Intellectual Property

### **Jurisdictional Comparison & Analytical Commentary on CRIT’s Impact on Intellectual Property Practice** The introduction of **CRIT**—a graph-based dataset designed to enhance cross-modal multi-hop reasoning in Vision-Language Models (VLMs)—raises significant **intellectual property (IP) considerations** across jurisdictions, particularly regarding **data ownership, copyright in AI-generated content, and liability for AI hallucinations**. In the **U.S.**, where AI-generated works face restrictive copyright protections (as seen in *Thaler v. Perlmutter*), CRIT’s synthetic data pipeline may trigger debates over **who owns the training data**—the researchers, the automated pipeline, or the underlying sources. **South Korea**, under its more permissive stance (e.g., allowing copyright in AI-generated works if human creativity is involved), may view CRIT’s manually verified test set as protectable, but disputes over **derivative works** could arise if CRIT’s outputs are used to train commercial VLMs. **Internationally**, under the **Berne Convention**, CRIT’s synthetic data may lack protection if deemed purely machine-generated, but jurisdictions like the **EU (under the AI Act)** could impose **liability frameworks** for AI-driven hallucinations, shifting enforcement risks to model developers rather than dataset creators. The broader implication is that **CRIT’s release may accelerate regulatory scrutiny** on AI training data provenance, potentially leading to **mandatory disclosure of

As a Patent Prosecution & Infringement Expert, I'll provide domain-specific expert analysis of the article's implications for practitioners. **Technical Analysis:** The article discusses a novel dataset and benchmark, CRIT, designed to evaluate the performance of Vision-Language Models (VLMs) in cross-modal multi-hop reasoning tasks. The CRIT dataset is built using a graph-based automatic pipeline to generate complex cross-modal reasoning tasks, addressing the limitations of existing multimodal benchmarks. This dataset and benchmark are likely to be used to evaluate the performance of VLMs in various applications, such as image and text analysis, and may lead to the development of new and improved VLMs. **Patent Implications:** The CRIT dataset and benchmark may have implications for patent applications related to VLMs and multimodal reasoning. For example, patent claims directed to VLMs may need to be revised to account for the limitations of existing multimodal benchmarks and the improved performance of VLMs on the CRIT dataset. Additionally, the CRIT dataset may be used as prior art to challenge the novelty and non-obviousness of patent claims directed to VLMs. **Case Law and Regulatory Connections:** The CRIT dataset and benchmark may be connected to the following case law and regulatory issues: * The CRIT dataset and benchmark may be relevant to the discussion of obviousness and non-obviousness in patent law, particularly in the context of artificial intelligence and machine learning inventions (e.g

### **Jurisdictional Comparison & Analytical Commentary on *HippoCamp* and Its IP Implications** The *HippoCamp* benchmark highlights critical gaps in AI-driven contextual reasoning, particularly in handling personal data—an area with significant implications for **data privacy, copyright, and trade secret protections** across jurisdictions. In the **U.S.**, where IP frameworks like the *Computer Fraud and Abuse Act (CFAA)* and *Defend Trade Secrets Act (DTSA)* govern unauthorized access to personal or proprietary systems, the benchmark’s findings on "long-horizon retrieval" and "cross-modal reasoning" could raise concerns about **unintended data exfiltration** by AI agents, potentially triggering liability under **copyright scraping laws (e.g., *Authors Guild v. Google*)** or **trade secret misappropriation claims**. **Korea’s** approach under the *Personal Information Protection Act (PIPA)* and *Unfair Competition Prevention Act (UCPA)* would similarly scrutinize AI agents’ handling of personal files, with stricter penalties for **unauthorized data processing**—though Korea’s **AI ethics guidelines** may encourage proactive compliance. At the **international level**, the *EU AI Act* and *GDPR* would impose **high-risk AI system obligations**, requiring transparency in training data sources and user consent for personal file interactions, while the *WIPO Copyright Treaty* and *TRIPS Agreement*

