DKL: Decoupled Knowledge Learning for Instruction-Tuned Language Models

RAG has become the de facto method for incorporating new, corpus-specific knowledge into an instruction following LLM (Instruct LLM). Although RAG-based prompting improves factual grounding, it fails when retrieval is incorrect or incomplete, leading to hallucinations. Finetuning methods such as RAFT and PA-RAG enhance RAG by injecting new knowledge into the model's parameters, but require generating a massive amount of synthetic QA that covers the entire corpus. Extended Pre-Training (EPT) on the text corpus avoids the need for comprehensive synthetic data generation but compromises an Instruct LLM's instruction-following capabilities, necessitating instruction fine-tuning (IFT) after pre-training. However, IFT is costly and may be infeasible due to the unavailability of an instruction-tuning corpus. In this work, we propose DKL-Decoupled Knowledge Learning for Instruction-Tuned Language Models. Instead of doing EPT on the Instruct LLM, DKL performs EPT on its corresponding base LLM to infuse new knowledge. These knowledge infused weights are then merged with the Instruct LLM, imparting new knowledge without affecting their instruction-following capabilities. DKL is a lightweight method that avoids expensive instruction fine-tuning and relies on model merging to infuse the new knowledge into the Instruct LLM without destroying its instruction following capabilities. Empirical results show that DKL improves RAG accuracy from 54.17 to 79.26 on retrieval failure cases, while outperforming prior approaches with substantially less training data.

CoMerge: Conflict-Driven Preference Optimization for Multi-Task Model Merging

Model merging provides an efficient paradigm for constructing multi-task large language models (LLMs) without full model retraining, yet it remains challenged by parameter interference. While existing methods aim to preserve the capabilities of individual expert models and mitigate interference, they generally do not directly learn from the potentially degraded behaviors exposed by naive merging. In this paper, we propose a conflict-driven preference optimization framework for model merging (CoMerge), which reformulates model merging as a preference optimization problem. The approach utilizes a self-supervised, conflict-driven strategy that leverages the defects of naive merging methods (e.g., task arithmetic) as hard negative samples to construct preference pairs without external annotations. By applying preference optimization to refine lightweight, tensor-wise merging coefficients, CoMerge enables the model to mitigate parameter-space conflicts while preserving task-specific capabilities. Extensive experiments show that CoMerge achieves an average normalized performance of 0.9968 on MergeBench, outperforming all evaluated data-free and data-driven model-merging baselines. Furthermore, on Llama-3.1-8B-Instruct, CoMerge yields marked improvements on conflict-sensitive tasks such as instruction following and safety, while remaining highly competitive with full-parameter fine-tuning despite optimizing only 1,445 scalar coefficients.

From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix

Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool. We consolidate traffic from over 200 internal applications onto a single model by closing quality gaps identified through production error analysis along three axes: instruction following, function-calling, and internal task distribution. Quality is tracked by offline benchmarks stratified to production traffic and scored by deterministic verifiers or calibrated LLM judges. Rather than optimising all objectives jointly, which introduces cross-domain reward interference, we train a separate GRPO expert per axis and merge them via two-stage SLERP. Each expert's reward exposes a distinct failure mode, namely semantic collapse, over-calling, and verbosity hacking, each requiring a domain-specific fix. In non-reasoning mode the recipe surpasses a ${\sim}7\times$ larger by total parameters baseline on the in-house Arena with 69.6 to 65.8, instruction following with 0.85 to 0.83, and function-calling with 0.79 to 0.77, while lifting general dialogue benchmarks. The model absorbs 50% of platform traffic, 116M requests per month, at a fraction of the serving cost.

How do World Models and Policies Compose in LLM Agents? A Joint Spectral and Behavioral Account

How do LLM agents come to both understand environments they act in and master tasks set within them? Through controlled experiments combining world-model training (next-state prediction) and policy training (reward maximization), we investigate this question. We dissect the resulting models through their additive parameter updates. Geometrically, we find effective world-model updates are low-rank and share an input-feature subspace with policy updates while writing to nearly orthogonal output directions, whether trained separately or sequentially. However, we find that, in projection interventions, the sequential update induces more robustness than separate policy RL when removing the world model's leading input directions, suggesting that it has learned alternative input pathways. Behaviorally, we find the sequentially trained agent explores a wider range of states and actions. Based on this, we ask: does policy training preserve world knowledge as well as it could? We probe this with training-free merging built on the geometrically motivated input basis plus an online world-model loss during policy RL, and show both improve over the untreated baseline. Our findings suggest world knowledge and task-directed ability can be learned in geometrically complementary forms, and that future post-training pipelines should consider how best to engineer the interface between them.

