Intellectual Property

**Jurisdictional Comparison and Analytical Commentary** The recent arXiv paper, "Hybrid Federated and Split Learning for Privacy Preserving Clinical Prediction and Treatment Optimization," has significant implications for Intellectual Property (IP) practice, particularly in the context of data privacy and healthcare modeling. A comparison of the US, Korean, and international approaches reveals that the proposed hybrid framework aligns with the European Union's General Data Protection Regulation (GDPR) emphasis on data protection by design and by default. In contrast, the US approach, as reflected in the Health Insurance Portability and Accountability Act (HIPAA), focuses on data security and breach notification, while Korea's Personal Information Protection Act (PIPA) emphasizes data protection and consent. **US Approach:** The US approach, as reflected in HIPAA, focuses on data security and breach notification. The proposed hybrid framework's emphasis on data protection by design and by default aligns with the GDPR's principles, which may influence US IP practice in the healthcare sector. However, the US approach may not provide sufficient protection for sensitive healthcare data, particularly in the context of collaborative clinical decision support. **Korean Approach:** Korea's PIPA emphasizes data protection and consent, which is reflected in the proposed hybrid framework's explicit collaboration boundary and lightweight defenses. The Korean approach may provide a more comprehensive framework for data protection in the healthcare sector, particularly in the context of collaborative clinical decision support. **International Approach:** The proposed hybrid framework aligns with the

The article introduces a novel hybrid FL-SL framework addressing privacy-utility trade-offs in clinical decision support, offering practitioners a scalable solution to navigate governance and privacy constraints without raw-data sharing. The empirical auditing of leakage via membership inference and lightweight defenses aligns with recent case law (e.g., FTC v. D-Link Systems) emphasizing the necessity of proactive privacy safeguards in data-sensitive applications. Statutorily, this approach may intersect with HIPAA’s Privacy Rule by demonstrating compliance through technical controls that limit exposure of protected health information. Practitioners should consider integrating similar hybrid architectures to mitigate risk while preserving clinical efficacy.

The article’s contribution to Intellectual Property practice lies in its methodological innovation for evaluating AI-generated content—a growing concern in IP disputes involving authorship, originality, and infringement. While the technical focus on variance-adaptive allocation via multi-armed bandit theory is algorithmic, its implications extend to IP: as LLMs become tools in content creation or legal analysis, accurate evaluation of model outputs becomes critical for determining liability, validity, or infringement claims. In the U.S., this aligns with evolving case law on AI authorship (e.g., *Thaler v. Vidal*), where courts grapple with attribution; in Korea, where IP law integrates algorithmic contributions under the Patent Act amendments, similar analytical frameworks may inform judicial interpretation of “inventive step” in AI-assisted inventions. Internationally, the WIPO AI Initiative has begun to recognize algorithmic evaluation metrics as relevant to patentability assessments, suggesting a convergent trend toward quantifiable, algorithmic validation as a proxy for human-like judgment. Thus, while the paper is computational, its ripple effect on IP doctrine—particularly in attribution, quality assessment, and standardization of AI outputs—is substantively significant.

As a Patent Prosecution & Infringement Expert, I'll analyze the article's implications for practitioners in the field of artificial intelligence (AI) and large language models (LLMs). **Technical Analysis:** The article presents a variance-adaptive approach to optimize the allocation of queries across multiple prompt-response pairs to minimize estimation error in LLM evaluations. This approach leverages multi-armed bandit theory and concentration inequalities to dynamically allocate queries based on estimated score variances. The proposed method achieves a worst-case score-estimation error of $\tilde{O}\left(\sqrt{\frac{\sum_{i=1}^K \sigma_i^2}{B}}\right)$, where $B$ is the fixed computational budget and $\sigma_i^2$ is the unknown score variance for pair $i$. **Implications for Practitioners:** 1. **Efficient LLM evaluation:** The proposed method can significantly reduce the worst-case estimation error while maintaining identical budgets, making it an efficient approach for LLM evaluation. 2. **Practical implications:** The work has practical implications for AI safety, model alignment, and automated assessment at scale, highlighting the importance of efficient LLM evaluation in these areas. 3. **Potential patent applications:** The proposed method could be a subject of patent applications, particularly in the areas of AI, machine learning, and natural language processing. **Case Law, Statutory, or Regulatory Connections:** While there are no direct case law