### **Expert Analysis of *HippoCamp* for Patent Practitioners** 1. **Technical & Legal Implications for AI/ML Patents** The *HippoCamp* benchmark highlights critical deficiencies in **multimodal large language models (MLLMs)** and **agentic systems** when handling **personal file management**, particularly in **long-horizon retrieval** and **cross-modal reasoning**. This aligns with existing patent trends in **AI-driven file systems** (e.g., USPTO Class 707/100-707/104 for database/file management) and **context-aware computing** (e.g., USPTO Class 706/46-706/50 for AI reasoning). The benchmark’s focus on **real-world personal data** (42.4 GB across 2K+ files) may raise **privacy and data security concerns**, potentially intersecting with **GDPR, CCPA, or trade secret protections** if such systems are commercialized. 2. **Potential Patentability & Prior Art Considerations** The benchmark’s **failure diagnosis framework** (46.1K structured trajectories) could inform **patent claims** related to **AI error correction, adaptive retrieval, or multimodal fusion techniques**. However, prior art in **personal file management agents** (e.g., USPTO 10,984,123 B

### **Analytical Commentary: Impact of ICML 2026 Tutorials on Intellectual Property (IP) Practice** The ICML 2026 Tutorials announcement highlights the evolving nature of academic and industry collaboration in machine learning (ML), raising important IP considerations regarding **open-access dissemination, proprietary knowledge protection, and collaborative innovation frameworks**. While the conference promotes **open educational resources (OER) and community-driven learning**, jurisdictions like the **US (patent-first approach), South Korea (balanced innovation policy), and international regimes (TRIPS/WIPO)** differ in how they balance **public disclosure (prior art) against patentability, trade secret protection, and collaborative R&D incentives**. The tutorial format itself—whether it involves **invited experts, community submissions, or rigorous peer review**—may influence **IP ownership of derivative works, licensing models, and the enforceability of open-source commitments**, particularly in cross-border collaborations where **Korean "creative commons" policies, US Bayh-Dole Act implications, and WIPO’s open-access principles** may lead to divergent legal interpretations. #### **Key Jurisdictional Comparisons:** 1. **United States:** - The **Bayh-Dole Act** encourages patenting of federally funded research, but ICML’s open-access policy may conflict with institutional IP policies if tutorials derive from patented work. - **Trade secret risks** arise if invited speakers disclose proprietary techniques without formal ND

### **Domain-Specific Expert Analysis for Patent Practitioners** This article highlights the **ICML 2026 Tutorials** selection process, emphasizing **rigorous peer review, community input, and expert-led instruction**—key considerations in **patent prosecution strategy**, particularly for **software and AI-related inventions**. The structured review process (invited, community-sourced, and peer-reviewed) mirrors the **USPTO’s subject matter eligibility (35 U.S.C. § 101) and obviousness (35 U.S.C. § 103) analyses**, where examiner discretion and prior art play pivotal roles. **Case Law & Regulatory Connections:** - **Alice Corp. v. CLS Bank (2014)** – The USPTO’s **Step 2B (inventive concept)** analysis aligns with ICML’s emphasis on **non-artificial evaluation** of tutorial proposals, ensuring substantive contributions beyond routine practice. - **USPTO’s 2019 Revised Patent Subject Matter Eligibility Guidance (PEG)** – The **machine learning tutorial topics** (e.g., numerical optimization, probabilistic numerics) must demonstrate **technological improvement** to overcome § 101 rejections, similar to how ICML evaluates **novelty and practical utility** in submissions. **Strategic Implications for Patent Practitioners:** 1. **Drafting AI/ML Patent Claims**

### **Jurisdictional Comparison & Analytical Commentary on the Impact of *IRENE* on Intellectual Property Practice** The development of *IRENE* (Information Bottleneck-guided EEG SeizuRE DetectioN via SElf-Supervised Learning) presents significant implications for **patent eligibility, trade secret protection, and data governance** across jurisdictions. In the **U.S.**, where patent eligibility under *Alice* and *Bilski* often hinges on whether AI-driven medical diagnostics are deemed "abstract" or "technological," *IRENE*’s structured approach to EEG graph optimization may strengthen patent claims by emphasizing its **specific application in seizure detection** rather than mere algorithmic abstraction. **South Korea**, under the *Patent Act* and *Unfair Competition Prevention Act*, would likely favor trade secret protection for proprietary EEG processing techniques, given its robust enforcement of trade secrets (e.g., Samsung’s past litigation over AI chip designs). At the **international level**, under the **TRIPS Agreement** and **WIPO’s AI guidelines**, patentability would depend on whether jurisdictions classify *IRENE* as a **technical solution** (favored in Europe under the EPO’s AI patent guidelines) or a **medical method** (often excluded in many jurisdictions). The rise of **self-supervised learning in healthcare** also raises **GDPR/PIPL compliance issues**, as EEG data

