level fusion. Specifically, we employ multiple encoders to extract different types of event representations and perform mutual knowledge distillation at both the feature-level and logit-level in this process. Afterwards, we perform a dynamic hierarchical model merge strategy on the model parameters from the first stage to yield a unified multi-modal RGB-Event representation model. Experimental results across multiple event-based datasets demonstrate that our proposed framework achieves significant performance improvements.
Oct 26, 2026

Preserve to Expand: Source-Aware Adaptation and Model Merging for Multilingual LLM-Based ASR

Expanding multilingual ASR systems to new languages is challenging when adaptation must preserve previously supported languages without replaying source-language training data. We study this problem in connector-based Speech LLMs, where a trainable connector maps speech representations into a frozen large language model. We evaluate target-update construction and merging strategies for adapting the connector while preserving source-language performance. In single-target expansion, target-only fine-tuning and standard TIES merging learn new languages but cause severe source-language forgetting, while SSU-Wanda+TIES gives the best source-preserving trade-off. In multi-target expansion, update merging provides stable source retention, whereas recursive expansion can improve target adaptation under favorable orders but is highly order-sensitive. These results show that multilingual ASR expansion requires balancing target acquisition, source retention, and interactions among target-language updates.
Oct 24, 2026

Error-Subspace Calibration for Multi-Task Model Merging

Model merging combines task-specific expert models into a single model without joint training on their original datasets, but often incurs a performance gap relative to the individual experts. Post-merging calibration can reduce this gap through task-specific representation alignment, but the alignment objective alone does not specify which feature directions to correct. We propose Error-Subspace Calibration (ESCal), a post-merging calibration framework that identifies low-dimensional error subspaces from weight differences between expert and merged models. At each intermediate block boundary, a bridge module projects the merged representation onto the next block's input-projection error subspace and maps these coordinates to an additive correction in the current block's output-projection error subspace. The subspace bases remain fixed, so calibration learns only these linear maps, yielding a parameter-efficient mechanism for feature correction. We find that ESCal brings the merged model closer to the experts in both representations and predictions, indicating more faithful recovery of task-specific behavior. Experiments on vision and natural language understanding benchmarks show that ESCal improves merged-model performance across diverse architectures and merging methods using a small calibration set for each task.
Oct 22, 2026

HimmBO: High-Dimensional Interactive Model-Merging Bayesian Optimization for Image Generation

Fine-tuning has become a common way to customize text-to-image models, leading to community collections of over 100K lightweight adapters. Adapters sharing a base model can be merged with continuous coefficients. This turns a large collection into a high-dimensional continuous design space, which recent work explores interactively using preferential Bayesian optimization (PBO). However, existing approaches are limited to a few dozen preselected adapters and require expensive image generation in every interaction round. We introduce HimmBO, a two-stage method that scales preference-driven adapter merging to much larger collections while reducing image-generation cost. Its screening stage recovers the active adapter set through level-set estimation over pre-generated images, requiring no image generation. The refinement stage then optimizes merge coefficients in the recovered low-dimensional subspace. Both stages use a style-aware kernel, allowing preference information to transfer between visually similar adapters. In simulated-user image-matching experiments, HimmBO achieves higher final similarity than all comparisons while reaching a given similarity with under one tenth of the image generations required by prior work. Its performance remains stable as the collection grows to 1080 adapters, while competing methods degrade substantially.
Oct 22, 2026

LoRA Superposition across Pretrained Models

Low-rank adaptation (LoRA) is usually tied to the pretrained backbone on which it is learned. We ask whether LoRAs trained independently for the same task on different pretrained backbones can be summed and applied unchanged to all contributing backbones. We call this construction LoRA superposition. Across seven ViT-B/16 CLIP backbones and eight image-classification tasks, the exact sum improves all 56 backbone-task pairs and achieves a mean accuracy gain of 30.21 percentage points over the unadapted backbone. The exact sum achieves a mean accuracy only 0.28 percentage points below that of individually trained solo LoRAs. In contrast, a leave-one-model-out (LOMO) sum produces a mean gain of only 0.06 percentage points over the unadapted backbones. Backbone-specific solo LoRAs are nearly orthogonal in parameter space, while pairwise cross-backbone responses are strongly concentrated on the diagonal. These results support a component-wise non-interference hypothesis, under which the frozen backbone acts as an implicit selector of its own component without explicit routing or cross-model alignment.
Oct 21, 2026

