FW
123 posts
(1/n) Introduce Online Experiential Learning toward the era of experience. Beyond offline pre-constructed training data, models can learn online from their own deployment experience across infinite, unsimulable real-world environments. Accumulate, consolidate, self-improve 🔄
💥 New example out! Deploy @Microsoft VibeVoice-ASR on Microsoft Foundry with @huggingface for multi-lingual STT! Structured output with Who (Speaker), When (Timestamps), and What (Content), up to 60 minutes in a single pass. Step-by-step in the thread 🧵
On-Policy Context Distillation for Experiential Learning: learning from experience (consolidated from trajectories) at test time.
Introduce Differential Transformer V2 (DIFF V2), an improved version of Differential Transformer. This revision focuses on inference efficiency, training stability, and architectural elegance. We verify the design on production-scale LLMs.
Microsoft just released VibeVoice-Realtime-0.5B huggingface.co/microsoft/Vibe…
🚀 We propose Generative Adversarial Distillation (GAD) 🤖 Designed to perform on-policy distillation from proprietary black-box LLMs. ➡️ Requires neither access to teacher logits nor alignment of tokenizer vocabularies. (1/n)
New @Microsoft paper teaches LLMs to organize reasoning into concurrent subtasks for faster, more accurate answers. It shows 28% lower wait time than typical parallel thinking while also boosting math accuracy. The big deal is simple, it turns coordination into a skill the model learns, so it decides when to split work, when to wait, and when to merge. The usual single chain wastes time because each step blocks the next. Fixed parallel plans also waste time because they cannot adapt to each query. The fix is an organizer that writes simple Fork and Join tags to start and merge worker thoughts. Workers chase sub-queries in parallel while the organizer keeps thinking and only pauses to Join. All control lives in plain text, so the base model stays unchanged. Training happens in 2 stages, first supervised traces that teach the tag format. Then reinforcement learning rewards correct final answers, clean format, and real concurrency. Speed is measured by the critical path through the Fork-Join graph, which matches true waiting. Across countdown puzzles, math questions, and Sudoku, the learned policy runs faster and fails less. The big idea is to learn organization itself rather than hard-code a script. ---- Paper – arxiv. org/abs/2510.26658 Paper Title: "The Era of Agentic Organization: Learning to Organize with Language Models"
Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other approaches for a fraction of the cost. thinkingmachines.ai/blog/on-policy…
Microsoft presents BitNet Distillation
Thinking Augmented Pre-training "we propose Thinking augmented Pre-Training (TPT), a universal methodology that augments text with automatically generated thinking trajectories. Such augmentation effectively increases the volume of the training data and makes high-quality tokens more learnable through step-by-step reasoning and decomposition." "Notably, TPT enhances the data efficiency of LLM pre-training by a factor of 3. For a 3B parameter model, it improves the post-training performance by over 10% on several challenging reasoning benchmarks."
VibeVoice-7B-Preview with 32K Context Length. Wan2.2 + VibeVoice is pretty nice github.com/microsoft/Vibe…
#Microsoft's VibeVoice-1.5B just turned my rig into a podcast studio. 4 voices. Zero API costs. Running locally on a consumer GPU. Generated 5 test podcasts instantaneously - they sound surprisingly human. Setup took 30 minutes: clone repo, load model (most of the time - rural Germany), feed script, press play. The open-source podcast revolution is here, and it fits in your home rig. Who needs cloud subscriptions when innovation runs at localhost? 🎙️ #GenAI #LocalAI #Podcasting #VibeVoice
Microsoft just dropped VibeVoice on Hugging Face A novel framework generating expressive, long-form, multi-speaker conversational audio like podcasts from text. Synthesizes up to 90 minutes of speech with up to 4 distinct speakers! huggingface.co/microsoft/Vibe…
Reinforcement Pre-Training New pre-training paradigm for LLMs just landed on arXiv! It incentivises effective next-token reasoning with RL. This unlocks richer reasoning capabilities using only raw text and intrinsic RL signals. A must-read! Bookmark it! Here are my notes:
⏰ We introduce Reinforcement Pre-Training (RPT🍒) — reframing next-token prediction as a reasoning task using RLVR ✅ General-purpose reasoning 📑 Scalable RL on web corpus 📈 Stronger pre-training + RLVR results 🚀 Allow allocate more compute on specific tokens
