https://xitter.cf/t.co/uJtHMTZCrZ
Joined May 2018
- Tweets130
- Following16
- Followers187
- Likes90
Agensh: "The key to scale multi-agent systems is self-organization."
Banger paper from Microsoft Research.
(bookmark it)
They run 1K+ coding agents at once to test a scalable self-organized multi-agent harness.
This is an interesting test because most multi-agent systems today have some hierarchy or structure.
Agensh has no central orchestrator.
It coordinates parallel coding agents through a shared state instead of a central orchestrator.
Each agent gathers context, claims a sub-task, does the work, shares what it found, verifies the result, and merges it, all asynchronously through a shared workspace and a message channel.
On the five hardest ProgramBench tasks with GPT-5.6-sol, increasing agents from 1 to 128 raises the mean final test-pass rate from 19.31% to 28.78%.
Larger teams also reach a given pass rate sooner.
On pandoc, 1,024 agents take the test-pass rate from 33.89% to 55.06%.
The authors also report forms of cooperation that the agents start on their own and that become standard practice as the team grows.
Paper: arxiv.org/abs/2609.26781
Chat with Paper: academy.dair.ai/papers/agens…
Introducing Agensh, a scalable, self-organized multi-agent harness and system designed to run agents at scale. See how 1,024 agens cooperate at aka.ms/agensh.
Great to see YOCO (decoder-decoder architecture) is used as the foundation for the new DeepSeek v4.1 Flash model architecture. YOCO will be the de facto transformer architecture in the era of agentic AI, offering order-of-magnitude advantages in prefilling and KV cache (memory).
DeepSeek V4.1 Flash has also embraced YOCO (You Only Cache Once; arxiv.org/abs/2405.05254 ) as the soul of its architecture, inheriting YOCO’s order-of-magnitude advantages in prefill and KV caching. #decoderdecoder
We’re thrilled to introduce VibeVoice-ASR-Streaming, the newest member of the VibeVoice model family, a unified streaming ASR model that continuously transcribes “who said what” as speech arrives, with support for custom hotwords and 10 languages.
LLM-as-a-Coach: Experiential Learning for Non-Verifiable Tasks
LLM-as-a-Coach: Experiential Learning for Non-Verifiable Tasks
"We propose Experiential Learning (EL), which repurposes the feedback model from an LLM-as-a-Judge into an LLM-as-a-Coach. The coach distills its assessment of each on-policy response into transferable experiential knowledge, which conditions a teacher model and is internalized by the policy through on-policy context distillation."
More #BitNet_LLMs and models
Great to see @TIIuae push the boundaries of LLM training with their Falcon‑H1 parallel hybrid design and BitNet ternary training—now integrated into NVIDIA Megatron Core.
Their work shows how foundation model builders can innovate on top of our scalable framework while improving efficiency and sharing reusable tools with the open source community. 🙌
👇 Read the full technical blog: developer.nvidia.com/blog/im…
Online Experiential Learning
VibeVoice-ASR is now officially in HF Transformers and Microsoft Foundry.
💥 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 🧵
#VibeVoice-ASR - your new ASR model in the era of LLM
FW retweeted
Microsoft just released VibeVoice-ASR on Hugging Face
A unified speech-to-text model that transcribes hour-long audio in one pass
With built-in speaker diarization, timestamps, and customizable user context
Anyways check out this 60s clip I made with a single image of @lexfridman and a 20s audio recording of his voice.
Full continuous shot with no cuts created with LongCat Avatar in ComfyUI.
VibeVoice for the audio.
New #VibeVoice model released - VibeVoice Realtime 0.5B for realtime, streaming, robust long-form TTS
Microsoft just released VibeVoice-Realtime-0.5B
huggingface.co/microsoft/Vib…
Generative Adversarial Distillation for Black-Box On-Policy Distillation of LLMs
FW retweeted
🚨 Microsoft Research just launched something that might define the next era of AI systems.
They call it 'Agentic Organization' and it’s not just a new model. It’s a new way for intelligence itself to organize.
Here’s what’s wild:
Most large language models still “think” like a single brain.
Step-by-step. Linear. Slow. Even “parallel thinking” just runs the same process twice and merges answers later.
Agentic Organization changes the entire game.
They built a new reasoning protocol called AsyncThink, where a model plays both roles an Organizer that breaks a complex problem into sub-queries, and Workers that solve those sub-parts at the same time.
Think of it like this:
Instead of one mind grinding through steps, AsyncThink forms a mini civilization of minds delegating, merging, adapting in real time.
And it learns this behavior through reinforcement learning literally learning how to organize its own thoughts.
The results are insane:
→ 28% lower inference latency than parallel thinking
→ Higher accuracy on math reasoning tasks
→ Zero-shot generalization to unseen problems like Sudoku
→ Learned organizational policies that evolve dynamically during reasoning
It’s like scaling from “an intelligent agent” → to “an intelligent organization.”
AsyncThink models don’t just reason faster they reason like teams do.
Fork. Think. Join. Verify. Iterate.
This is a glimpse of post-LLM intelligence systems that don’t just think, they coordinate thought.
And if that holds, the future of AI might look less like a single brain… and more like a company of minds.
Paper: The Era of Agentic Organization: Learning to Organize with Language Models
ALT Title page for a research paper on "Agentic Organization" with abstract text and diagrams comparing sequential, parallel, and asynchronous agentic thinking.
The Era of Agentic Organization
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"
FW retweeted
On-policy + Reverse KLD = MiniLLM (arxiv.org/abs/2306.08543). Really nice blog by @thinkymachines. Exciting to see it being offered as a service!
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-…