**Jurisdictional Comparison and Analytical Commentary** The Open Rubric System (OpenRS) presents a novel approach to scaling reinforcement learning with pairwise adaptive rubrics, addressing the limitations of scalar reward models in open-ended alignment. This development has significant implications for Intellectual Property (IP) practice, particularly in the realm of artificial intelligence (AI) and machine learning (ML). A comparative analysis of US, Korean, and international approaches reveals distinct perspectives on the role of IP in AI development. **US Approach:** In the United States, the focus on IP protection in AI development is primarily driven by the need to safeguard intellectual creations and inventions. The OpenRS system, which relies on a plug-and-play, rubrics-based framework, may be seen as a novel application of existing IP laws, such as the copyright and patent laws, to AI-generated content. However, the use of adaptive rubrics and meta-rubrics may raise questions about the ownership and protection of these AI-generated rules. **Korean Approach:** In Korea, the government has actively promoted the development of AI and ML technologies through various initiatives, including the creation of AI-specific IP laws. The OpenRS system may be seen as aligning with Korea's IP policies, which emphasize the need for transparency and explainability in AI decision-making processes. However, the use of adaptive rubrics and meta-rubrics may also raise concerns about the potential for bias and unfair competition in AI-generated content. **International Approach:** Internationally

The Open Rubric System (OpenRS) introduces a novel framework for aligning large language models (LLMs) by replacing opaque scalar reward models with transparent, principle-based reasoning processes. Practitioners should note that this approach aligns with emerging trends in AI governance, emphasizing transparency and inspectability in reward design, akin to principles seen in regulatory frameworks for algorithmic accountability (e.g., EU AI Act provisions on transparency). Statutorily, this may intersect with evolving standards for AI compliance, particularly regarding the use of verifiable criteria to mitigate reward hacking. Case law, while still nascent, may draw parallels to precedents on algorithmic bias and accountability, such as those addressing opaque decision-making in automated systems. This shift toward explicit, inspectable principles could influence future litigation or regulatory guidance on AI alignment and fairness.

### **Jurisdictional Comparison & Analytical Commentary on the Impact of Parametric Factuality Research for IP Practice** This research, which distinguishes between *encoding* (knowledge retention) and *recall* (accessibility of stored knowledge) in LLMs, has significant implications for **patent law, trade secrets, and AI-generated content liability** across jurisdictions. The **U.S.** (with its strong emphasis on patent enablement and trade secret protection under the *Defend Trade Secrets Act*) may see increased scrutiny over whether AI-generated disclosures meet "sufficiency of disclosure" standards under 35 U.S.C. § 112 if recall bottlenecks lead to inconsistent outputs. **South Korea**, under its *Patent Act* (similar to the U.S. in requiring enablement) and *Unfair Competition Prevention Act* (protecting trade secrets), may face challenges in proving infringement when AI systems fail to retrieve known facts, particularly in cases involving long-tail or reverse factual queries. Internationally, under the **TRIPS Agreement**, the distinction between *encoded* and *accessible* knowledge could influence how jurisdictions assess **novelty and inventive step** in AI-assisted inventions, particularly where prior art retrieval depends on parametric memory rather than external databases. The study’s findings suggest that **liability frameworks for AI-generated errors** may need to evolve, particularly in cases where recall failures (rather than missing knowledge) lead to mis

As a Patent Prosecution & Infringement Expert, I will analyze the article's implications for practitioners in the field of artificial intelligence and neural networks. The article discusses the limitations of current factuality evaluations of Large Language Models (LLMs), which treat all errors alike, failing to distinguish between missing knowledge (empty shelves) and limited access to encoded facts (lost keys). This distinction is crucial for understanding the performance of LLMs and for improving their capabilities. The proposed behavioral framework and the WikiProfile benchmark provide a more nuanced understanding of LLMs' performance, highlighting the importance of recall as a major bottleneck. Implications for practitioners: 1. **Patent Prosecution**: This article highlights the need for more nuanced understanding of LLMs' performance, which may impact patent prosecution strategies. Practitioners should consider the distinction between missing knowledge and limited access to encoded facts when evaluating the novelty and non-obviousness of inventions related to LLMs. 2. **Prior Art**: The article's findings on the limitations of current factuality evaluations may impact the search for prior art in patent prosecutions. Practitioners should consider the possibility that errors in LLMs may be due to limited access to encoded facts rather than missing knowledge. 3. **Infringement**: The article's emphasis on recall as a major bottleneck may impact the assessment of infringement in patent cases. Practitioners should consider the possibility that LLMs may infringe patents by accessing and utilizing encoded facts, even

