CMU's IdeaTaste Reveals AI Research Agents Hit Only 18% of Human Judgment
A new CMU benchmark freezes the literature at 2025 and checks whether LLM research proposals match what human communities actually published in 2026.
20 stories
A new CMU benchmark freezes the literature at 2025 and checks whether LLM research proposals match what human communities actually published in 2026.
An agentic loop autonomously discovered a lossless 12-bit weight format and new verification tricks, more than doubling speculative decoding speedups without any training.
A CMU pipeline turns handheld human demos into whole-body mobile robot policies that avoid walls, reach into fridges, and fold boxes at 87% task progress.
A new transformer architecture keeps a shared pool of blocks and lets a router pick which one runs at each depth, mixing dense, looped, and mixture-of-experts behavior under one knob.
A community-tuned drafter for Qwen3.8-27B's NVFP4 build pushes DGX Spark decode from 12.5 to 48.2 tokens per second, bit-exact lossless.
A 4.5 MB Rust binary bundles a browser kanban board with an MCP server so Claude Code and Codex share the same task list as you.
Speridlabs' open image-to-3D model Idol lands in ComfyUI, turning one picture into a million-Gaussian splat on 7.2 GB of VRAM.
A Japanese ASR model from Third Intelligence hits 6.12% CER at roughly 1/70th the cost of GPT-Transcribe, matching commercial APIs.
An open-source multi-agent plugin turns DeepSeek Harness into a team of persistent bots with memory, group chats, and bot-to-bot handoffs.
A new paper shows transformers trained on interacting agents build separate internal belief subspaces per agent, scaling linearly instead of exponentially.
OpenGVLab released InternVL2-1B, a sub-billion-parameter vision language model pairing InternViT-300M with Qwen2-0.5B for edge-friendly multimodal reasoning.
NVIDIA shipped an NVFP4 quantized DeepSeek V4.1 Flash that trims 75 GB off disk while matching the original on six benchmarks.
A refusal-stripped build of Qwen3.5-4B shrinks to 2.7GB, keeps benchmark scores intact, and runs offline on an 8GB laptop.
OhMyGame is an open-source desktop studio where an AI agent drafts the design doc, generates art, writes code, and playtests the game for you.
Scenario Labs open-sourced 70+ Agent Skills that teach Claude, Cursor, and Codex to generate production-ready images, video, audio, and 3D through MCP.
A tuned imatrix GGUF repack of Spark-X2.5-4B lands a 4B coder on 6 GB GPUs with usable speed at 131k context.
Amazon's 21M-parameter time series foundation model delivers zero-shot forecasts up to 250x faster than original Chronos while beating specialized baselines.
A new MCP server called BLA shields JavaScript bundles, Electron apps, WASM, and native binaries from LLM-powered reverse engineering agents like REA.
A new Claude Code skill records animations from any site, video, or gif, then rebuilds and verifies the copy frame by frame until the motion matches.
Exo lets an agent inspect and rewrite its own harness code at runtime, cutting model costs by double digits while matching task completion.