arXiv cs.LG

Recursive transformers for semiconductor thermo-mechanical reliability

Kart-leong Lim

To address the issues of parameter redundancy in traditional Transformer surrogate models, which are prone to overfitting and computing power waste in engineering scenarios where large amounts of simulation data are difficult to obtain, this paper evaluates three types of recursive Transformers (including a self-developed deep recursive version), compares their performance, parameter count and computing power consumption, and provides a model selection guide for resource-constrained scenarios. Verified by advanced semiconductor packaging thermo-mechanical reliability analysis and Laplace PDE solving, recursive weight-sharing Transformers can balance accuracy, parameter efficiency and computing cost in small-data engineering surrogate modeling.

Regularizing modality contribution drift in multimodal continual learning

Zhen Zhang, Jielei Chu, Bin Liu…

Aiming at the flaw of existing multimodal continual learning solutions that ignore cross-task modality relative contribution drift (MCD), this study quantifies MCD evaluation metrics, proposes the CMCDR regularization method adapted to scenarios with or without old samples. It preserves historical modality contribution structures by comparing old sample probes and distilling the contribution distribution of old models from current samples respectively, and its universality and effectiveness are verified on multimodal class incremental and continual visual question answering tasks.

DoTime: A Synthetic Benchmark Generator for Interventional and Counterfactual Time Series

Dennis Thumm, Billy Tim Anthony, Ying Chen

To address the problem that existing time series causal inference benchmarks are mostly observational, small-scale, domain-specific, and lack support for intervention and counterfactual evaluation, this paper launches the open-source, extensible DoTime generator, which can generate multivariate time series structural causal models with interventions, adds unique capabilities such as continuous intervention windows and mechanism switching, and is equipped with supporting benchmark suites and baseline tools. Experiments verify that with the same model capacity, models trained with interventions have significantly better direction accuracy than purely observation-trained models.

OpenAI

Advancing responsible AI across Europe

OpenAI

This is OpenAI’s practice explanation on responsible AI governance in Europe: Currently, OpenAI has built a compliance system from four dimensions: system security protection, operation security control, technical transparency construction, and data source traceability, to meet EU AI regulatory requirements. As the EU AI Act legislation progresses, OpenAI will continue to iterate on relevant practices to support the implementation of responsible AI governance in Europe.

Building abundant intelligence

OpenAI

Titled Building Abundant Intelligence, this research targets the pain points of current advanced AI: limited capability ceiling, high implementation cost, and narrow applicable scenarios. It proposes a full-stack optimization technical route to simultaneously realize AI capability leaps and lower deployment costs, covering more differentiated application needs. The goal is to break the resource barrier of high-level AI, and promote advanced intelligence to transform from a scarce service to an inclusive and accessible public resource.

Anthropic News

Introducing Claude Opus 5

Anthropic

Claude Opus 5 is a major iterative version of Anthropic’s high-end Opus large model product line, delivering step-function performance upgrades. This version specifically optimizes long-term operation support capabilities, which can stably adapt to the long-cycle operation needs of Agent-based intelligent agents; it also significantly improves processing performance for programming and professional domain tasks, better meeting the needs of complex development and professional scenarios across various industries.

Inviting hard questions

Anthropic

Titled Inviting hard questions, this announcement’s core initiative is to openly solicit all kinds of difficult and sharp questions about the AI field from the public across society. The research team promises that in the entire process of responding to and answering these questions, all work details of the demonstration research will be fully disclosed, with a transparent and traceable process, so as to respond to public concerns about AI development and promote AI research to adapt to public demands.

Google DeepMind

Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration

Google DeepMind

Gemini’s dedicated robotics technology solution Gemini Robotics ER 2 has achieved step-function technological breakthroughs in three major directions: first, upgraded video understanding capabilities for robotics scenarios, second, optimized task tool orchestration and scheduling capabilities, and third, innovative multi-robot collaboration mechanisms. This solution can support robots to complete autonomous reasoning and multi-robot cooperation, implement and solve various practical tasks in real scenarios, providing a new technical foundation for robotics applications.