GRAM: Group-wise Rank-Aware Model Merging via Subspace Alignment

Merging different models into a single model is highly desirable for multi-task deployment, particularly in regulated domains like healthcare. In such settings, models are often locally fine-tuned due to strict data privacy policies, necessitating that aggregation operates exclusively on the fine-tuned weights. Existing methods apply a uniform merge across all tasks, ignoring the geometric relationships among their learned subspaces. When task updates occupy misaligned subspaces, projecting them onto a shared low-rank basis discards task-relevant directions, a failure mode named subspace interference. Through this study, we show that the per-task projection error of a merged model is governed by the principal-angle alignment between each task and the shared subspace, so that grouping geometrically compatible tasks provably reduces this error within each group. Motivated by this observation, we propose Group-wise Rank-Aware Model Merging (GRAM), which uses the principal-angle geometry of task vectors to decide which tasks should be merged together and then performs a standard SVD-based merge within each coalition; the pipeline requires only the fine-tuned weights at merge time. With only one additional model (\nK=2\nK{=}2), GRAM significantly mitigates subspace interference and consistently outperforms prior merging methods across language understanding (GLUE with Flan-T5), 8-task CLIP vision, and medical instruction tuning (Qwen2-7B-Instruct with four English/Chinese medical LoRAs).
Oct 15, 2026

Training-Free On-Policy Distillation

On-policy distillation consolidates specialist models into one student at the price of rollouts, teacher passes and gradient steps for every teacher. We show that the consolidation does not need training: with forward passes only, the capabilities of separately tuned specialists can be aggregated into one model that is, on average, more accurate than either specialist alone. When specialists share a base model, distillation has an exact layer-wise solution: each linear layer of the fused model must reproduce every specialist's output on the inputs that specialist receives while reading its own answers, a least-squares problem whose solution needs only the second moments of those inputs. One forward pass over a few hundred rollouts provides them. We call this training-free on-policy distillation. What the rollouts contain matters more than the solver: answers truncated in repetition loops are few among a mathematics specialist's rollouts but carry a large share of their tokens, and removing them lifts the fused model on AIME25 from task arithmetic's level to the specialist's at matched precision. On Olmo-3-7B, a few GPU-hours of forward computation per specialist yield a generalist that keeps most of each specialist's gain on each of seven mathematics and code benchmarks.
Oct 15, 2026

RecSoup: Efficient Model Merging Approach for Controllable Multi-Objective Recommendation

Recommender systems often need to balance multiple conflicting objectives. Controllable Multi-Objective Recommendation (CMOR) enables dynamic adjustment of objective trade-offs at inference time without retraining, adapting to shifting user or system requirements. Existing CMOR approaches typically rely on output-level reranking or parameter-generating modules, which either limit the scope of feasible trade-offs or incur substantial training and inference cost. Model merging provides a promising alternative by reusing multiple objective-specific models via parameter combination. However, its direct application to recommender systems is challenging due to the lack of a shared linear regime and cross-objective interference. To address these issues, we propose RecSoup, a lightweight two-stage approach for effective CMOR via specialized model merging. RecSoup first mitigates parameter interference across objectives through a pruning strategy that jointly considers objective-specific importance and cross-objective relevance. It then performs adaptive parameter merging via a learnable preference-to-coefficient mapping, allowing dynamic trade-off control at inference time without joint retraining or additional generative modules. Theoretically, we prove that RecSoup outperforms the parameter generation paradigm in effectiveness and efficiency, as it can fully navigate the Pareto front without smoothness constraints. Extensive experiments on multiple benchmarks demonstrate that RecSoup achieves improved Pareto front quality and effective controllability while delivers 165--228x faster inference-time latency than parameter generation approaches. Code is available at https://github.com/35798642/RECSOUP.git.
Oct 11, 2026

