All Updates
Page 905 of 1661
April 20, 2026
Anthropic, White House Eye Reconciliation on AI Risks
Anthropic's CEO met with US government over cybersecurity risks in their latest model, Mythos. Following tensions with the Department of Defense, both sides may seek cooperation for trust recovery and safety. This comes amid rising concerns.
Replacements for Banned Claude Code
User banned from Claude Pro seeks local alternatives matching Claude's reasoning and Code's CLI/file workflows. Needs agent-like terminal access for lesson planning, content creation, DJ sets, Obsidian vaults. Tried ChatGPT but misses the feel; open to paid tools.
Teacher vs Chatbot: AI Classroom Journey Podcast
A novice teacher recounts first-time classroom struggles intensified by AI chatbots in a Guardian podcast. Introducing AI felt like downing coffee during a panic attack. By Peter C Baker, read by Adam Sims.
Subliminal Unsafe Behavior Transfer in AI Distillation
Researchers demonstrate the first empirical evidence that unsafe behaviors in AI agents transfer subliminally via model distillation, even with keyword filtering on safe trajectories. In API settings, student agents inherit 100% deletion bias; in Bash, 30-55% chmod preference. This shows explicit sanitation fails against implicit biases in trajectory dynamics.
Stein Variational Boosts Black-Box Optimization
This new method integrates the Stein operator into Estimation-of-Distribution Algorithms (EDAs) to create particle repulsion, enhancing exploration in multimodal combinatorial landscapes. It prevents premature convergence and excels on large-scale benchmarks, outperforming or matching state-of-the-art approaches. The approach leverages Stein variational gradient descent for discrete black-box problems.
Siemens Eyes US/China Over EU for AI Investments
Siemens AG will prioritize AI investments in the US and China if the EU does not relax its AI regulations. CEO Roland Busch made this statement, highlighting regulatory hurdles in Europe. This reflects growing frustration with Europe's AI framework.
PBRC Stops Conformity in Agent Beliefs
PBRC 是一種協議級機制，用於多代理系統中嚴格分離開放通訊與可接受的信念變更，透過預註冊證據觸發器避免從眾導致的錯誤收斂。合約固定一階證據觸發、可接受修訂運算子、優先規則與後備政策，每項實質信念變更需提供外部驗證證據。論文證明其安全性，並展示審核性與邏輯規格化。
Milkyway Evolves Agents for Future Predictions
Milkyway is a self-evolving LLM agent system that updates a persistent harness using internal feedback from repeated predictions on unresolved questions. It improves factor tracking, evidence gathering, and uncertainty handling without changing the base model. It achieves top scores on FutureX (60.90) and FutureWorld (77.96).
MCTS for Bilevel Agent Skills Optimization
Researchers introduce a bilevel optimization framework for LLM agent skills, using Monte Carlo Tree Search in the outer loop to determine skill structure and LLMs in the inner loop to refine content. This addresses the interdependence between structure and components in skill design. Experiments on an open-source Operations Research QA dataset show improved agent performance.
LLM Reasoning: Latent, Not Chain-of-Thought
This position paper argues that LLM reasoning is best studied as latent-state trajectories rather than surface chain-of-thought (CoT). It formalizes three hypotheses separating latent states, explicit CoT, and serial compute, with evidence favoring latent trajectories (H1). Recommends focusing on latent dynamics and disentangling evaluations for better interpretability.
LACE Enables Cross-Thread LLM Reasoning
LACE transforms isolated LLM reasoning into coordinated parallel processes via cross-thread attention, allowing paths to share insights and correct errors. It uses a synthetic data pipeline to train collaborative behavior absent in natural data. Experiments show over 7-point accuracy gains versus standard parallel search.
KWBench: LLM Unprompted Problem Recognition Benchmark
KWBench introduces a benchmark for evaluating LLMs' ability to recognize problems in knowledge work scenarios without explicit prompts. It features 223 tasks from fields like acquisitions and fraud analysis, based on game-theoretic patterns, with a three-tier scoring rubric. Top models achieve only 27.9% pass rate unprompted, highlighting evaluation gaps.
GIST: Multimodal Spatial Grounding Topology
GIST transforms consumer-grade mobile point clouds into semantically annotated navigation topologies via 2D occupancy maps, topological layouts, and lightweight semantic overlays. It powers tasks like intent-driven semantic search, one-shot localization (1.04m error), zone classification, and visually-grounded instruction generation. Evaluations show superior performance over baselines and 80% navigation success with verbal cues.
DeepER-Med: Agentic AI for Medical Research
DeepER-Med introduces an agentic AI framework for evidence-based medical research with modules for planning, collaboration, and synthesis. It includes DeepER-MedQA, a dataset of 100 expert-level questions from real scenarios. The system outperforms top platforms in insight generation and aligns with clinical recommendations in 7 of 8 cases.
Claude Adds Context Sharing & Skills for Excel/PowerPoint
Anthropic enhanced Claude for Excel and Claude for PowerPoint with multi-file context sharing. The new 'skills' feature enables one-click execution of routine workflows. This eliminates the hassle of switching between Excel and PowerPoint.
Canadian AI Register Obscures Accountability
Canada launched its first Federal AI Register in November 2025, listing 409 AI systems. Analysis reveals 86% are internal efficiency tools, but the register obscures human discretion, training data, and uncertainty management. It frames AI as reliable tooling rather than contestable decision-making.
Algebraic Invariants Enhance LLM Reasoning
Presents a symbolic reasoning framework for LLMs operationalizing abduction, deduction, and induction via Peirce's tripartite inference. Enforces consistency with Gamma Quintet invariants, including Weakest Link bound to prevent error propagation in chains. Verified through 100 properties and 16 fuzz tests over 10^5 cases, offering a benchmark foundation.
Xiaomi miclaw first to pass CAICT Claw eval
Xiaomi's miclaw mobile intelligent agent is among the first to pass the CAICT Claw evaluation for smartphone intelligent assistants. This regulatory approval validates advanced on-device AI systems. It is powered by Xiaomi's in-house MiMo large model.
iQiyi Overhauls for Full AI Content Creation
iQiyi Inc. anticipates AI generating entire films and shows from scratch soon. This vision drives the streaming service's largest overhaul in 16 years. It marks a monumental shift in the content industry.
SGOCR: Grounded OCR Dataset Pipeline Released
Independent researcher releases SGOCR, an open-source pipeline and V1 dataset for spatially-grounded OCR-focused VQA with rich metadata. Pipeline uses Nvidia nemotron-ocr-v2 for extraction, Gemma4/Qwen3-VL for anchors, and Gemini-2.5-flash for verification. Developed via agentic loops and custom optimization for VLM training.