We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control

Google DeepMind

Google recently officially launched the new generation AI music generation model Lyria 3.5 on its Flow Music platform. The core upgrades of this version cover four major directions: the overall musical coordination of generated works is greatly improved, the accuracy of lyric creation and the naturalness of vocal quality are both significantly optimized, while user creative control is further strengthened, which can adapt to the differentiated music generation needs of ordinary users and professional creators.

Hugging Face Blog

GPU Management: Why Idle GPUs Are the New Grounded Aircraft

Hugging Face

This paper analogizes idle GPUs to civil airliners that lose money as soon as they are grounded, pointing out that the current average idle rate of GPUs in global data centers exceeds 30%, with core causes including coarse scheduling granularity, video memory fragmentation, and task adaptation mismatch. The paper proposes optimization schemes of fine-grained time-sharing scheduling, dynamic shared video memory pool, and heterogeneous task co-location, which can increase GPU utilization by more than 60% and reduce computing power costs by 40%, highlighting the urgent value of fine-grained GPU management.

The OlmoEarth Platform: Geospatial inference at planetary scale

Hugging Face

Only the title of this paper is provided for now, the specific text of the abstract is missing, so the request for translation and summarization cannot be completed~ Please supplement and upload the full English content of the abstract, and we will sort out its core methods and conclusions as required, and output a concise and clear summary of about 120 words for you.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This paper focuses on the AI alignment problem, starting from the perspective of virtue ethical agency, refuting the traditional assumption that “rational agents need to be oriented towards ultimate goals”: it points out that human rationality is embodied in actions adapting to the practical network composed of behavior patterns, evaluation criteria, resources, etc., rather than pointing to fixed goals. It argues that the decision-making logic of rational AI needs to match human practical logic, which is not only a requirement for ethical alignment, but also related to the core security attributes of AI.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

This paper sorts out the evolution of the concept of Recursive Self-Improvement (RSI): In 1965, I.J. Good first proposed the idea of ultra-intelligent machines, referring to systems that can surpass all human intellectual activities and independently design better machines; in 2008, Eliezer Yudkowsky clarified that the core of RSI is the feedback loop where AI iterates and optimizes its own cognitive architecture relying on existing intelligence. Current RSI in the AI field includes both models directly rewriting their own weights, and can also broadly cover the behavior of models optimizing their own training pipelines.

量子位 QbitAI

SIGGRAPH时间检验奖揭晓：这项研究，提前十年押中了物理AI

量子位 QbitAI

The 2026 SIGGRAPH Test of Time Award was awarded to the 2016 character motion generation research of the team led by University of Hong Kong professor Taku Komura. At a time when AI was still focused on the image field, this research was the first to apply deep learning systems to 3D motion generation. It uses convolutional autoencoders to learn low-dimensional motion spaces from motion capture data, and can generate and edit natural humanoid motions according to high-level instructions, laying an early foundation for the current core direction of physical AI.

刚刚，即梦 Seedance 2.5来了！我狂测测测测……

量子位 QbitAI

ByteDance recently officially released its self-developed AI video model Seedance 2.5, which has been simultaneously launched on its official experience portal Jimeng AI. The model can natively output 30-second CG-quality videos, supports generation of up to 3 minutes long, and has capabilities including local editing, reference of up to 50 materials, and timestamp control with accuracy within 1 second. It solves the previous pain points of AI videos being too short and having poor controllability, and adapts to the professional creation needs of film and television, advertising, and gaming.

视频后期，危！MiniMax H3手绘即特效，多模态的「Coding时刻」来了

量子位 QbitAI

MiniMax released its first open-source video model H3, breaking the previous industry paradigm where AI videos can only generate materials. It integrates end-to-end full-process capabilities including editing logic, subtitle layout, transitions, soundtrack, hand-drawn effects, etc., and can directly generate publishable finished products with 2K resolution from text input. Currently, H3 tops the global video editing leaderboard, is recognized as SOTA and open-weight, shaking the inherent perception that post-production is an exclusive human moat.