All news

A malicious AGENTS.md file in a Downloads folder made a coding agent install a rogue npm package. Researchers reproduced it against OpenAI Codex and showed a related trick against Claude.

Joint caching of large models with parameter sharing and inference offloading in mobile edge computing.

DoorDash uses a multi-agent LLM system to clean up stale feature flags at scale.

Shows routing to the most predictable model is suboptimal for LLM routing.

DIAMM splits multimodal model inference across edge devices to cut latency.

A glibc update in Flatpak runtime 21.08 broke Chromium audio because of a sandbox syscall filter, and the post traces the cause.

SuperMimic uses one differentiable world model for humanoid motion control.

FairCache shows prefix and context caching makes multi-tenant LLM serving unfair and proposes cache-aware scheduling to fix it.

Shows offline tests cannot always verify agent updates stay safe after deployment.

PrefixQuant adds prefixed tokens to absorb activation outliers, enabling simpler static LLM quantization.

Models future evolution to improve autoregressive chunked video generation.

RIFT compresses draft-model KV cache for speculative decoding with complementary training.

Item-level adaptive recursion spends more compute on hard items in sequential recommendation.

AfterDeleteBench shows deleted data persists in LLM agent memories and derived summaries.

Meta Expert shares experts across layers to cut MoE storage and retrieval.

Tests whether frontier models can autonomously build and improve their own software harness.

Lessons on Docker oddities learned while writing a container runtime.

Fine-tuning Gemma 4 26B-A4B to think 39 to 69 percent less with the same accuracy.

Agent Mosaic detects confidentiality leaks in agentic LLM systems.

Learns robot behavior priors from non-robot synthetic data.

Agentic world action models for robot manipulation.

Continued pretraining for embodied agents in many virtual worlds.

Training-free output-driven allocation of block-sparse attention in video diffusion transformers.

Motion-guided VLM picks frames for efficient video anomaly detection instead of dense per-frame inference.

Fly.io explains why it built its own block device layer for Sprites instead of polishing a homemade filesystem.

Qwen-Audio-3.1-ASR technical report.

Stratified sparse attention narrows the gap to dense attention at extreme sparsity for long contexts.

Google and Meta open-sourced TorchTPU, a native PyTorch backend for TPUs, and Meta plans to move larger LLM workloads onto TPUs; AWS released TorchNeuron for Trainium the same week.

Skeleton-grounded latent action model for robot world models.

An incremental approach to building a JIT compiler.