DARTS: Decoder-Aware Representation Tuning via Surgery for Model Merging

Model merging combines multiple task-specific fine-tuned LLMs into a single multi-task model without additional training. However, merged models are known to suffer from representation bias: systematic drift between the merged model's hidden states and those of each individual source model. Prior work (Yang et al., 2024a) study and mitigate this bias for encoder-based vision models using a lightweight correction module trained with L1 loss. However, such bias is not studied for decoder models due to their autoregressive nature. We analyze the problem of representation bias in decoder models, and show two challenges absent in encoders: (1) the causal attention mask causes bias to accumulate across token positions, requiring position-dependent correction; and (2) not all token positions are equally important, i.e., high-entropy (decision-critical) positions matter far more than low-entropy ones. To address these challenges, we propose Decoder-Aware Representation Tuning via Surgery (DARTS). DARTS employs a novel entropy-weighted L1 loss to upweight correction at high-entropy positions where errors most affect generation quality, and a per-position additive bias that captures position-dependent error without overparameterization. We perform extensive evaluation on three domains: code generation (HumanEval), mathematical reasoning (GSM8K), and instruction following (AlpacaEval) on Llama-2-7B models, and show DARTS achieves significant improvement over the standard surgery approach while adding negligible parameters ($0.1\%$ of total parameters).

Consolidating RLVR Capabilities Across Domains: A Deep Dive into Fusion Paradigms

Reinforcement learning with verifiable rewards (RLVR) improves specific capabilities of large language models, but covering multiple capabilities often involves training separate domain experts and subsequently consolidating them. We organize three fusion paradigms by the artefacts they reuse: Merge combines expert task vectors, Mix RL pools their datasets, and multi-teacher on-policy distillation (MOPD) uses both. Because they have largely been studied in isolation, how they compare and how to choose among them remain unclear. We compare all three using shared experts and data across model scales and a multi-domain benchmark suite. Although their average performance differs by at most 1.4 points, the gap reaches 8.6 points on a single benchmark, with domain-level variation tracking cross-domain relations visible in task-vector geometry. Training dynamics expose distinct constraints: Mix RL depends on domain mixture proportions, MOPD remains bounded by its teachers, and Merge compresses all expert updates into one. All three improve single-sample accuracy without measurable gains in solution coverage or losses in held-out capabilities. These results yield a practical guideline: use Merge when experts already exist and cheap fusion is paramount; Mix RL when training a unified model without experts, with domain proportions adjusted for cross-domain transfer; and MOPD when preserving domain-specific gains matters more than surpassing teachers or minimizing end-to-end cost.

When Muon Meets Task Interference: A Spectral Perspective on Continual Learning and Model Merging

Continual learning (CL) and model merging (MM) both aim to obtain a single model that performs well across multiple tasks, challenged respectively by catastrophic forgetting and weight-disentanglement error. In the literature, these difficulties are merely treated separately and mitigated through a variety of solutions, while the geometry induced by the base optimizer is treated as an implementation detail. In this work, we show that the two difficulties are in fact two instances of the same phenomenon: a parameter update useful for one task shifts the model's outputs on another. We formalize this shared phenomenon as \textit{task interference} and reduce it to a common layer-wise Frobenius inner product $\langle ΔW_\ell, J_\ell(x)\rangle_F$. This quantity, in turn, is utilized to expose the role of the optimizer. We theoretically derive an upper bound that isolates the spectral norm $\|ΔW_\ell\|_2$ as an optimizer-controllable factor of task interference, and a per-mode analysis shows that this bound tracks the dominant part of the empirical interference. Specifically, we then identify the recent Muon optimizer as a mechanism that regulates this factor by construction. Our work reveals that its elegant control on spectral norm tightens the interference bound for both CL and MM, positioning Muon as a principled optimizer-centric approach complementary to existing solutions. Our theoretcal analysis is well validated by experimental results. Replacing the AdamW optimizer with Muon improves accuracy by up to +5.02 points on the eight-task model-merging benchmark across three CLIP backbones. For continual learning, Muon also delivers uniformly positive gains across ten class-incremental protocols, three task-incremental protocols, and the 11-task MTIL benchmark.

