Daily archive

Friday — April 17, 2026

100 stories — deduplicated across sources, ranked by significance, every source cited.

SIGNIFICANCE

★ Top story · ModelsApr 17

Qwen Team Open-Sources Qwen3.6-35B-A3B: A Sparse MoE Vision-Language Model with 3B Active Parameters and Agentic Coding Capabilities

Qwen Team Open-Sources Qwen3.6-35B-A3B: A Sparse MoE Vision-Language Model with 3B Active Parameters and Agentic Coding Capabilities MarkTechPost

QQwen News MMarkTechPost Read original ↗

ModelsQwen 3.6 35B-A3B

Friday — April 17, 2026

Qwen Team Open-Sources Qwen3.6-35B-A3B: A Sparse MoE Vision-Language Model with 3B Active Parameters and Agentic Coding Capabilities

Dive into Claude Code: The Design Space of Today's and Future AI Agent Systems

A Chinese AI called 'Qwen3.6-35B-A3B,' which is more powerful than Gemma4, has been released as an open model. - GIGAZINE

Claude Opus wrote a Chrome exploit for $2,283

Alibaba (09988) Qwen 3.6-35B-A3B Model Open Source Release - 富途牛牛

Introducing Claude Design by Anthropic Labs

The Specification Trap: Why Static Value Alignment Alone Is Insufficient for Robust Alignment

Beijing brands Meta's Manus acquisition as "conspiratorial" and bars founders from leaving China

LLMs Gaming Verifiers: RLVR can Lead to Reward Hacking

Finance ministers and top bankers raise serious concerns about Mythos AI model - BBC

AutoRAN: Automated Hijacking of Safety Reasoning in Large Reasoning Models

When Slower Isn't Truer: Inverse Scaling Law of Truthfulness in Multimodal Reasoning

Threshold Differential Attention for Sink-Free, Ultra-Sparse, and Non-Dispersive Language Modeling

Agentic Microphysics: A Manifesto for Generative AI Safety

A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation

Startup Offers AI Tools to Find New Jobs for Workers Displaced by AI

HWE-Bench: Benchmarking LLM Agents on Real-World Hardware Bug Repair Tasks

Reward-Aware Trajectory Shaping for Few-step Visual Generation

Representation Interventions Enable Lifelong Knowledge Memory Control in LLMs

AIBuildAI: An AI Agent for Automatically Building AI Models

SpaceMind: A Modular and Self-Evolving Embodied Vision-Language Agent Framework for Autonomous On-orbit Servicing

Seeing Through Experts Eyes A Foundational Vision Language Model Trained on Radiologists Gaze and Reasoning

From Tokens to Layers: Redefining Stall-Free Scheduling for MoE Serving with Layered Prefill

CoopEval: Benchmarking Cooperation-Sustaining Mechanisms and LLM Agents in Social Dilemmas

MemoSight: Unifying Context Compression and Multi Token Prediction for Reasoning Acceleration

GAMBIT: A Gamified Jailbreak Framework for Multimodal Large Language Models

HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds

PUFFIN: Protein Unit Discovery with Functional Supervision

Latent-Condensed Transformer for Efficient Long Context Modeling

Maximal Brain Damage Without Data or Optimization: Disrupting Neural Networks via Sign-Bit Flips

IG-Search: Step-Level Information Gain Rewards for Search-Augmented Reasoning

Stability and Generalization in Looped Transformers

To See or To Please: Uncovering Visual Sycophancy and Split Beliefs in VLMs

Mechanistic Decoding of Cognitive Constructs in LLMs

Prompt Injection as Role Confusion

SAM3-I: Segment Anything with Instructions

GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens

PixelDiT: Pixel Diffusion Transformers for Image Generation

How Retrieved Context Shapes Internal Representations in RAG

Dense Neural Networks are not Universal Approximators

Your LLM Agents are Temporally Blind: The Misalignment Between Tool Use Decisions and Human Time Perception

Acceptance Dynamics Across Cognitive Domains in Speculative Decoding

Implicit Neural Representations: A Signal Processing Perspective

ELMoE-3D: Leveraging Intrinsic Elasticity of MoE for Hybrid-Bonding-Enabled Self-Speculative Decoding in On-Premises Serving

Attribution, Citation, and Quotation: A Survey of Evidence-based Text Generation with Large Language Models

Awakening Dormant Experts:Counterfactual Routing to Mitigate MoE Hallucinations

Route to Rome Attack: Directing LLM Routers to Expensive Models via Adversarial Suffix Optimization

Hijacking Large Audio-Language Models via Context-Agnostic and Imperceptible Auditory Prompt Injection

KnowRL: Exploring Knowledgeable Reinforcement Learning for Factuality

OccuBench: Evaluating AI Agents on Real-World Professional Tasks via Language Environment Simulation

Compressing Sequences in the Latent Embedding Space: $K$-Token Merging for Large Language Models

The AI Research Assistant: Promise, Peril, and a Proof of Concept

CODESTRUCT: Code Agents over Structured Action Spaces

MCP-Flow: Facilitating LLM Agents to Master Real-World, Diverse and Scaling MCP Tools

NaturalGAIA: A Verifiable Benchmark and Hierarchical Framework for Long-Horizon GUI Tasks

Seeing Through Circuits: Faithful Mechanistic Interpretability for Vision Transformers

Large Language Model Post-Training: A Unified View of Off-Policy and On-Policy Learning

Attack Selection Reduces Safety in Concentrated AI Control Settings against Trusted Monitoring

The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows

The Agentification of Scientific Research: A Physicist's Perspective

Layered Mutability: Continuity and Governance in Persistent Self-Modifying Agents

Mind DeepResearch Technical Report

Sparse Growing Transformer: Training-Time Sparse Depth Allocation via Progressive Attention Looping

Topology-Aware Active Learning on Graphs

BarrierBench: Evaluating Large Language Models for Safety Verification in Dynamical Systems

H2VLR: Heterogeneous Hypergraph Vision-Language Reasoning for Few-Shot Anomaly Detection

Adaptive Canonicalization with Application to Invariant Anisotropic Geometric Networks

Scepsy: Serving Agentic Workflows Using Aggregate LLM Pipelines

The Fourth Challenge on Image Super-Resolution ($\times$4) at NTIRE 2026: Benchmark Results and Method Overview

FedIDM: Achieving Fast and Stable Convergence in Byzantine Federated Learning through Iterative Distribution Matching

DA-Cramming: Enhancing Cost-Effective Language Model Pretraining with Dependency Agreement Integration

Measuring the metacognition of AI

TurboTalk: Progressive Distillation for One-Step Audio-Driven Talking Avatar Generation

Improving Language Models with Intentional Analysis

Towards Design Compositing

SRMU: Relevance-Gated Updates for Streaming Hyperdimensional Memories

Benchmarking Optimizers for MLPs in Tabular Deep Learning

Safe Reinforcement Learning using Action Projection: Safeguard the Policy or the Environment?

MetaMuse: Algorithm Generation via Creative Ideation