All Updates
Page 1037 of 1587
March 31, 2026
Survey of Uncertainty-Aware XAI
This arXiv paper surveys uncertainty-aware explainable AI (UAXAI), detailing how uncertainty is integrated into explanations via Bayesian, Monte Carlo, and Conformal methods. It outlines strategies like trustworthiness assessment, model constraining, and uncertainty communication. The work critiques fragmented evaluations and advocates unified principles linking uncertainty to robustness and human decisions.
Neuro-Symbolic AI for Compliant Process Predictions
Presents a neuro-symbolic approach using Logic Tensor Networks (LTNs) to incorporate domain-specific process constraints into predictive monitoring, overcoming sub-symbolic methods' compliance issues. Follows a four-stage pipeline: feature extraction, rule extraction, knowledge base creation, and knowledge injection. Achieves higher compliance and accuracy than baselines in experiments.
Multiverse: Text-Guided Cross-Game Level Blending
Multiverse is a language-conditioned generator for blending game levels across multiple games using textual prompts. It learns a shared latent space to align text instructions with level structures and employs multi-positive contrastive supervision for cross-domain links. This enables controllable blending via latent interpolation and zero-shot compositional generation.
MediHive: Decentralized Agents for Medical Reasoning
MediHive introduces a decentralized multi-agent framework for medical QA using LLMs, featuring self-assigning agents, evidence-based debates, and iterative fusion for consensus. It addresses limitations of centralized systems like scalability issues and single points of failure. The system achieves 84.3% accuracy on MedQA and 78.4% on PubMedQA, outperforming baselines.
Gaps in EU AI Act Transparency Rules
EU AI Act Article 50 II requires dual human- and machine-readable labels for AI-generated content starting August 2026, clashing with generative AI constraints. Analysis of fact-checking and synthetic data generation reveals post-hoc labeling is insufficient due to non-deterministic outputs and workflows. Three structural gaps demand treating transparency as core architecture.
FormalProofBench Tests AI Graduate Math Proofs
FormalProofBench is a new private benchmark evaluating AI models' ability to produce formally verified graduate-level math proofs in Lean 4. Problems are sourced from qualifying exams and textbooks in analysis, algebra, probability, and logic. Frontier models achieve up to 33.5% accuracy, with analysis of tool-use, failures, cost, and latency.
daVinci-LLM Advances Open Pretraining Science
daVinci-LLM combines industrial-scale resources with full openness to explore LLM pretraining science. It releases a 3B model trained from scratch on 8T tokens using Data Darwinism framework and two-stage adaptive curriculum. Over 200 ablations reveal processing depth as a key scaling factor and domain-specific strategies.
Claude Code Source Code Leaked
Anthropic accidentally leaked 59.8 MB of Claude Code's TypeScript source code via an npm package source map file. The 512,000-line codebase reveals innovative memory architecture to combat context entropy in AI agents. This $2.5B ARR product's IP exposure aids competitors like Cursor.
Apple MacBook Gets Phone Chip at ยฅ4599
Apple integrates mobile phone chips into MacBook, launching a ยฅ4599 model alongside Air. This shift from traditional laptop chips enables affordable entry into Apple ecosystem. Subjective review compares Air and Neo experiences.
AlignOPT: LLM-GNN for COPs
AlignOPT aligns LLMs with graph neural solvers to overcome limitations of language-only approaches in combinatorial optimization. LLMs encode textual COP descriptions, while GNNs model graph structures for integrated representations. It achieves state-of-the-art results and strong generalization to unseen instances.
53x Faster Bitboard Tetris AI
This arXiv paper introduces a high-performance Tetris AI framework using bitboard optimizations for a 53-fold speedup over OpenAI Gym-Tetris. It features an afterstate-evaluating actor network and buffer-optimized PPO, achieving 3,829 scores on 10x10 grids in 3 minutes. The framework includes a Python-Java Gym interface for easy RL integration.
Mobvoi Cuts 2025 Loss 90%, AI Hardware Grows
Mobvoi reported 2025 revenue of 3.23 billion RMB, narrowing losses by 90.5% and adjusted losses by 72.6%. AI hardware revenue reached 1.74 billion RMB, up 14.4% YoY. AI software gross margin held at 86.7% with ops costs down 57.4%.
Apple's Buybacks Beat AI Bets
Apple spent $700B on stock buybacks over a decade, prioritizing financial engineering over risky AI investments like rivals. It shifts iPhone production to India for supply chain stability and earns $900M in 2025 from AI app commissions on App Store. This positions Apple as a 'tollbooth' for AI services without core innovation.
Hotels Rush All-In on AI
International hotels are urgently committing fully to AI amid competitive pressures. The article emphasizes that AI serves as a tool rather than an end goal itself.
OpenClaw Hype Lessons for Enterprise
OpenClaw's current hype acts as an early market signal rather than a mature product. Embedding it in enterprise environments faces hurdles like security gaps, compliance issues, and high token costs.
LiDAR Profits Surge as Carmakers Struggle
Car manufacturing is unprofitable, but LiDAR vendors are thriving. Hesai prioritizes profit margins, while RoboSense aggressively scales production and market share.
Global AI Governance Race Heats Up
Countries worldwide are finalizing AI governance plans. These reveal AI's productivity gains alongside global safety risks. A battle for AI rules is set to unfold.
ZTE Deepens ByteDance Tie-Up for AI Phones
ZTE announced deeper collaboration with ByteDance and other ecosystem partners to advance R&D and rollout of next-generation Doubao AI phones. This aligns with ZTE's 'AI for All' strategy.
Twitter's 20 Years Echo Internet Lifecycle
Twitter marks 20 years as X in 2026 with declining US usage at 22%, surpassed by Threads. Founder Jack Dorsey reposts first tweet, calling it 'unfinished.' Article traces evolution from open SMS tool to algorithm-dominated platform, mirroring internet's shift.
Meta Tests Cheap Instagram Plus Story Peeker
Meta is testing Instagram Plus subscription in Philippines and Mexico at 65 PHP (~RMB 7.41)/month for regular users. Features include unlimited Story audience lists, repeat view counts, anonymous peeking at others' Stories, and 24-hour extensions. It's cheaper than Verified service and blends creator/consumer tools, but still in early testing.