When Privacy Hurts Mergeability: Geometry-Aware Model Merging under Differential Privacy

Model merging promises to construct a single multi-task model from independently fine-tuned task models without accessing the original task data. This makes it attractive when task data cannot be centralized, but released task models may still leak private fine-tuning data. Differential privacy (DP) provides a principled mechanism for limiting such leakage, yet its effect on model merging remains poorly understood. In this paper, we study the geometry of differentially private model merging and identify two geometric obstacles that make private task models difficult to merge: \emph{local sharpness}, which makes task losses sensitive to the parameter displacement induced by merging, and \emph{reference drift}, which measures the displacement of private task models from the shared pretrained initialization and amplifies cross-task interference. Based on these observations, we propose \textbf{DP-Merging}, a geometry-aware framework that improves the mergeability of differentially private task models. DP-Merging uses a DP-compatible sharpness-aware objective to guide each private task model toward flatter loss regions, and a reference-based alignment regularizer to keep task models close to the shared pretrained initialization. We derive a merge-gap upper bound showing that reducing local curvature and reference drift tightens the bound on the loss increase induced by merging. Experiments on vision and language tasks across multiple privacy budgets show that DP-Merging consistently improves private merged-model performance while preserving the privacy guarantees of the underlying DP fine-tuning procedures.

A Single Suffix to Break Them All: Basin-Aware Jailbreaks for Merged Model Families

Model merging enables combining multiple fine-tuned models without additional training, but its safety implications remain poorly understood. Prior work primarily attributes merging risks to unsafe constituent models, implicitly assuming that merging individually aligned models preserves safety. In contrast, we show that model merging reveals a previously overlooked jailbreak risk rooted in the pretrained foundation model, even when all constituent models are individually safety-aligned. Motivated by this observation, we study a new threat setting where an attacker constructs jailbreak prompts that generalize across merged models sharing the same pretrained backbone, without access to the exact merging coefficients or constituent checkpoints. To exploit this phenomenon, we propose \textbf{Basin-Aware Jailbreak (BAJ)}, which formulates jailbreak generation as a min--max optimization over the merging space to produce transferable adversarial suffixes across merged model families. Experiments across diverse backbones and merging settings show that BAJ achieves consistently high transfer success rates and remains effective under existing defenses.

GRIP: Granular Reward-Guided Parameter Interpolation for Efficient Reasoning

Reasoning-oriented large language models often achieve strong problem-solving performance by generating long chains of thought, but this behavior substantially increases inference cost and latency. In contrast, instruction-tuned models tend to answer more concisely, yet often lack comparable reasoning ability. This accuracy-efficiency mismatch motivates a lightweight approach that combines the strengths of both models without full model retraining. In this paper, we propose GRIP (Granular Reward-guided Interpolation of Parameters), a reward-guided parameter interpolation framework for efficient reasoning. Given a reasoning model and an instruction model with identical architectures, GRIP assigns learnable interpolation ratios to individual modules and optimizes only these ratios while keeping both source models frozen. The interpolation ratios are trained with a reward signal that favors responses that are both correct and concise. Experiments show that GRIP achieves a better accuracy-efficiency trade-off than fixed or search-based merging baselines and further reveals module-wise fusion patterns associated with efficient reasoning.

Escaping Low-Dimensional Overlap: Multi-Task Model Merging via High-Dimensional Sparse Disentanglement

Model merging provides an efficient way to construct multi-task generalist models without additional training, but its performance often degrades under severe task interference. Task interference in model merging primarily stems from \textit{superposition}, where task-specific features become entangled within the parameter space. This entanglement renders conventional decomposition methods insufficient for effectively isolating useful task directions from interfering components. In this paper, we propose a sparse-representation-based merging framework that uses Sparse Autoencoders (SAEs) to project task vectors into a high-dimensional sparse feature space, enabling feature-level disentanglement before fusion. To reduce computational overhead, we further introduce a lightweight Group-Ranked Zeroth-Order Optimizer (GR-ZOO) to identify task-critical layers for selective merging. Experiments on both Qwen2.5-1.5B and Qwen2.5-7B demonstrate that our method consistently outperforms representative baselines, including Task Arithmetic, TIES-Merge, DARE, Fisher-Merge,and several recent training-free merging methods, across mathematical reasoning, code generation, instruction following, and general knowledge tasks. In a highly conflicting four-task setting on Qwen2.5-1.5B, our method further achieves a 2.78\% improvement over the strongest baseline.