### **Expert Analysis of Patent & IP Implications for Practitioners** #### **1. Patentability & Novelty Considerations** The paper introduces **IRENE**, a novel framework combining **Information Bottleneck (IB) theory** with **self-supervised graph learning** for EEG-based seizure detection. Key novel aspects include: - **Dynamic graph denoising** tailored to EEG noise characteristics (unlike prior art relying on statistical correlations or predefined similarity measures). - **Joint optimization** of graph structure and spatial-temporal representations under the IB principle. - **Graph Masked AutoEncoder (GMAE)** for structure-aware signal reconstruction. **Potential patent claims** could focus on: - A **method for EEG seizure detection** comprising steps of: - Constructing a denoised dynamic graph from EEG signals using an IB-guided approach. - Applying a self-supervised GMAE to learn compact representations. - A **system** comprising EEG sensors, a processor, and memory storing executable instructions for the claimed method. **Prior Art Risks:** - **Graph-based EEG processing** (e.g., dynamic functional connectivity graphs) is well-known (e.g., *Bashivan et al., 2016*). - **Self-supervised learning in EEG** (e.g., contrastive learning for seizure detection) has been explored (*Tang et al., 2021*). - **Information Bottleneck in neural networks

The paper "Manifold Generalization Provably Proceeds Memorization in Diffusion Models" offers a fascinating theoretical lens through which to understand the generative capabilities of diffusion models, particularly their ability to produce novel outputs even with "coarse" training. This insight has significant implications for intellectual property (IP) practice, particularly in the realm of copyright and inventorship, by refining our understanding of what constitutes "originality" and "creation" in AI-generated content. The core argument – that diffusion models capture the *geometry* of data rather than merely memorizing its *distributional structure* – directly challenges the simplistic notion that AI models are merely sophisticated copy machines. If a model is indeed learning underlying manifold regularities and generating outputs based on these learned geometric principles, rather than reproducing specific training data points, it strengthens the argument for the *originality* of AI-generated works. This theoretical underpinning could influence how courts and IP offices assess copyrightability, potentially shifting the focus from direct input-output comparisons to the sophistication of the generative process and the novelty of the resulting output. **Jurisdictional Comparisons and Implications Analysis:** The implications of this research diverge across jurisdictions, reflecting their varying approaches to AI and IP. * **United States:** The U.S. Copyright Office (USCO) has, to date, taken a relatively restrictive stance, emphasizing the need for human authorship in AI-generated works. The USCO's current guidance suggests that works "produced by a machine

This article's findings regarding diffusion models' ability to generate novel samples from coarse scores, by capturing data geometry rather than fine-scale distribution, has significant implications for patent practitioners in the AI/ML space. **Implications for Practitioners:** 1. **Claim Scope and Enablement (35 U.S.C. § 112):** The concept of "coarse scores capturing the geometry of the data" suggests that claims directed to AI models might be enabled even if they don't explicitly define every fine-grained parameter or training detail. If the core innovation lies in *how* the model learns and leverages data geometry for generalization, rather than precise density estimation, then broad claims focusing on this geometric learning could be defensible. Conversely, if an inventor claims a specific "fine-scale distributional structure," but the underlying model operates on coarse geometric principles, the claim might lack adequate written description or enablement for the *actual* invention. This connects to cases like *Ariad Pharmaceuticals, Inc. v. Eli Lilly and Co.* regarding written description, and *The Medicines Co. v. Hospira, Inc.* on enablement, where the specification must teach one of ordinary skill in the art how to make and use the invention without undue experimentation. 2. **Inventive Step/Non-Obviousness (35 U.S.C. § 103):** The article highlights that this generalization behavior "is a phenomenon not accounted