arXiv cs.LG (Machine Learning)

Vision-Based Runtime Monitoring under Varying Specifications using Semantic Latent Representations

Bardh Hoxha, Oliver Sch”on, Hideki Okamoto…

This paper addresses the demand for past-time signal temporal logic certified runtime monitoring for visual inputs, and proposes two types of reusable monitors that do not require retraining for each formula: a semantic base monitor that covers all target formula fragments with a single conformal calibration, and a rolling prediction monitor adapted for short scenarios. Tests show both methods meet conformal coverage requirements, and the certified bound of the semantic base monitor is 4 times tighter than that of the rolling version in long cycles.

Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders

William Lehn-Schi{\o}ler, Magnus Ruud Kj{\ae}r, Rahul Thapa…

Aiming at the problem that EEG foundation models have excellent clinical performance but opaque internal mechanisms that hinder clinical adoption, this paper uses TopK sparse autoencoders to extract sparse features from three EEG Transformers with different architectures, evaluates monosemy and disentanglement in combination with clinical classification systems, proposes a concept shift selectivity metric, locates three types of representation states and clinical entanglement defects such as age-pathology confusion, and maps hidden layer operations to interpretable EEG spectral features through a spectrum decoder.

Rethinking Molecular OOD Generalization via Target-Aware Source Selection

Zhuohao Lin, Kun Li, Jiameng Chen…

To address the challenges in AI drug discovery: high difficulty in predicting extreme out-of-distribution molecular properties, semantic overlap in existing evaluation benchmarks, and frequent negative transfer in traditional domain adaptation, this study proposes the SCOPE-BENCH benchmark divided by physicochemical clustering, and the POMA framework based on reinforcement learning source selection and dual-scale adaptation. Tests show that the error of existing SOTA models increases by an average of 5.9 times on the new benchmark, and POMA reduces the mean absolute error by 6.2% compared to the baseline.

OpenAI Official Updates

A new personal finance experience in ChatGPT

OpenAI

This new personal finance feature for ChatGPT Pro users in the U.S. is currently in preview. Its core capability allows users to securely link their personal financial accounts, and the system can combine users’ actual financial status, preset financial goals and priority preferences to output customized AI financial analysis insights and provide financial guidance tailored to individual circumstances, offering scenario-based intelligent financial assistance for paid users.

Databricks brings GPT-5.5 to enterprise agent workflows

OpenAI

Recently, big data vendor Databricks announced the deployment of GPT-5.5 for enterprise agent workflow scenarios. Previously, this model set a new state-of-the-art record on OfficeQA Pro, the authoritative benchmark for evaluating complex office tasks, outperforming previous generation models in both accuracy of complex office tasks and completion rate of cross-tool collaboration, which can greatly improve the automation level of enterprise office processes and effectively reduce costs and increase efficiency.

Anthropic News

Introducing Claude Opus 4.7

Anthropic

Anthropic’s latest large model product Claude Opus 4.7 is now officially fully available. Compared with the previous baseline version Opus 4.6, this version has targeted capability upgrades, with core improvements focused on advanced software engineering task processing capabilities, especially in the most difficult task scenarios in this field, the performance improvement is particularly significant, making it more suitable for professional developers’ complex task processing needs.

Introducing Claude Design by Anthropic Labs

Anthropic

Anthropic Labs recently launched the new product Claude Design, which supports users to collaborate with the Claude large model on visual creation, and can produce a variety of polished professional visual works including graphic designs, product prototypes, presentation slides, and single-page promotional materials. This product expands the capability boundary of large models, extending their content production capability from the text field to professional visual design scenarios, which can effectively lower the threshold for visual creation.

Google DeepMind

AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields

Google DeepMind

This paper introduces the application value of the intelligent coding agent AlphaEvolve: its core adopts a dedicated algorithm architecture driven by Google’s Gemini large model, focusing on low-threshold professional code generation capabilities. It has now achieved large-scale cross-domain empowerment, covering needs such as improving R&D efficiency in commercial scenarios, optimizing infrastructure operation and maintenance, and accelerating the implementation of cutting-edge scientific research, providing a new technical path for cost reduction and efficiency improvement across industries.

Enabling a new model for healthcare with AI co-clinician

Google DeepMind

This study focuses on the implementation path of AI-assisted healthcare, with the core goal of developing a new medical collaboration model of “AI co-clinician” and exploring a new medical service system. In this model, AI does not replace clinicians, but acts as a diagnosis and treatment collaboration partner to make up for the shortcomings of manual decision-making, which can improve the scientificity of diagnosis and treatment decisions and consultation efficiency, providing a feasible direction for building a universal and accurate medical service paradigm.

Hugging Face Blog

Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality

Hugging Face

The newly released Granite Multilingual Embedding Model R2 uses the fully open source Apache 2.0 license, has a lightweight parameter scale of less than 100M, supports a 32K ultra-long context window, and is suitable for multilingual semantic representation scenarios. Tests show its retrieval accuracy is the best among embedding models with parameter sizes below 100M currently available, enabling low-cost implementation of cross-language long document retrieval, semantic matching and other tasks, with no commercial use restrictions.

Unlocking asynchronicity in continuous batching

Hugging Face

Currently, only the paper title is provided, and the corresponding English abstract content is not attached, so translation, extraction and summarization work cannot be completed. Please supplement the full English abstract of this “Unlocking asynchronicity in continuous batching” article, and I will highlight the core methods and conclusions as required, and output a concise Chinese summary of about 120 words.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This paper studies the AI alignment problem from the perspective of virtue ethics, refutes the traditional assumption that “rational agents need to be anchored to ultimate goals”, and points out that human rationality stems from the adaptation of behavior to the practice network composed of action tendencies, evaluation standards, etc., rather than pointing to fixed goals. It proposes that to realize AI adaptation to human needs, the AI decision-making logic should match the practical action logic of human beings, taking into account both ethical and safety alignment requirements.

QbitAI

Huawei Cloud INSPIRE Creator Conference theme forum agenda announced: Unveiling new Agentic AI layout

QbitAI

Huawei Cloud’s upcoming INSPIRE Creator Conference will announce its full-stack Agentic AI layout: covering a computing power base optimized for software-hardware collaboration, an industry-adapted one-stop training and inference platform, the newly launched AgentArts enterprise-level intelligent agent development platform with supporting full-lifecycle AI security capabilities. It will also roll out industry zones, partner and university developer programs, building a full-stack technology + ecosystem closed loop to promote large-scale deployment of Agentic AI.

Need is all you need: After AI takes over coding, is this the only most valuable ability left for programmers?

QbitAI

The AI programming track has now shifted from competing on code generation speed to full-link delivery capability from requirement to launch. Alibaba recently released Qoder 1.0, which completes the upgrade from a traditional AI IDE to an autonomous agent development workbench: its core upgrades the Quest function to an independent window with independent task status, supports seamless switching between task delegation and collaborative programming, and can also realize cross-project multi-task parallel processing.

Ronglian Cloud releases “digital employee” level AI Agent platform, reshaping LLM-powered contact centers

QbitAI

At the 2026 China Customer Service Festival, Ronglian Cloud released a new generation of AI Agent smart contact platform. The platform adopts a “single Agent + multiple Skills” architecture, built on an omni-channel CC + CRM base, with three core capabilities: omni-channel access, unified workbench, and Agent-driven full-process autonomy. It can act as an independently responsible “digital employee”, promoting contact centers to shift from passive response to active customer operation and supporting business growth.

---