CORAM: Coherent Orthogonal Rotation for Model Merging

Merging finetuned models combines specialized capabilities without joint training or access to the original data. Most methods operate by linear arithmetic in Euclidean weight space, which cannot carry the geometry of the update. Orthogonal Model Merging (OrthoMerge) uses a single orthogonal transform for each weight matrix, but such a transform cannot change singular values. We propose CORAM, which partitions each target matrix into row slices, represents every expert slice by its singular value decomposition in the corresponding base-model SVD frame, and merges the task-specific factors on their corresponding manifolds. Because manifold averaging contracts the merged update, CORAM applies an amplification coefficient $λ=κ\hat{c}$. The scale c_hat is estimated from the expert and merged update norms and is approximately $\sqrt{N}$ for $N$ experts with comparable update magnitudes. The restoration strength kappa is selected from the dispersion of expert updates without evaluating candidate merged models. This rule remains within 0.72 points of the best swept value on all evaluated suites. CORAM also includes spread slicing to distribute highly updated rows across slices and a residual pathway for non-target layers. Across four suites covering three model families, 3B to 9B scales, and language and vision-language experts, CORAM improves over OrthoMerge by 0.25 to 1.35 points and matches or exceeds the strongest weight-space baselines.

Mint-Agent: Introducing Finance-Native Agentic Foundation Models

Financial agents must do more than recall domain knowledge: they must be both reliable, executing precise operations over grounded evidence, and executive, sustaining long-horizon research whose conclusions remain auditable. We present Mint-Agent, a family of finance-native agentic models designed around these two scales of financial intelligence. Mint-Agent is built upon three pillars: data, harness, and algorithm. Our data engine constructs clean, specialized tasks for atomic financial capabilities and long-horizon agentic execution from real-world financial sources. MintHarness enables stable interaction with open-ended environments and maintains auditable evidence trails across extended research trajectories. Our training recipe combines SFT, critical-step OPD, and RLVR to develop separate financial reasoning and agentic execution experts, which are then unified through model merging and multi-teacher on-policy distillation into compact, general-purpose financial agents. This pipeline yields two flagship models, Mint-Cu (9B) and Mint-Ag (27B). Across professional financial benchmarks, our models demonstrate two defining strengths: (1) Reliability: Mint-Ag achieves 98.33% on RFC-Bench, surpassing GPT-5.6-Sol and Claude-Opus-4.8 by 3.66 and 3.00 points; and (2) Executability: Mint-Cu reaches 69.86% on FinSearchComp T2, outperforming Agents-A1-35B and Nex-N2-mini by 22.83 and 12.78 points, while Mint-Ag achieves 76.00% and 60.49% on FinanceAgentBench v1.1 and v2, respectively. These results establish a path toward trustworthy financial intelligence in which domain expertise, long-horizon execution, and auditable evidence are jointly engineered as a unified foundation for frontier agentic models.

Multi-Objective Bayesian Optimization for Model Merging

Model merging combines trained models directly in weight space, offering a compute-efficient alternative to additional fine-tuning. Selecting merge parameters is nevertheless difficult because downstream evaluations are expensive, gradients are unavailable, and source capabilities can conflict. We formulate merge-parameter selection as a black-box multi-objective optimization problem and introduce MOBO-Merge, a merge-operator agnostic framework that uses multi-objective Bayesian optimization to approximate the Pareto front under a limited evaluation budget. We evaluate Qwen3-4B and Llama-3.1-8B in two-model instruction-math and three-model instruction-math-code settings using Linear, SLERP, TIES, and block-wise merge operators. On held-out benchmark partitions, MOBO-Merge obtains higher mean hypervolume than random search in 11 of 12 reported comparisons. The gain is small for one-dimensional Linear interpolation but substantially larger for several TIES, block-wise, and three-objective searches. No merge operator is uniformly best: TIES leads in three of four family-setting combinations, whereas Block-Linear 4x is strongest for the Llama three-model merge. These results show that multi-objective Bayesian optimization is valuable as a search layer for expressive merge parameterizations.

CABS+: Efficient and Scalable Model Merging via Conflict-Aware Sparsification and Adaptive Weight Allocation

Model merging has recently attracted significant attention as a promising paradigm for constructing unified multi-task models without requiring additional retraining. However, parameter conflicts and knowledge interference across tasks often degrade merged-model performance. Prior work introduced Conflict-Aware and Balanced Sparsification (CABS), which reduces parameter interference through structured pruning and sequential masking. However, CABS relies on grid search to determine scaling coefficients, resulting in exponential time complexity, while its optimization objective can be dominated by high-performance tasks, leading to suboptimal overall performance. To address these limitations, we extend CABS and propose CABS+. Specifically, Adaptive Weight Allocation (AWA) optimizes merging coefficients via a gradient-free search scheme to reduce time complexity, while an asymmetric fitness function promotes more comprehensive performance gains across tasks. Moreover, we conduct a systematic empirical study of key factors influencing model merging performance and propose Relative Synergy Score (RSS) to quantify model mergeability and guide model selection. We compare CABS+ with state-of-the-art model merging methods, including CABS, AdaMerging, and WUDIMerging, across 27 datasets and 5 models covering large language, small-scale language, and vision models. Extensive experiments verify the effectiveness and efficiency of CABS+. Compared with AdaMerging and WUDIMerging, CABS+ improves overall performance by 16.97% and 12.93%, respectively, exhibits stronger stability and robustness across varying task numbers and model architectures, uses less than 25% of the GPU memory required by AdaMerging, and achieves nearly a 4x speedup in merging time over WUDIMerging.