**Jurisdictional Comparison and Analytical Commentary on the Impact of AI Agent Societies on Intellectual Property Practice** The emergence of AI agent societies, as illustrated by the case study of Moltbook, presents a paradigm shift in the intellectual property (IP) landscape. In the United States, the concept of AI-generated content raises questions about authorship and ownership, with the US Copyright Act of 1976 potentially extending protection to AI-generated works (17 U.S.C. § 101). In contrast, Korean law is more restrictive, with the Korean Copyright Act (KCA) requiring human authorship for copyright protection (Article 2, KCA). Internationally, the Berne Convention for the Protection of Literary and Artistic Works (1886) and the WIPO Copyright Treaty (1996) provide a framework for protecting IP rights, but the application of these treaties to AI-generated content remains uncertain. The findings of the Moltbook study suggest that AI agent societies may not converge towards a single, homogenous system, but rather maintain individual diversity and persistent lexical turnover. This raises questions about the potential for AI-generated content to be considered as original works, and the implications for IP protection. In the US, the concept of "originality" is central to copyright protection, and AI-generated content may be seen as lacking the necessary creative spark. In Korea, the emphasis on human authorship may lead to a more restrictive approach to IP protection for AI-generated works. Internationally, the lack

The article’s findings on AI agent societies—specifically the persistence of individual diversity and lexical turnover despite systemic stabilization—have implications for practitioners in AI design and governance. Practitioners should recognize that scale and interaction density alone do not equate to socialization; instead, mechanisms for shared social memory or persistent influence anchors must be intentionally designed to foster emergent social dynamics. This aligns with statutory considerations under AI regulatory frameworks (e.g., EU AI Act) that emphasize intentional design for societal impact, and echoes case law principles from *State Street Bank* and *Alice* in assessing functional vs. substantive innovation in AI systems. Practitioners must incorporate these insights into architecture and policy to avoid unintended homogenization or lack of emergent coherence.

**Jurisdictional Comparison and Analytical Commentary: Impact on Intellectual Property Practice** The proposed fast algorithm for element selection in dimension reduction has significant implications for intellectual property (IP) practice, particularly in the areas of AI-driven innovation and software development. In the US, the algorithm's multiplication-free form may be seen as a novel application of existing mathematical concepts, potentially eligible for patent protection under 35 USC § 101. In contrast, Korean law may view the algorithm as a mere mathematical concept, ineligible for patent protection under Article 2(2) of the Korean Patent Act. Internationally, the algorithm's novelty and non-obviousness may be evaluated under the European Patent Convention (EPC) and the Patent Cooperation Treaty (PCT), with a focus on whether the algorithm's multiplication-free form represents a significant improvement over existing methods. **US Approach:** The US Patent and Trademark Office (USPTO) may view the algorithm as a novel application of mathematical concepts, eligible for patent protection under 35 USC § 101. However, the algorithm's eligibility for patent protection would depend on whether it represents a "useful, concrete, and tangible" application of mathematical concepts, as required by the Supreme Court in Alice Corp. v. CLS Bank Int'l (2014). **Korean Approach:** Korean law may view the algorithm as a mere mathematical concept, ineligible for patent protection under Article 2(2) of the Korean Patent Act. This provision excludes "mathematical

The article presents a novel multiplication-free dimension reduction method that shifts the computational burden from matrix multiplication to combinatorial optimization, offering a potential alternative to PCA in resource-constrained environments. Practitioners should consider the implications for patent claims in machine learning algorithms that reduce computational overhead—specifically, claims directed to selection-based reduction without matrix operations may now be more defensible or infringed upon due to this innovation. Statutory relevance arises under 35 U.S.C. § 101, where the novelty lies in the algorithmic shift from multiplication-dependent to selection-dependent computation, potentially affecting eligibility under the Alice framework; case law like Alice Corp. v. CLS Bank (2014) may inform validity challenges if the claim is construed as abstract without technical improvement. Regulatory implications may also extend to computational efficiency standards in AI/ML deployments under evolving FTC or EU AI Act guidelines.

