arXiv cs.LG (Machine Learning)

Temporal Contrastive Transformer for Financial Crime Detection: Self-Supervised Sequence Embeddings via Predictive Contrastive Coding

Danny Butvinik (NICE Actimize), Yonit Marcus (NICE Actimize), Nitzan Tal (NICE Actimize)…

This paper proposes a Temporal Contrastive Transformer (TCT) adapted for financial transaction sequences, which uses self-supervised contrastive learning to train temporal behavior embeddings to support downstream anti-fraud tasks. Experiments show that the embeddings it generates alone achieve a prediction AUC of 0.8644, and can automatically learn domain features close to manually designed ones. However, combining it with manual features does not bring performance gains, and it is still weaker than the strong manual feature baseline at present, providing a feasible direction for reducing the reliance on feature engineering in financial anti-fraud.

Teaching Language Models to Forecast Research Success Through Comparative Idea Evaluation

Srujan P Mule, Aniketh Garikaparthi, Manasi Patwardhan

To address the bottleneck of massive AI-generated research ideas and high experimental verification costs, this study built a dataset of more than 11,000 pairs of research ideas with real implementation results, and trained an 8B parameter small model through supervised fine-tuning and reinforcement learning with verifiable rewards. The fine-tuned model reaches an accuracy of 77.1%, outperforming GPT-5’s 61.1%; the reinforcement learning version balances 71.35% accuracy and interpretability, with excellent robustness, and can be used as a low-cost pre-screening tool for research ideas.

The Attribution Impossibility: No Feature Ranking Is Faithful, Stable, and Complete Under Collinearity

Drake Caraker, Bryan Arnold, David Rhoads

This explainable AI study proves that no feature attribution ranking can meet the requirements of faithfulness, stability, and completeness at the same time in scenarios with feature collinearity. There are only two types of methods with no intermediate state: “faithful and complete but unstable” and “stable but requiring reporting of symmetric feature ties”. The researchers proposed the Pareto-optimal DASH ensemble attribution method, completed the first formal verification of an impossibility conclusion in the XAI field through Lean4, and also found that 68% of public datasets have attribution instability, and SHAP fairness audits are unreliable under collinearity.

OpenAI Official Updates

OpenAI named a Leader in enterprise coding agents by Gartner

OpenAI

Recently, authoritative consulting firm Gartner released its 2026 Magic Quadrant report for enterprise AI coding agents, and OpenAI has been named a Leader. Its core coding large model Codex has been officially recognized for its outstanding technological innovation and mature enterprise-level large-scale deployment capabilities. This rating also confirms OpenAI’s dual leading position in technology and commercialization in the ToB AI development tool track.

How Virgin Atlantic ships faster with Codex

OpenAI

This case sorts out Virgin Atlantic’s AI development practices: To launch the revised mobile app before the fixed deadline of the holiday travel season, the team introduced the AI programming tool Codex to support development work. In the end, it not only delivered smoothly on schedule, but also achieved nearly 100% unit test coverage, with no highest priority (P1) defects after launch, verifying the efficiency improvement value of AI programming tools for high-demand commercial projects.

Anthropic News

Introducing Claude Opus 4.7

Anthropic

Anthropic’s latest large model Claude Opus 4.7 is now officially fully available. As an upgraded version of the previous generation Opus 4.6, the core optimization direction of this model is high-level software engineering capabilities, with particularly significant improvements in the most difficult R&D tasks such as complex code development and troubleshooting of difficult problems, which can better meet the high-intensity technical needs of professional developers.

Introducing Claude Design by Anthropic Labs

Anthropic

Anthropic Labs has newly launched the product Claude Design, whose core feature is supporting users to create collaboratively with the Claude large model. This tool can produce various types of professional-level visual outputs, covering scenarios such as graphic design, product prototypes, presentation slides, and single-page promotional materials. Users do not need to be proficient in complex design tools to efficiently obtain well-structured, mature visual works, lowering the threshold for creating high-quality visual content.