SuperAI: Merging Capabilities Across Heterogeneous Foundation Models

Enabling the reuse and composition of knowledge and capabilities distributed across foundation models is an important step toward building scalable artificial intelligence (AI) systems. However, architectural differences and incompatible representations make it difficult to transfer and integrate these capabilities through a unified cross-model interface. To address this challenge, we propose SuperAI, a cross-architecture multi-model fusion framework that integrates complementary capabilities from multiple source models into a host model through orthogonal alignment of low-dimensional task-relevant representations. Specifically, we first extract task-relevant representations from the source models and the host model using shared anchor samples, and apply principal component analysis (PCA) to construct low-dimensional functional subspaces with the same dimensionality. Next, we use orthogonal Procrustes mappings to align the source and host coordinate systems, establishing cross-model correspondences while preserving the internal geometry of each source subspace. Finally, we train a lightweight residual multilayer perceptron (MLP) with joint supervision on task outputs and source-model increments, learning input-dependent residual increments that are injected into the host model for multi-model capability fusion. Experiments on three host models across 15 capability domains show that SuperAI achieves higher average domain scores than the evaluated cross-architecture fusion methods and better preserves general capabilities as the number of source models increases. Ablation studies further support the effectiveness of orthogonal alignment, nonlinear residual mapping, and joint supervision.
Oct 10, 2026

CrossMean: Model Merging by Pairing Merged and Expert Activations

Model merging combines fine-tuned experts into one network without retraining, and closed-form methods compute each merged layer from statistics of its input activations. However, these statistics are measured on a single expert, while in the merged model, each layer receives activations from the merged layers that precede it. Recent work re-measures these statistics in the merged model but still forms them from a single network. We formulate closed-form methods as a linear system and show that they use the same matrix on both sides. We then introduce CrossMean, which replaces this matrix on one side with a bilinear statistic that pairs merged and expert activations. From the same statistics, we also derive a label-free criterion that indicates the task sets in which accounting for the merged activations improves the merge more. Empirically, task sets with a lower criterion value show a larger improvement. On the eight-task benchmark of each vision backbone that we tested and on BERT encoders, CrossMean improves on merging with statistics measured on the experts, and with all linear layers merged, it also improves on re-measuring these statistics in the merged model. These results suggest that merging benefits from pairing merged and expert activations.
Oct 9, 2026

UniMergeTrack: Hierarchical Model Merging for General RGB-X Visual Object Tracking

RGB-X tracking leverages complementary modalities, such as thermal infrared, event, or depth, to improve robustness under challenging conditions. Existing general-purpose RGB-X trackers typically follow two paradigms: independent task-specific training and joint training across modality combinations. However, the former suffers from knowledge isolation and model redundancy, while the latter introduces optimization interference across heterogeneous tasks. In this work, we present UniMergeTrack, a new model merging paradigm that preserves task-specific optimization while enabling effective cross-task knowledge integration. UniMergeTrack first trains modality-specific trackers independently and then performs hierarchical model merging. Specifically, transferable backbones are merged across RGB-X tasks via high-pass-filtered merging, and complemented by lightweight task-specific experts that preserve modality-specific cues. Tracking heads are merged within each task using optimal transport-guided channel alignment to mitigate hidden channel mismatch. A subsequent distribution alignment fine-tuning stage further improves compatibility between the unified backbone and task-specific heads. Extensive experiments on the LasHeR, RGBT234, VisEvent, VOT-RGBD22, COESOT, and CDTB benchmarks demonstrate that UniMergeTrack achieves state-of-the-art or competitive performance across RGB-Thermal, RGB-Event, and RGB-Depth tracking tasks, while exhibiting strong generalization to unseen datasets. The source code will be released upon acceptance. The source code of this paper will be released on https://github.com/Event-AHU/OpenEvTracking.
Oct 7, 2026