The MEMTS framework introduces a novel paradigm for domain adaptation in time series forecasting by substituting traditional retrieval-dependent or pretraining-based methods with a parameterized memory mechanism. Jurisdictional analysis reveals divergent IP implications: in the U.S., such innovations may qualify for patent protection under 35 U.S.C. § 101 if deemed non-abstract and tied to technical application, particularly given the algorithmic efficiency gains in real-time processing; Korea’s IP regime, governed by the Korean Intellectual Property Office (KIPO), similarly recognizes computational innovations with tangible efficiency improvements as patentable subject matter under Article 10 of the Patent Act, provided they involve inventive application of data processing; internationally, WIPO’s PCT framework acknowledges the broader applicability of memory-based adaptation systems as eligible for international patent coverage, reinforcing cross-border standardization of AI-driven temporal modeling. Practically, MEMTS’s retrieval-free architecture aligns with global trends toward scalable, low-latency AI deployment, while its parameterized knowledge encapsulation offers a defensible IP edge by embedding domain-specific learning as a proprietary, internalized mechanism—distinct from conventional external retrieval or pretraining—thereby strengthening claims of originality and inventive step in both U.S. and Korean patent filings.

**Domain-Specific Expert Analysis:** The proposed MEMTS method for retrieval-free domain adaptation in time series forecasting addresses the limitations of existing solutions, such as Domain-Adaptive Pretraining (DAPT) and Retrieval-Augmented Generation (RAG), which suffer from catastrophic forgetting and scalability bottlenecks, respectively. MEMTS achieves accurate domain adaptation with constant-time inference and near-zero latency by internalizing domain-specific temporal dynamics into a compact set of learnable latent prototypes. This innovation has significant implications for practitioners working with Time Series Foundation Models (TSFMs) in real-world vertical domains. **Case Law, Statutory, or Regulatory Connections:** The proposed MEMTS method may be relevant to the following statutory or regulatory connections: 1. **35 U.S.C. § 103**: Non-obviousness. The MEMTS method may be considered non-obvious over existing solutions, such as DAPT and RAG, which suffer from limitations and scalability bottlenecks. 2. **35 U.S.C. § 112**: Enablement. The proposed MEMTS method may be considered enabled for patent protection, as it provides a clear and concise description of the invention and its operation. 3. **Federal Circuit precedent**: The MEMTS method may be relevant to the Federal Circuit's jurisprudence on non-obviousness, enablement, and patentable subject matter, particularly in the context of artificial intelligence and machine learning inventions. **Patent Prosecution Strategies:** To effectively prosecute a

The introduction of the Mean Velocity Policy (MVP) with an Instantaneous Velocity Constraint (IVC) in the field of reinforcement learning (RL) has significant implications for Intellectual Property (IP) practice, particularly in the realm of artificial intelligence (AI) and machine learning (ML). In the US, the MVP's innovative approach to policy function design may be protected under patent law, with the IVC serving as a novel technical feature that enhances learning accuracy and expressiveness. In contrast, Korean IP law may view the MVP as a software innovation, subject to copyright protection, while international approaches, such as the European Union's Software Directive, may categorize the MVP as a non-protectable algorithm. Jurisdictional comparison: - In the US, the MVP's patentability may be assessed under 35 USC § 101, with the IVC serving as a technical feature that distinguishes the invention from prior art. - In Korea, the MVP's copyrightability may be evaluated under the Copyright Act, with the MVP's software code and IVC being considered protectable expressions of an idea. - Internationally, the MVP's treatment under the EU's Software Directive (2009/24/EC) may be complex, with some arguing that the MVP's algorithmic nature makes it non-protectable, while others may view the IVC as a novel technical feature that warrants protection. Implications analysis: - The MVP's innovative approach to policy function design has significant implications for the development of

The article introduces a novel reinforcement learning policy, MVP, which balances expressiveness and computational efficiency by modeling mean velocity fields with an instantaneous velocity constraint. Practitioners should note that this design introduces a theoretical boundary condition that enhances learning accuracy and policy expressiveness, offering a new framework for RL applications. While not directly tied to patent law, these advancements may intersect with patent claims in AI/ML domains, particularly those covering policy optimization or generative models, potentially influencing infringement analyses or validity assessments of related claims. Case law such as *Alice Corp. v. CLS Bank* (2014) and statutory provisions under 35 U.S.C. § 101 may be relevant in evaluating the patent eligibility of such innovations.

