A 4B compiler trains parameter-efficient adapters for a frozen 0.6B interpreter, matching Qwen3-32B direct prompting while using roughly 1/50th inference memory at 30 tok/s on a MacBook M3. FuzzyBench dataset of 10M examples released.
Treats LLM memory management as a trainable skill via two automated loops — memory structure optimization and proficiency enhancement. Improves a 32B open-weight model roughly 2x-4x on Crafter, MiniHack, and NetHack, matching Claude Opus 4.5 and Gemini 3.1 Pro Thinking.
MrFlow combines low-res generation with pixel-space super-resolution and noise injection, achieving 10x end-to-end speedup on FLUX.1-dev and Qwen-Image. Orthogonal to timestep distillation for a combined 25x acceleration. SOTA among training-free methods.
RL framework using distribution-wise rewards with subset-replace strategy mitigates mode collapse. Improves FID-50K from 8.30 to 5.77 on SiT and from 3.74 to 3.52 on EDM2.
LOCOS identifies attention heads responsible for non-literal context synthesis via OV-circuit analysis. Mean-ablating top heads collapses ROUGE-L from 0.401 to 0.000 on Qwen3-8B, while the strongest baseline retains 0.292.
Uses VLMs to establish instance-level correspondences between text prompts and visual conditions with adaptive mask refinement. Outperforms SOTA in multi-instance controllable generation. Accepted at ECCV 2026.
FlashMorph formulates hybrid attention layer selection as budget-constrained subset optimization using morphable models with linearization regularization. Discovers effective hybrid configs preserving long-context recall. By ByteDance Seed.
Introduces bounded contract approach with typed retrieval for Llama agents in Slay the Spire 2. No-store baseline wins 3/10 games vs 6/10 with strategic skills enabled. 298 trajectories released.
China's U-World revealed the U1, a full-size ultra-bionic humanoid robot built for mass production. The robot can read emotions, recognize individuals, and recreate a person's face and voice, marking a major step in commercial humanoid robotics.
This week's major AI developments: Anthropic released Sonnet 5, OpenAI announced the GPT-5.6 family, Google launched NotebookLM Shorts and Nano Banana 2 Light, Gemini arrived on Mac, and Fable 5 returned with new restrictions.
Open-source tool that lets multiple Claude Code sessions running in the same checkout see each other's in-flight work and message each other mid-turn, enabling parallel feature development without branches or worktrees.
Open-source screen memory tool that captures and analyzes screen activity locally using Gemma 4. Fully on-device with searchable AI memory, MCP server support, and runs on hardware as modest as a $150 GPU.
Anthropic shipped a richer admin analytics dashboard, model-level access entitlements, and configurable spend-threshold alerts for Claude Enterprise, giving IT and finance teams granular cost oversight over agentic AI usage.
Elvis Saravia discusses multimodal prompting beyond text as the future for interacting with coding agents, sharing implementation details and a video walkthrough.
Simon Willison observes that with current AI systems, developers can generate thousands of lines of production code per day, undermining arguments that writing code yourself is faster.