Orientation, not magnitude: the causal structure of task-vector interference in merged language models

Model merging by task arithmetic works until it doesn't, and the field diagnoses why with magnitudes: layerwise representation bias, deviations from cross-task linearity, parameter overlap. Tracking the exact layerwise cross-term of merged LLMs through a factorial ledger and intervening on it directly, we find magnitude insufficient - and inconsistent across model families - as a diagnostic axis. An exact decomposition of the layerwise flux shows it is dominated by amplifying transport of the existing cross-term (~65-70% in both families, gain >1 per late block), and erasing the term is undone by propagation - rebuilt to 99% of its norm at cosine 0.99 - unless applied near the output; a basin test with six starting displacements establishes the carried direction as an attractor of the forward pass. That direction is causally load-bearing: erasure along it removes expressed interference dose-dependently and saturates at exact erasure, while norm-matched wrong-direction controls fail or backfire. Instruction wrappers gate the effect: the same erasure finds 13x less relative interference to remove under a wrapper that internally amplifies the cross-term, because the wrapper drowns the interaction in a template-pinned main effect rather than shrinking it - a structure that replicates across further instruction templates but not under a length-matched control. Magnitude, by contrast, is at best a coarse correlate, and the striking +-15% "universality" of naive bfloat16 generation turns out to be quantization roughness. Task pairs whose local cross-term generation differs by at most 1.9x differ by 14x-337x in causally removable interference. All 46 predictions were preregistered and frozen before their data; falsifications, including of our own headline expectations and of behavioral recovery under a validated continuous endpoint, are reported as such.

HyperFix: Combinatorial Nonlinear Correction for Task Vector Merging

Task vectors enable model merging without joint retraining. In practice, the subset of task vectors to be merged may vary, but many existing methods use scalar tuning for a particular subset, requiring repeated tuning across subsets and restricting task vector merging to linear rescaling. We therefore formulate merging across varying task subsets as a combinatorial correction problem and introduce HyperFix, a lightweight hypernetwork that predicts subset-conditioned nonlinear corrections in weight space. Trained once on singleton, pair, and triple subsets from a task bank, HyperFix generalizes to larger subsets without per-subset optimization. Our local perturbation analysis bounds the residual correction beyond linear merging and motivates learning it from small task updates. Experiments across diverse benchmarks show that HyperFix outperforms existing task vector merging methods while reducing tuning cost.

CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation

Large language models (LLMs) have advanced code generation, where executable feedback provides a more reliable learning signal than textual imitation alone. Hardware verification is an important application of code generation and accounts for a substantial fraction of modern chip design effort, with high-coverage testbench stimulus generation as a key task. We present CHORUS, a post-training framework that pushes performance beyond what a conventional supervised fine-tuning (SFT)-to-reinforcement learning (RL) pipeline achieves. CHORUS builds on two observations. First, staged SFT produces behaviorally diverse checkpoints, and dense-reward RL turns them into strong experts with comparable aggregate performance but distinct task-level strengths. Second, these complementary strengths can be exploited through either training-free model merging or further post-training to outperform the best individual expert. By consolidating the resulting specialists into a single 4B model, CHORUS achieves 88.0% Pass@1 on CVDP-ECov, outperforming DeepSeek-R1 (671B) by 13.5 percentage points.