Google DeepMind

We’re launching the Google DeepMind Accelerator program in Asia Pacific to tackle environmental risks

Google DeepMind

Google DeepMind recently launched the Asia Pacific Accelerator program, which is open for applications from science and innovation teams in the Asia Pacific region focusing on environmental issues. Selected teams will receive resource support including exclusive DeepMind AI technology authorization, free computing quota, and one-on-one guidance from domain experts, aiming to rely on AI technology to tackle challenges in scenarios such as extreme weather warning, carbon emission accounting, and ecological restoration, and systematically reduce various environmental risks in the Asia Pacific region.

Fast-tracking genetic leads to reverse cellular aging

Google DeepMind

This study aims to accelerate the discovery of genetic targets for reversing cellular aging. Biologists carried out screening with the help of the “Co-Scientist” research assistance tool, and successfully discovered a new regulatory factor, which has been verified to effectively rejuvenate human cells. This achievement greatly shortens the R&D cycle of anti-aging targets, providing a new research direction for the subsequent development of aging intervention programs and anti-aging drugs.

Hugging Face Blog

Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models

Hugging Face

This research from NVIDIA’s Nemotron Labs focuses on high-speed text generation. Targeting the pain point of high per-token inference latency of autoregressive models, it optimizes the modeling logic of discrete text diffusion models, greatly reduces the number of sampling iteration steps, and realizes full-sequence parallel generation. Tests show that under the premise that its generation quality matches that of autoregressive models with the same parameters, the inference speed is increased by dozens of times, marking a key step towards light-speed text generation.

Specialization Beats Scale: A Strategic Variable Most AI Procurement Decisions Overlook

Hugging Face

This paper addresses the common misconception in current AI procurement that model parameter scale is given priority, pointing out that “scenario specialization level” is a core decision variable that has been ignored for a long time. Through empirical comparison across multiple business scenarios, specialized small and medium-sized models adapted to segmented needs outperform large-parameter general models in the same field in terms of accuracy, inference efficiency, deployment cost, and data security. It is recommended that procurement prioritize scenario adaptability rather than blindly pursuing large scale.

The Gradient

After Orthogonality: Virtue-Ethical Agency and AI Alignment

The Gradient

This AI alignment study from the perspective of virtue ethics challenges the goal-oriented rationality hypothesis: it proposes that human rationality is not anchored to final goals, but adjusts behavior to adapt to a self-reinforcing practice network composed of actions, tendencies, evaluation criteria, and resources. The paper argues that to achieve alignment between AI and human collaboration, compliance, and even basic safety, the AI decision-making logic needs to match this practice-oriented action paradigm of humans.

QbitAI

Former head of Meituan Waimai enters catering embodied model sector, Yuanjie Intelligence secures tens of millions in seed round financing

QbitAI

Yuanjie Intelligence, an embodied intelligence company founded by Wang Dong, former technical head of Meituan Waimai (disciple of Academician Zhang Bo), recently completed a seed round of financing worth tens of millions of yuan. The team did not follow the trend of the general humanoid robot track, but anchored the restaurant back kitchen scenario with higher commercialization certainty, focusing on solving pain points such as meal handover errors and omissions, and industry labor shortages. It has obtained cooperation intentions from multiple leading enterprises, and the financing will be used for core product R&D and rollout.

Too hard to raise lobsters? Zhou Hongyi built a cloud office for lobsters, with professional trainers to train lobsters online

QbitAI

At present, the Agent track is seeing rapid framework iteration, but a wave of user abandonment is prominent, with core pain points of high threshold, high cost, and insecurity. 360 has launched the Secure Lobster Cloud version and Lobster Coach: the former provides a full set of cloud resources, enabling continuous operation without running on local devices; the latter handles complex operations such as agent training and process optimization, complementing the infrastructure for Agent implementation, greatly lowering the usage threshold, and when used properly, it can be equivalent to a human team.

Fei-Fei Li makes another move, the ImageNet for spatial intelligence is here

QbitAI

Fei-Fei Li’s team released ESI-Bench, the world’s first embodied spatial intelligence evaluation benchmark with a closed perception-action loop, breaking the previous evaluation logic of passively providing images, forcing AI to take the initiative to obtain information to answer questions. More than 3,000 evaluation tasks are designed around the four core spatial cognitive abilities of humans. Actual tests show that current AI only has outstanding visual capabilities, and there is still a large gap from real spatial intelligence capable of active exploration.

---