What Belongs in AGENTS.md, Design Docs and Tests?
Why agent instructions accumulate and how to decide what to keep, move, or remove.
20 stories
Why agent instructions accumulate and how to decide what to keep, move, or remove.
Microsoft Research unveils Quine, a multimodal biological world model that ranks compounds in silico before scientists commit them to wet-lab experiments.
RLinf open-sourced RPent, a recursive agent framework that wraps frozen vision-language-action models with planners and memory to reliably run manipulation tasks.
NVIDIA published a 4-bit NVFP4 build of Qwen2.5-VL-7B-Instruct that runs on Blackwell Tensor Cores through TensorRT-LLM, cutting memory roughly 3.5x.
A new architecture called SOMA shards models into independent experts, slashing the gradient variance that has blocked zero-order methods from scaling to pretraining.
Pleias unveils SYNTH, an open fully synthetic Wikipedia-derived corpus, plus Baguettotron and Monad models that match larger baselines with 10-140x less data.
A new paradigm lets language models edit their own context file with arbitrary Bash commands, beating hand-designed summarization harnesses across long-horizon agent tasks.
Jeff packs zero-shot classification into 0.8B parameter models that decide in 22ms, trained on one workstation GPU in about two hours.
Apple researchers show a 9B model can crack tasks it failed 128 times in a row by backpropagating short self-written notes instead of full solutions.
Magpie is an open-source menu bar app that lets any coding agent talk to any model provider through a single local gateway, mixing subscriptions and API keys on demand.
The startup spun out of gameplay platform Medal just closed a fresh round to open access to action foundation models trained on billions of clips.
A CTranslate2 conversion of OpenAI's Whisper medium delivers up to 4x faster transcription with lower memory, and it's crossed 830K downloads.
NVIDIA's BigVGAN v2 vocoder now runs at 44kHz with 512x upsampling, a fused CUDA kernel for 1.5-3x faster synthesis, and training across diverse audio.
A trio of surgical tweaks to PPO's critic eliminates training collapse in LLM reinforcement learning, delivering up to 14.89% score gains without touching the actor.
Huihui released an uncensored 180B version of Qwen3.8-Flash-Next using abliteration, with GGUF quants supporting a 262K context window.
Renmin University researchers release a plugin for Claude Code and Codex that builds and evolves a semantic layer over your data, grounded in actual agent interactions.
LibraryDesignBench tests whether AI coding agents can build libraries that other agents actually find usable, exposing why reuse keeps breaking down.
Ten Claude Sonnet 5.5 agents collaborated for 15 hours to produce a 17,895-line Lean proof for the Thomson problem at N=7.
A 12B multimodal classifier fine-tuned from Gemma 4 that returns calibrated probabilities across text, image, audio and video inputs.
Supersonic Labs released Julia-1, a 144M-parameter open decision model that turns a state, question, and options into one typed choice, running in-browser on WebGPU.