[論文レビュー] ReasonAny: Incorporating Reasoning Capability to Any Model via Simple and Effective Model Merging

Junyao Yang, Chen Qian+4 more|arXiv (Cornell University)|Jan 9, 2026

Explainable Artificial Intelligence (XAI)被引用数 0

ひとこと要約

tldr: ReasonAny introduces a training-free approach to merge reasoning capability with any base model by simple model merging, resolving parameter conflicts with an exclusion mechanism while preserving safety alignment.

ABSTRACT

Large Reasoning Models (LRMs) with long chain-of-thought reasoning have recently achieved remarkable success. Yet, equipping domain-specialized models with such reasoning capabilities, referred to as "Reasoning + X", remains a significant challenge. While model merging offers a promising training-free solution, existing methods often suffer from a destructive performance collapse: existing methods tend to both weaken reasoning depth and compromise domain-specific utility. Interestingly, we identify a counter-intuitive phenomenon underlying this failure: reasoning ability predominantly resides in parameter regions with low gradient sensitivity, contrary to the common assumption that domain capabilities correspond to high-magnitude parameters. Motivated by this insight, we propose ReasonAny, a novel merging framework that resolves the reasoning-domain performance collapse through Contrastive Gradient Identification. Experiments across safety, biomedicine, and finance domains show that ReasonAny effectively synthesizes "Reasoning + X" capabilities, significantly outperforming state-of-the-art baselines while retaining robust reasoning performance.

研究の動機と目的

• Motivate efficient development of reasoning-enabled models without full retraining.
• Enable synthesis of reasoning skills with domain-specific capabilities in a single step.
• Address parameter conflicts during model merging with a simple exclusion-based approach.
• Assess impact on safety alignment and potential reduction in jailbreaking or safety degradation.

提案手法

• Propose a simple model-merging framework to inject reasoning capability into any model.
• Use an exclusion process to resolve conflicting parameters between the reasoning and domain-specific subspaces.
• Focus on training-free synthesis by avoiding full retraining of base models.
• Identify and preserve safety alignment parameters during the merge to mitigate jailbreak risks.
• Discuss computation trade-offs due to gradient-based attribution used in the identification phase.

実験結果

リサーチクエスチョン

• RQ1 Can ReasonAny effectively integrate reasoning capabilities into arbitrary base models without retraining?
• RQ2 Does the exclusion-based merging reduce parameter conflicts without compromising domain-specific performance?
• RQ3 How well does ReasonAny preserve safety alignment and mitigate safety degradation after merging?
• RQ4 What are the computational overheads associated with the identification phase compared to weight-averaging methods?

主な発見

• ReasonAny enables training-free synthesis of reasoning with domain-specific capabilities in base models.
• The exclusion mechanism resolves parameter conflicts between reasoning and domain subspaces to minimize interference.
• ReasonAny preserves safety alignment parameters, reducing risks of jailbreaking or safety degradation relative to other merging techniques.
• The approach highlights computational overhead from gradient-based attribution during the identification phase.
• Limitations include potential overlap between subspaces in complex tasks and current focus on two-model merging.
• Broader impact suggests lower resource use and carbon footprint for developing reasoning-enabled models.