When Do Task Vectors Interfere? Mapping the Validity Boundaries of Weight-Space Composition

Task arithmetic treats fine-tuning displacements as composable directions in weight space, yet it remains unclear when parameter addition reflects predictable changes in model function. We separate parameter geometry from functional geometry and measure pairwise functional non-additivity over a two-dimensional task-vector surface, using a first-token predictive-distribution interaction ratio conditioned on an input distribution and evaluated with norm-matched controls, three training seeds, and response-only fine-tuning. On Qwen2.5-1.5B, code+safety is more non-additive than the matched code+math control on code and instruction prompts, but not on math prompts. In a prospectively specified six-task expansion, all eight high-versus-low comparisons of unseen task pairs have the predicted sign. The primary ordering further persists under full-parameter fine-tuning at 0.5B, Qwen2.5 LoRA scale tests up to 7B, and a Llama-3.1-8B cross-architecture audit. External validation exposes a sharper boundary: raw public code, instruction, and safety prompts preserve the continuous contrast, whereas an instruction-style wrapper collapses it on the identical public-code prompts, and EvalPlus pass@1 interactions do not robustly reproduce it. Weight-space composition therefore supports coarse, input- and format-conditioned functional statements across adaptation methods, scales, and one additional model family, not a universal merging-performance predictor.

DiffSafeMerge: Mitigating Backdoor Inheritance in Diffusion Model Merging

Unconditional diffusion checkpoint merging assumes benign sources, yet a compromised public checkpoint can transfer a dormant backdoor while clean generation appears normal. Mitigation is difficult without knowing the compromised source, trigger, or target, and broad sanitization may degrade image quality. We introduce DiffSafeMerge (DSM), which uses a small unlabeled clean set and fixed, attack-agnostic stress probes to score source blocks, shrink suspicious contributions toward a trusted reference, and select attenuation under a clean denoising-loss budget. We evaluate four attacks, two datasets, and 21 target conditions. Intended merging already has zero worst-target ASR in 10 of 14 source cases; DSM preserves these outcomes and records no target match in the remaining four over three seeds, including three with baseline ASR of 48--100\%. Among methods with zero worst-target ASR on both datasets, DSM obtains the lowest case-averaged FID in the matched seed-0 comparison.

Signature-Guided Capacity Occupancy for Dense Expert Merging

Dense expert merging combines domain-specialized language models into one single checkpoint, typically by admitting task-vector support in weight space. However, this admission is governed by three decisions that existing methods answer only partially: where to open layer capacity from cross-expert conflict, who should occupy that capacity based on domain demand, and how to admit the resulting support without relying on costly recipe search. To tackle these issues, we propose SigMerge (Signature-Guided Capacity Occupancy), a structured capacity assignment framework for dense expert merging. Starting from a dense base merge, conflict signatures set each layer's capacity from cross-expert conflict, positive base-merge deficits set each domain's share of that capacity, and a sequential occupancy rule admits each expert delta up to the resulting layer-domain budget. Across 21 paired settings spanning seven dense base merges and three model pools, SigMerge improves every one (by 15.0% on average) and achieves the best average rank (1.67) among six merging methods, outperforming three categories of merging baselines.

When Skills Meet Safety: Benchmarking and Characterizing the Adaptive Jailbreak Robustness of Skill-Merged LLMs

Model merging has become the default way to give an aligned language model new skills without retraining: a practitioner folds task vectors from math, code, or domain specialists into a safety-aligned base using task arithmetic, TIES, or DARE. This convenience is known to carry a safety cost, but almost all of that evidence rests on static refusal tests: fixed harmful prompts scored for compliance. We argue this is misleading. Because safety alignment is "shallow," concentrated in the first few generated tokens, a merged model's static refusal can stay clean while a real adaptive attack still breaks it. We introduce SkillSafe-Bench, a controlled benchmark that scores skill-merged models on static refusal, adaptive jailbreak robustness, and capability retention under a conservative two-judge AND rule. Across six open-weight bases (five families, two scales), static safety does not predict robustness to attack: under a semantic template attack, safe-looking merges on the fragile bases (both Qwen scales and Gemma) are jailbroken 60-76% of the time while others (Llama, Phi-4) stay robust. We further show the static effect of merging is base-conditional, characterize same-recipe abliteration-style safety erosion through a data-free geometric signal (the overlap of a task vector with a safety subspace), and outline SubSafe-Merge, which projects this overlap away to remove that erosion at held capability. Adaptive evaluation is not optional for merged LLMs: the models that most need it look safe under static screening.

AgentPatch: Coarse-to-Fine Weak-Task Repair for Merging Agentic Multimodal Large Language Models

Agentic multimodal large language models (MLLMs) extend multimodal perception and reasoning with planning, tool use, and interaction in dynamic environments. Yet current models are specialized for particular tools or environments, complicating consolidation into a single generalist. We formulate Agentic MLLM Merging and identify two challenges: asymmetric capability preservation, whereby capabilities with different interaction complexity are retained unevenly, producing weak tasks after merging, and behavior-critical forgetting, whereby losing decisive actions can derail long-horizon execution. We propose AgentPatch, a training-free coarse-to-fine repair framework. It selects a stable merged backbone, restores diluted weak-task-specific signals through Weak-Task Unique Residual Recovery, and applies an Agent-Guided Behavior-Critical Patch that recovers decisive behaviors under explicit capability protection. AgentPatch produces a single static checkpoint without routing or ensembles. Experiments across six agentic and multimodal benchmarks show that AgentPatch improves diverse merged backbones, alleviates weak-task degradation, and better balances weak-task recovery with the preservation of complementary search and agentic visual processing capabilities. Code is available at https://github.com/ziboshao/AgentPatch.