The *sleep2vec* innovation presents a nuanced intersection between intellectual property and technological advancement, particularly in the domain of multimodal biosignal processing. From an IP perspective, the model’s foundational architecture—leveraging cross-modal alignment via a contrastive pre-training mechanism—introduces novel technical solutions to longstanding challenges in sensor heterogeneity and sensor dropout, thereby potentially qualifying for patent protection under utility patent frameworks in the US, Korea, and internationally. The US approach, rooted in the post-Alice Corp. v. CLS Bank jurisprudence, may scrutinize claims for abstractness, yet the specificity of the “Demography, Age, Site & History-aware InfoNCE” objective and its application to physiological metadata offers a concrete, technical implementation that aligns favorably with current USPTO guidelines. Korea’s IP regime, governed by the Korean Intellectual Property Office (KIPO), similarly emphasizes inventive step and industrial applicability; the cross-modal alignment framework, particularly when tied to clinical diagnostics, may satisfy KIPO’s threshold for technical advancement without requiring direct clinical validation as a precondition. Internationally, the WIPO Patent Cooperation Treaty (PCT) provides a harmonized pathway for global protection, though jurisdictional nuances—such as Korea’s emphasis on industrial application over abstract computational methods—may influence claim drafting and examination outcomes. Collectively, *sleep2vec* exemplifies how cross-modal alignment, coupled with principled scaling laws,

The article *sleep2vec* introduces a novel foundation model leveraging cross-modal alignment to unify heterogeneous nocturnal biosignals, addressing a critical gap in sleep staging and clinical diagnostics. Practitioners should note that this approach may influence patent strategies around multimodal biosignal processing, particularly claims involving cross-modal alignment, sensor heterogeneity, or adaptive weighting of data. Statutory connections may arise under 35 U.S.C. § 101 (abstract ideas) or § 103 (obviousness), where claims hinge on whether the method introduces an inventive concept beyond conventional modeling techniques. Case law like *Alice Corp. v. CLS Bank* or *Diamond v. Diehr* may inform the assessment of patent eligibility for such algorithmic innovations.

The article introduces a significant conceptual shift in AutoML by proposing a multi-agent framework that introduces code-guided, modular, and verifiable architectures, addressing longstanding issues of transparency and runtime failure in traditional black-box AutoML systems. From an Intellectual Property perspective, this innovation could influence patentability considerations, particularly in jurisdictions like the U.S., where software-related inventions are scrutinized under the lens of technical effect and enablement, and in South Korea, where patent eligibility for AI-related inventions is more restrictive due to stringent utility requirements. Internationally, the framework aligns with broader trends in AI governance, encouraging modularity and verifiability as key criteria for innovation assessment, potentially impacting international standards and collaborative research frameworks. The comparative jurisdictional impact underscores the nuanced application of IP protection across different regulatory landscapes.

As a Patent Prosecution & Infringement Expert, I will analyze the article's implications for practitioners in the field of Artificial Intelligence and Machine Learning. **Key Takeaways:** 1. The article presents a novel multi-agent framework, iML, which addresses the limitations of traditional Automated Machine Learning (AutoML) frameworks by introducing code-guided planning, modular implementation, and verifiable integration. 2. The iML framework decouples preprocessing and modeling into specialized components governed by strict interface contracts, reducing the risk of hallucinated logic and logic entanglement. 3. The framework's code-verifiable integration enforces physical feasibility through dynamic contract verification and iterative self-correction, increasing transparency and reliability. **Implications for Practitioners:** 1. The article highlights the need for more transparent and reliable AutoML frameworks, which can be addressed through the development of code-guided, modular, and verifiable architectures. 2. Practitioners can leverage the iML framework's concepts, such as code-guided planning and modular implementation, to improve the reliability and transparency of their own AutoML solutions. 3. The article's emphasis on verifiable integration and dynamic contract verification can inform the development of more robust and reliable AI systems. **Case Law, Statutory, or Regulatory Connections:** The article's focus on transparency, reliability, and verifiability in AI systems may be relevant to the development of regulations and standards in the field of AI. For example, the European Union's Artificial