All Updates
Page 1181 of 1480
March 12, 2026
IH-Challenge Boosts LLM Instruction Hierarchy
IH-Challenge is a new reinforcement learning dataset designed to improve instruction hierarchy in frontier LLMs, enhancing defenses against jailbreaks and prompt injections. Fine-tuning GPT-5-Mini on it yields +10% robustness across benchmarks, slashes unsafe behavior to 0.7%, and maintains helpfulness. The dataset is released on Hugging Face for further research.
HyMEM Supercharges GUI Agents
HyMEM is a graph-based memory system for GUI agents that combines discrete symbolic nodes with continuous trajectory embeddings, inspired by human memory. It enables multi-hop retrieval, self-evolution through node updates, and dynamic working-memory refreshing. Experiments demonstrate it boosts Qwen2.5-VL-7B by +22.5%, matching or surpassing Gemini 2.5 Pro Vision and GPT-4o.
HEAL Breaks Teacher Ceiling in Reasoning Distillation
HEAL is an RL-free framework that distills reasoning from Large Reasoning Models to smaller ones by overcoming rejection sampling limits and the 'Teacher Ceiling'. It combines GEAR for entropy-guided trajectory repair, PURE for filtering genuine breakthroughs, and PACE for progressive curriculum learning. Experiments show superior performance over SFT and baselines on multiple benchmarks.
FAME: Scalable Minimal NN Explanations
FAME proposes formal abstract minimal explanations for neural networks using abstract interpretation, scaling to large models while minimizing explanation size. It employs dedicated perturbation domains and LiRPA-based bounds to discard irrelevant features without needing traversal order. Benchmarks show superior explanation size and runtime over VERIX+.
Agentic Center Automates Data Product Optimization
Researchers propose Agentic Control Center, a system using specialized AI agents to automate data product improvements in a continuous optimization loop. It surfaces questions, monitors multi-dimensional quality metrics, and incorporates human-in-the-loop controls. This transforms data into observable, refinable assets balancing automation with trust.
Calterah Raises $140M+ in Series E
Calterah, a mmWave radar CMOS chip specialist, completed over 1 billion RMB in Series E funding. The round attracted national industrial funds, local government funds, and industrial investors. Huaxing Securities acted as the exclusive financial advisor.
NVIDIA AI-Q Tops DeepResearch Benches I & II
NVIDIA's AI-Q model has achieved the #1 ranking on both DeepResearch Bench I and II. The Hugging Face blog details how this superior performance was attained. This marks a significant benchmark breakthrough for NVIDIA.
US Iran Strikes Herald AI War Dawn
US strikes on Iran under Trump are touted as the prelude to 'AI war' due to AI's full-chain integration enabling rapid victories. Anthropic and Palantir emerge as key players in this narrative. The piece spotlights Palantir's role in intelligence.
Craveable Brands Eyes AI Despite Failure Risks
Craveable Brands is cautiously exploring AI adoption while monitoring high project failure rates. The company hopes tech partners will help de-risk implementation. This reflects common enterprise concerns in AI rollout.
UBTech Beyond the Dance
UBTech, famous for its dancing robots on the Spring Festival Gala, signals deeper ambitions. The company contemplates strategies beyond entertainment performances. More reflections on its future direction.
GitHub Feb 2026: Six Service Incidents
In February 2026, GitHub faced six incidents causing degraded performance across its services. The availability report summarizes these outages. It was posted on the GitHub Blog.
Social Sciences' First AI Crisis
Vibe Research's essay explores the inaugural AI crisis in social sciences, entered in ๅฐๆฐๆดพ's 2025 annual contest under #TeamSilicon25. The contest innovates with 'only use AI' and 'no AI' tracks for creative submissions. It reflects the author's views on AI's disruptive impact.
RTX PRO 6000 MoE Benchmark: 50.5 tok/s Max
8+ hour benchmark of MoE backends for Qwen3.5-397B NVFP4 on 4x RTX PRO 6000 yields 50.5 tok/s with Marlin TP=4. NVIDIA CUTLASS kernels fail on SM120 due to initialization bugs, forcing FP16 fallback. MTP regresses performance by 22%.
Jiangyin 500M RMB AI Fund Launches
Jiangyin Guolian Xizhou AI Industry Equity Fund (LP) recently registered with 500 million RMB capital. Executed by Wuxi Guolian Xizhou Private Fund Management Co., it focuses on equity investments and asset management. Key partners include Wuxi Guolian Xizhou PE Fund, Jiangyin Qilian Industry Fund, and Jiangyin Huigang Investment.
CAICT Launches Claw Agent Standards
China Academy of ICT has started compiling Claw intelligent assistant agent series standards to build a robust framework. They seek industry units and experts for 'Claw Product Trustworthy Capability Requirements'. The standard covers user permissions, execution transparency, behavior risks, and platform trustworthiness.
Magic Atom Raises 10.5B for Embodied AI
Magic Atom has secured 10 billion RMB in fundraising and 500 million RMB in financing, totaling 10.5 billion RMB. The funds target the endgame of embodied intelligence. This positions Magic Atom as a prime sample for commercializing embodied AI.
Oil Prices Surge, EV Costs Rise on Memory & Lithium
China's oil prices rise 4th straight time due to Middle East tensions, prompting gas station rushes. EV owners face price hikes from surging memory chip and lithium carbonate costs, affecting brands like Zeekr and NIO.
Seagate: Iran War Spares AI Supply Chain
A Seagate executive states the Middle East conflict, including Iran war, won't significantly disrupt the technology supply chain short-term. This addresses fears over vital materials like helium. AI infrastructure remains unaffected for now.
Ex-Huawei AI Chief Launches Beta Infinity with ยฅ100M Seed
Former Huawei AI leader Liu Wulong founded Beta Infinity, targeting consumer-grade embodied robots with personalization and autonomous evolution. The startup raised nearly 100M RMB in seed funding from Hongtai Fund, Zhengjing Fund, and others. Backed by elite team from Huawei, ByteDance, and DJI, it's partnering with top industrials for quick market entry.
AI: New Species, Not Actor Replacement
Director Zhao Bao Gang views AI as a new species with unique emotional logic, not a human actor substitute, arriving in long dramas within 2-5 years. AI slashes short drama costs by 70% and flattens industry pyramids into hourglasses by automating mid-level roles. Future AI content may evolve beyond traditional viewing into immersive experiences.