
Apple's MixAtlas Cuts Multimodal Training Steps in Half With Smarter Data Mixing
Apple researchers built a two-axis search that finds better multimodal training mixes using tiny 0.5B proxies, then transfers them to 7B models.
20 stories

Apple researchers built a two-axis search that finds better multimodal training mixes using tiny 0.5B proxies, then transfers them to 7B models.

A new framework upgrades cheap scalp EEG toward the fidelity of implanted electrodes by combining MRI-derived cortical geometry, a pretrained EEG foundation model, and a diffusion generator.

A new framework splits shared and private multimodal signals, then routes them through Fourier and hypergraph branches to push conversational emotion recognition past strong baselines.
InCoder-32B-Thinking teaches a 32B model to reason like a hardware engineer by replaying error-correction loops from real industrial toolchains, beating Claude Sonnet 4.6 on CAD and chip design tasks.

A community-abliterated build of Gemma 4 E2B strips all 465 test refusals while keeping vision, audio and a 131K context window intact.

The-Vibe-Company/companion turns a folder of TOML files into a private fleet of specialized AI agents with memory, tools, and Tailscale-only access.

lat.md replaces flat AGENTS.md files with a bidirectional knowledge graph that agents keep in sync, so context survives across sessions.

AI4Finance's benchmark library gives trading agents hundreds of gym-style market environments, unified data pipelines, and reproducible baselines under one MIT-licensed roof.

A new continual learning framework treats replay memories like matter that freezes over time, cutting forgetting by up to 80%.

Allen AI's OLMo-core is a fully open PyTorch library for training frontier LLMs, now powering OLMo-2 and OLMo-3 at 40%+ GPU efficiency

pyannote.audio 4.0 ships community-1, a new open-source diarization model with 50% fewer speaker errors and native Whisper integration

The ONNX ecosystem's model converter library ships 1.16.0 with LGBMRanker export, custom objective support, and better XGBoost categorical handling.

Ben Burtenshaw's open-source repo upgrades Karpathy's autoresearch loop into a 4-agent system that runs ML experiments on HuggingFace GPUs autonomously, 24/7.

A community study of Claude Code's internal architecture hit 12k stars, mapping the 12 layered mechanisms that turn a while-loop into a production coding agent.
A community GGUF conversion of Google's multimodal Gemma 4 E4B-it lands on Hugging Face, unlocking local llama.cpp inference across laptops and consumer GPUs.
ProactAgent teaches LLM agents to decide when to consult memory using reinforcement learning, cutting interaction steps by 33% and beating all baselines on three benchmarks
RAT is a multi-language agent framework that automatically configures executable environments for any code repository, beating prior baselines by 36.1% on environment setup success rate.
Open-H-Embodiment releases 780 hours of surgical robot data across 50 institutions, enabling the first open foundation model to autonomously complete suturing end-to-end.
A new metric and training framework teaches ASR models to say 'I don't know' instead of confidently hallucinating wrong words
Brain-CLIPLM reframes EEG-to-text decoding as semantic compression, recovering keyword anchors from brain signals before reconstructing full sentences with an LLM