A Unified Model for Cross-Domain Clone Detection via Model Merging

The growing diversity of code clone types, from syntactic copies to cross-language semantic clones to AI-generated duplicates, has created a fragmentation crisis in clone detection. Current deep learning detectors are domain specialists that degrade significantly outside their training distribution, with F1 drops exceeding 70% across domains. Deploying multiple specialized models is impractical, yet training a single cross-domain detector requires simultaneous access to all training data. To address this, we investigate model merging, a family of post-hoc techniques that operate solely on trained checkpoints. We evaluate parameter merging with five task-vector methods, architecture merging via greedy layer stitching, and cross-tokenizer alignment across four code models, three benchmarks, and twelve configurations. Same-base TIES merging creates effective cross-domain detectors, validated across two model families and three random seeds, reaching 0.865 combined F1 on UniXcoder, 93% of multi-task performance without any training data at the merging step. WUDI achieves the highest in-distribution combined F1 at 0.899, but TIES generalizes better to unseen AI-generated clones, making it our recommended method. Cross-base merging yields only marginal and high-variance gains across all five methods, indicating that task vector compatibility through a shared pre-trained base is the binding factor for effective merging. Merged detectors also outperform zero-shot code LLMs on GPTCloneBench at lower inference cost and generalize up to 4x better than multi-task training to unseen AI-generated clones, suggesting a trade-off between in-domain performance and OOD robustness. This work provides one of the first systematic empirical studies of model merging for software engineering and a practical recipe for building cross-domain clone detectors.

SAFE-Merge: Data-Free Continual Model Merging with General Knowledge Preservation

Data-free continual model merging must incorporate a stream of specialized models while retaining both pretrained general knowledge and previously acquired tasks, without access to task data. Existing methods mainly merge task updates by suppressing interference among downstream tasks; while this protects previously acquired tasks, it overlooks the safety of the pretrained knowledge itself, whose erosion degrades generalization to held-out distributions and weakens the foundation for future task acquisition. We propose SAFE-Merge, a simple data-free continual-merging framework that first decides which parameter updates are safe to retain, and then recovers the task information lost through masking. Specifically, to ensure safety, risk-aware sparse masking selects parameter updates that carry task-specific information while posing low risk to general knowledge. Masked low-rank recovery then compensates for the lost task information using only the same retained parameter updates, while leaving all masked-out parameters strictly unchanged. Finally, the combined update is fused into the backbone, incurring no additional inference cost. Across vision and language benchmarks, SAFE-Merge consistently achieves the best H-score. On longer CLIP task sequences, it substantially improves H-score over NUFILT while also achieving the highest accuracy.

Sharpness-aware Model Merging with Salience Recovery for LLM-based Cross-Domain Sequential Recommendation

LLM-based Cross-Domain Sequential Recommendation (CDSR) leverages LLMs to enhance target performance via deep semantic reasoning, alleviating the dependency on overlapping users. Among LLM-based paradigms, model merging is particularly promising for multi-domain scenarios due to its superior scalability and flexibility in integrating diverse knowledge sources. However, our empirical investigations reveal two critical bottlenecks: (1) cross-domain knowledge conflict; and (2) performance saturation in multi-domain fusion. Our analysis attributes these phenomena to parameter-level misalignment and statistical homogenization during the merging process. To address these bottlenecks, we propose SharpRec, Sharpness-aware Model Merging with Salience Recovery for LLM-based CDSR, a framework designed to lift the performance upper bound of merged models. SharpRec incorporates two synergistic modules: Sharpness-aware Geometric Alignment to establish a stable geometric foundation for interference-free fusion; and Preference Salience Activation to effectively recover the distinctive features essential for bolstering target domain performance. Extensive experiments in both dual-domain and multi-domain scenarios demonstrate that SharpRec consistently outperforms state-of-the-art baselines.

