All Updates
Page 447 of 863
March 20, 2026
Cursor Model Beats Opus 4.6, Prices Slashed
Cursor launched its self-developed AI model surpassing Opus 4.6 on benchmarks. It features a massive price cut, heightening excitement in ambient programming. The upgrade introduces a new reinforcement learning method.
TeachingCoach: AI Chatbot Guides Instructors
TeachingCoach is a fine-tuned chatbot providing real-time, pedagogically grounded guidance to higher education instructors. It uses a data-centric pipeline to extract rules from educational resources and generate synthetic dialogues for fine-tuning. Expert evaluations show it outperforms GPT-4o mini in clarity and responsiveness, with user studies noting depth-efficiency trade-offs.
Skele-Code: No-Code Agentic Workflow Builder
Skele-Code offers a natural-language and graph-based interface for non-technical users to build AI agent workflows via interactive notebooks. Steps convert to code, with agents used only for generation and error recovery, avoiding orchestration costs. It reduces token usage through context-engineering, yielding modular, shareable workflows usable as agent skills.
Retrieval Boosts LLM Agent Generalization
Researchers propose combining fine-tuning and retrieval for LLM agents to achieve better generalization on unseen tasks. They develop a superior LoRA-based SFT recipe and analyze optimal retrieval strategies for storage, querying, and selection. The integrated pipeline outperforms state-of-the-art methods.
Multi-Trait Steering Reveals Harmful AI Interactions
Researchers developed MultiTraitsss framework using subspace steering to create 'Dark models' that simulate harmful human-AI interactions linked to mental health crises. Evaluations confirm these models produce cumulative harmful behaviors over multi-turn conversations. The study proposes protective measures to mitigate such risks in LLMs.
Efficient AI Reliability Amid Error Propagation
AI systems in smart cities face reliability issues from error propagation across stages, challenged by data scarcity, interdependence, and complexity. This paper uses a physics-based AV simulation with error injector to generate data and develops a new framework modeling error propagation. Parameters are estimated via a computationally efficient composite likelihood EM algorithm, proven on AV perception systems.
EDM-ARS Automates EDM Research Pipelines
EDM-ARS is a multi-agent system automating end-to-end educational data mining research with five LLM-powered agents. It generates complete LaTeX manuscripts including validated ML analyses and Semantic Scholar citations. Released open-source with architecture details and a roadmap for expansions like causal inference.
Dynamic Clustering Speeds Dense Crowd Prediction
Proposes a novel cluster-based approach for efficient trajectory prediction in dense crowds by dynamically grouping individuals with similar attributes. This plug-and-play method uses group centroids in place of individual inputs, reducing computational costs and memory usage. It matches state-of-the-art accuracy while enabling faster processing in noisy, massive tracking scenarios.
DEAF Benchmark Exposes Audio MLLMs' Text Reliance
DEAF introduces over 2,700 conflict stimuli across emotional prosody, background sounds, and speaker identity to test Audio MLLMs' acoustic processing. A multi-level evaluation framework disentangles text bias from prompt influence. Seven models show text dominance despite acoustic sensitivity, revealing gaps in genuine audio understanding.
CORE: Robust OOD Detection via Orthogonal Scoring
CORE disentangles confidence and membership signals by decomposing penultimate features into orthogonal subspaces: classifier-aligned confidence and residual. It scores each subspace independently and combines via normalized summation for robust OOD detection. Achieves SOTA performance across five architectures and benchmarks with negligible overhead.
Continually Self-Improving AI
This arXiv paper identifies three key limitations of modern LLMs: data-inefficient knowledge acquisition, reliance on finite human data, and human-confined training pipelines. It proposes synthetic data amplification for efficient updates from small corpora, self-generated data to bootstrap pretraining without distillation, and test-time search over algorithm spaces. These steps aim to enable continually self-improving AI systems.
ADM: Efficient Training for Geometric & Neuromorphic AI
This arXiv paper introduces Adaptive Domain Models (ADM), a novel training architecture using posit arithmetic, type systems, and hypergraphs for memory-efficient, grade-preserving AI training. It enables Bayesian distillation to bootstrap domain-specific models from general ones and warm rotation for interruption-free updates. Applicable to geometric and neuromorphic AI with verifiable correctness.
Access Control for Agentic AI Websites
Researchers address gaps in delegating critical tasks to agentic AI via websites due to poor access controls. They propose a fine-grained access control design, including website implementation and modifications to open-source authorization protocols. Evaluation confirms effective AI agent usage.
Alibaba Shares Sink on Earnings Reality Check
Alibaba's shares sank following its earnings report, described as a 'reality check' by Macquarie analyst Ellie Jiang. Despite challenges, she views Alibaba as best positioned for AI advancements. The company faces scrutiny amid broader market pressures.
Tencent Maps V11 AI for Multi-Person Trips
Tencent Maps released major V11.0 version on March 19, targeting multi-person travel. It launches AI one-stop itinerary management for planning, sharing, and recording before, during, and after trips. Focuses on high-frequency group outing scenarios.
Huawei Bridges Digital Divide with Connectivity
Huawei is committed to global connectivity to bridge the digital divide, as AI and digital technologies advance rapidly. Over 2.2 billion people remain offline, missing essential online services like education and finance. This initiative aims to connect the unconnected worldwide.
AI Short Drama Explodes Globally
AI-generated short drama '้ๅป็ ' produced in 5 days at low cost garnered 500M plays and spread overseas. Chinese short dramas hit $10.88B revenue in H1 2025, up 249%, but face rising costs and low 10-15% hit rates. Platforms rely on heavy ad spends amid intense competition.
CATL HK Rally Sets Record Premium
Shares of Contemporary Amperex Technology Co. briefly traded at a record premium in Hong Kong over mainland-listed peers. Strong earnings and energy supply shocks boosted demand for the battery giant.
Delton Shares Jump 106% HK Debut
Delton Technology Guangzhou Inc.โs shares rallied 106% in their Hong Kong trading debut. The Chinese circuit board maker raised HK$3.3 billion ($421 million) via the listing.
ChiNext Index Rockets Over 3%
ChiNext index surges over 3%, Shanghai Composite rises 0.12%, Shenzhen Component up 1.36%. Photovoltaic, energy storage, power, and compute hardware sectors lead gains across nearly 2000 rising stocks.