AI Research Pulse

Kimi Delta Attention (KDA) replaces most full-attention layers with a hybrid linear mechanism that scales linearly while preserving exact...

Traditional benchmarks like SWE-bench hand models complete specs upfront and score one-shot results. SlopCodeBench instead reveals requirements...

Agentic AI's thousands of daily autonomous actions outpace documents and review boards, requiring governance built into the platform's real-time...

FAOC integrates RL flexibility with OC safety by mapping abstract actions to guaranteed-feasible OCP parameters in constrained systems.

• Key...

An end-to-end RL policy for quadrotors uses differentiable simulation, ToA maps as privileged information, and a yaw alignment loss to navigate large...

Meaningful AI safety assessment requires examining deployment environments, safeguards, moderation layers, and governance mechanisms alongside the underlying model.

• Blind spots persist: Microsoft’s EXTRA program funds 18 global labs with unrestricted grants to expose failure modes missed by internal teams.
  -...

Physics-informed ML models for detection tasks can swap deep learning for Random Forest or XGBoost when interpretability requirements rule out black-box approaches.

• Google joins EU code committing to watermarking and metadata labels for generative content, positioning itself ahead of the AI Act's transparency...

Kimi K3's hybrid attention stack slashes long-context memory demands: three-quarters of layers use constant-size KDA state instead of growing KV...

Transformers are accelerating nuclear reactor simulations with high accuracy, yet protein-folding tools reveal how easily AI ignores physical laws.

-...

DQN and Double DQN solve the environment with reported advantages in sample efficiency and stability over policy-gradient methods.

Static certifications miss the core problem. SOC 2 and ISO 42001 treat safety as a fixed design-time property, but autonomous agents introduce...

A new survey unifies progress reward modeling for robotics into three connected steps—model interface, internal construction methods, and supporting...

A novel Risk-Aware SAC integrates risk perception, obstacle motion extrapolation, and heading constraint rewards into the SAC framework for unmanned vehicle path planning. This preserves entropy while enhancing safety in dynamic environments.

• NVIDIA's $5B SSI deal grants 10x Vera Rubin compute after internal research review, following an OpenAI model escape incident.
• Untrusted advice...

A new dRAE autoencoder discretizes high-dimensional visual representations via Hyper-Spherical Quantization, addressing codebook collapse that has...

• StateAct prioritizes program state over pixels to improve reliability in long-horizon computer-use agents.
• MAPD uses structured protocols to...

Naive velocity matching under classifier-free guidance creates Negative Branch Asymmetry (NBA) when the teacher’s negative branch holds privileged...

DecoupleMix replaces heuristic VLM pretraining mixtures with a decoupled optimization framework that separates inter-class capability ratios from...