Rethinking Expert Training for Model Merging with Prompt Learning

Model merging aims to combine multiple domain-specialized experts trained from a shared foundation model into a single multi-task model. Existing approaches largely focus on improving the merging procedure itself and typically assume experts obtained through full-parameter fine-tuning. In this work, we revisit expert training for model merging. We first show that prompt-based adaptation provides a strong baseline: independently learned prompts can be exploited across tasks while keeping the backbone fixed, avoiding the interference introduced by weight merging. Building on this observation, we introduce Dual-Tuned Experts (DTEs), a two-stage training strategy that first learns prompts and then fine-tunes the vision encoder. This reduces the magnitude of task-specific parameter updates and produces experts with higher merge compatibility. Experiments across multiple CLIP architectures, full fine-tuning, and LoRA experts show that DTEs consistently improve merged performance of standard merging approaches and remain effective even when combining heterogeneous sets of experts.

Asymmetric Collapse in Model Merging: When Refusal Over- writes Recognition

Model merging is often used to combine capabilities from separately fine-tuned models without additional training, but it is unclear whether standard merging methods preserve multiple safety-relevant behaviors simultaneously. We study this question through a controlled case study using two Gemma-3-1B-IT finetunes on two complementary safety objectives: CARES harm-level classification and WildJailbreak adversarial refusal. We merge the two fine-tunes using Linear, SLERP, TIES, and DARE-TIES, and evaluate the merged models on classification accuracy, attack resistance, and benign compliance. Across all four methods, attack resistance transfers significantly more than classification accuracy: merged models retain 81-85% jailbreak refusal rates while CARES accuracy falls to at most 12.9%. Weight-space measurements suggest that this asymmetry is not caused by strongly opposing task-vector directions: the two task vectors are nearly orthogonal (cosine similarity 0.011). Instead, the refusal fine-tune induces consistently larger per-layer task-vector magnitudes, causing magnitude-sensitive methods to favor refusal updates. These results show that standard model merging can collapse safety recognition into broad refusal when safety-relevant task vectors differ substantially in scale.

Enough is as good as a feast: A Comprehensive Analysis of How Reinforcement Learning Mitigates Task Conflicts in LLMs

Model merging plays a crucial role in consolidating multiple specialized models into a single, unified model, especially in the era of large language models (LLMs). Recent research has primarily focused on developing strategies to enhance merging performance with the trained models, while the impact of training paradigms, such as supervised fine-tuning (SFT) and reinforcement learning (RL), on the effectiveness of model merging remains underexplored. In this study, we systematically explore the merging behavior of RL-trained LLMs compared to those trained with traditional SFT. Through comprehensive evaluations across five representative tasks, we find that RL significantly reduces task conflicts and results in less performance degradation after merging, making RL-trained models particularly well-suited for this process. To unearth the reasons behind the superior suitability of RL for model merging, we conduct extensive empirical experiments and theoretical analyses. Our findings highlight three key factors: (1) On-policy training data in RL control the gradient updates in a smaller magnitude, reducing the risk of overwriting existing knowledge for other tasks in the model. (2) The RL optimization objective, which favors ``\textit{enough is as good as a feast}", progressively reduces the magnitude and the number of conflict parameter updates as the model converges. (3) Joint optimization of positive and negative examples in RL steers the model towards an unbiased task-specific parameter subspace, ensuring robust performance while further preventing parameter conflicts.

TwistedMerge: Certified Higher-Order Diagnostics and Abstention for Model Merging

Model merging combines independently trained or fine-tuned models, but pairwise alignability does not imply globally consistent alignment. We formulate merging as a finite descent problem in which checkpoints are local objects, alignment maps are transitions, and cycle products are residuals. TwistedMerge is a conservative certification pipeline that separates fixed-chart averaging, synchronization-removable gauge inconsistency, a certified central obstruction on a specified comparison complex, and nonabelian holonomy. A residual is promoted to a cohomology class only after inverse-consistency, coefficient-identification, centrality, and closure tests; otherwise the method abstains and returns an ordinary or synchronized fallback. We prove a constant-edge no-go result, frozen-complex three-way and predeclared-family error-control theorems, and a refinement test for comparison-complex sensitivity. A planted neural alignment defect is removed by cycle-consistent synchronization, showing that a nonzero cycle score alone is not a higher obstruction. Controlled central systems recover the predicted non-coboundary and projective-rank behavior, while noisy estimates move from certification to abstention without false lifts on the tested controls. A trained low-rank-adapter audit shows that naive factor averaging depends on the chosen GLr representative, whereas global factor synchronization and dense-delta SVD are stable. On natural checkpoint collections, cycle residuals do not predict merge degradation and no natural central or period-index class is certified. The results position descent theory as a falsifiable certification and abstention framework.