CHASE: Channel-Aligned Structure Exploitation for Geometry-Aware Model Engineering

Geometric and Spectral Alignment (GSA) characterizes trained networks through spectral concentration, physical-channel alignment, support structure, and changes in singular bases. In this paper, we propose CHASE (Channel-Aligned Structure Exploitation) to use these structures in practical model design. CHASE covers six applications across model modification, reconfiguration, and compression. CORA, COEC, and CORAM apply GSA to parameter-efficient finetuning, structured-pruning compensation, and model merging. We further develop three new methods. CAGA uses GSA to identify multi-head attention heads that can share a KV representation and constructs the shared key and value heads through geometric alignment and low-rank subspace extraction. SAKV uses GSA to determine which adjacent layers can share a low-rank KV-cache representation and the retained rank for each layer group. CAPS uses GSA spectral structure to group output neurons and selects retained input channels separately for each group. Results from CORA, COEC, and CORAM establish the effectiveness of GSA for adaptation, pruning compensation, and model merging. Experiments on CAGA show that geometric shared-head construction substantially improves MHA-to-GQA conversion, and SAKV and CAPS improve over representative baselines for KV-cache compression and structured pruning. These results show that the structures identified by GSA can be used directly to design methods for a range of model operations.
Oct 6, 2026

InterCorrect: Intersection-Aware Correction of Demographic Model Merging for Fair ASR

Automatic Speech Recognition (ASR) systems often show uneven performance across demographic groups, and errors can be especially difficult to address for speakers belonging to multiple demographic groups. This work studies demographic-aware model merging for fair Speech-LLM-based ASR. Starting from a SLAM-ASR-based model, we fine-tune only the connector on demographic-specific subsets and merge the resulting subgroup-adapted connectors into a global model. We then identify critical cross-axis demographic pairs using subgroup WER and task-vector conflict, and apply intersection-specific correction vectors to the global merged model. Experiments on Fair-Speech show that global demographic merging improves overall WER over the base model, while intersection correction provides additional gains for several merging strategies. In particular, TIES with WER-based correction achieves the best overall WER, reducing it from 7.38% to 5.13%. Subgroup and disparity analyses further show that the proposed approach improves performance across demographic axes, while highlighting that lower average WER does not always imply reduced subgroup disparity.
Oct 6, 2026

Enhancing Diffusion Language Models with Autoregressive Post-Training Weights

Diffusion language models (dLLMs) have emerged as a promising alternative to autoregressive (AR) language models, offering flexible token-update orders and parallel decoding. Recent dLLMs are often initialized from pretrained AR models before diffusion conversion in order to inherit their learned representations. After the conversion, however, they typically ignore the extensive post-training ecosystem of their AR ancestors. In this work, we show that these existing AR post-training weight updates can instead be effectively recycled to enhance diffusion models. Despite the changes by AR-to-diffusion conversion, directly adding an AR post-training weight update to a diffusion base model remains effective, bringing its performance close to that achieved by direct diffusion post-training. Notably, AR and diffusion post-training updates are nearly orthogonal in weight space, yet induce substantially more aligned representation changes in the diffusion model. Their distinct updates are also complementary: composing their weights can retain gains from both regimes and further improve the post-trained diffusion model. Based on these findings, we propose A2D, a simple training-free framework for enhancing diffusion models with existing AR post-training resources. A2D can transfer capabilities from AR post-trained models to diffusion base models, and further improve already post-trained diffusion models by composing AR and diffusion post-training updates. Across various dLLMs, including Dream, DreamReasoner, DiffuCoder, Dream-Coder, Nemotron-Labs-Diffusion, and DiffusionGemma, A2D reliably improves instruction following, mathematical reasoning, and coding with both supervised fine-tuning and reinforcement learning updates, without additional training, or inference-time computation.