2026-09-18

The model is derived from Qwen3.8-27B, a 27B hybrid-attention causal language model (architecture unchanged), but uses ternary weights to shrink model size down to - Collection: - Demo:

Google Research has introduced Retrieve-for-Train (R4T), a framework for search that returns coherent, diverse result sets. It trains a fan-out language model with RL once, using groundedness, diversity, and alignment rewards. That model then synthesizes training data for a 53.9M-parameter diffusion retriever. The retriever generates all retrieval directions in a single pass, running 12× to 20× faster than autoregressive fan-out. No code or model weights have been released yet. The post Google R

Recent work extends recursive self-improvement (RSI) to agent harnesses for long-horizon coding and terminal tasks, enabling agents to improve execution mechanisms from experience. However, generalizable harness RSI remains challenging. First, evolving harnesses on evaluation benchmarks or their subsets makes it difficult to distinguish reusable improvements from benchmark-specific adaptation. Second, single-trajectory updates can conflate systematic harness deficiencies with instance-specific r

Base Labs, the research group Baseten spun up earlier this year, will develop and publish methods for training and monitoring open models.

arXiv:2609.17699v1 Announce Type: new Abstract: We present NeMo Data Designer (NDD), an open-source, general-purpose framework for multi-modal synthetic data generation (SDG). Designed to be intuitive to use, NDD provides a declarative configuration format in which human and/or agent users define each dataset column, with column types spanning text, code, structured outputs, images, embeddings, and statistical samplers that are explicitly configured to steer dataset diversity. Additional column

Recently, linear attention layers have been increasingly adopted to replace softmax attention at scale for long-context modeling. However, existing context extension approaches typically apply continued pretraining directly without modifying these layers, overlooking the spectral properties of linear attention state dynamics. In this work, we study long-context extension of Gated DeltaNet (GDN) from a spectral perspective of transition matrix and identify two essential factors governing long-ran

Direct preference alignment methods are widely used to align large language models (LLMs) with human preferences because of their computational and memory efficiency. However, likelihood displacement motivates alternative ways to extract information from preference pairs with small likelihood margins. In this paper, we propose and analyze Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method based on comparison oracles. ComPO extracts directional information from thes

OpenAI can disclose misalignment before fixes exist. Its 6 initial reports include fabricated data and leaked API keys. The post OpenAI Releases a Model Misalignment Disclosure Framework With 3 Review Tracks and 6 Incident Reports From RL Training appeared first on MarkTechPost .

Nunchux AI has released VC-Attention, a training-free low-bit attention kernel built for video Diffusion Transformers (DiTs). It targets 2 problems at once: value quantization error and a slow softmax stage. Why Attention is the Video Bottleneck Video DiTs flatten a clip into 1 sequence of spatiotemporal tokens and run full self-attention at every layer. A […] The post Nunchux AI Introduces VC-Attention: A Training-Free Low-Bit Attention Kernel That Speeds Up Video Diffusion Transformers appeare

People kept commenting and asking about single AMD 1xR9700 cards in the comments and discord. Well, I finally had time to do some optimizations for 1xR9700 owners and performance has doubled across the board. You can see the results in BetterBench above if you like visuals or below if you're more into text. These results were measured running Unsloth's Qwen3.8 27b NVFP4 . Decode ┌───────────────┬───────────────┬──────────────────┐ │ category │ update p50 ms │ decode t/s (med) │ ├───────────────┼

Reinforcement learning with verifiable rewards is typically performed on-policy, keeping training data close to the current policy but limiting learning to trajectories that the policy can discover itself. Off-policy methods such as supervised fine-tuning, on the other hand, can leverage external knowledge beyond the base model's capabilities, but may suffer from large distribution shift. The key challenge is thus to expand exploration without sacrificing learnability. In this work, we introduce

I find another new model at HF: "Xing4.0-29B-A4B is a next-generation large language model in the Xing series (formerly TeleChat ), developed by China Telecom Artificial Intelligence Technology Co., Ltd. With 29B total parameters and only 4B activated per token, it natively supports a 256K context length, extensible to 512K. It is the first model of this scale trained entirely on the Ascend NPU platform with the MindSpore framework, and deeply optimized for complex engineering tasks. For more in

Newly unsealed court filings show Microsoft privately called OpenAI's data practices "theft" while both companies scraped paywalled Times content, built datasets from it, and warned internally it would gut publishers.

The round values the data center giant at $30.9 billion.

Today, I’m talking with Mustafa Suleyman, the CEO of Microsoft AI. As you’re no doubt aware, the biggest story in tech right now is the spiraling debate about AI safety and regulation. It should come as no surprise that Mustafa has strong opinions on how AI should be built and regulated. Microsoft just published a […]

Action tokenizers play a central role in autoregressive vision-language-action (VLA) models, determining both the targets for policy training and the executable commands recovered from predicted tokens. Their fidelity is commonly evaluated using pointwise reconstruction metrics such as mean squared error (MSE), yet small individual errors do not fully characterize how faithfully action adjustments across demonstrations are preserved. After compression, similar actions may still cluster around a

Hey everyone, Jovan from UkisAI here, a small lab building the tech to make tiny frontier LLMs possible (and doing it open-source!) The purpose of this post is simply to thank the community for all the amazing finetunes, quantizations and overall improvements over our original release which made our model get attention and the support for us to continue building in this direction! If it weren't for you guys going out of the way to contribute we wouldn't have half the results of this. For context

The testing & evaluations platform for MCP servers Discussion | Link

arXiv:2609.17695v1 Announce Type: new Abstract: A large language model (LLM) agent can follow more graph paths without acquiring more independent evidence. GraphEcho tests whether agents mistake these repeated encounters for additional corroboration. The benchmark varies path counts and evidential origins while holding evidence content fixed, and evaluates both judgments and active exploration. Controlled synthetic experiments reveal model-dependent judgment shifts, but redundant supporting path

arXiv:2609.17696v1 Announce Type: new Abstract: Document repositories evolve continuously. Guidelines and policies are revised, superseded, and re-uploaded, so the same content recurs in different wording and newer versions refine or contradict earlier ones. These inconsistencies belong to the growing collection rather than to any single document, yet existing work treats versioning, duplicate detection, and contradiction detection as isolated pairwise tasks and stops once a pair is labeled. We

arXiv:2609.17804v1 Announce Type: new Abstract: Large language models solve grade-school math word problems with high accuracy, yet a single irrelevant clause inserted into the problem can collapse it. We reconcile these observations with a mechanistic account. We show that the model's internal computation decomposes into a four-stage sequential pipeline, Schema Abstraction, Operation Planning, Operand Binding, and Computation, each stage producing a distinct intermediate representation in an id

Enabling robots to adapt to unfamiliar environments as readily as humans remains a moonshot goal of embodied AI. No finite collection of demonstrations can cover every task and situation a robot will encounter, making the ability to learn from context at deployment essential for generalization. Such in-context learning (ICL), however, remains largely beyond the reach of existing robotic policies. The broad agentic capabilities of commercial vision-language models (VLMs), such as GPT-6 Astra, rai

Cloud Macs your coding agents can actually build on Discussion | Link

Automated QA for AI agents. Stop shipping on vibes. Discussion | Link

Hey all, Henry from Cactus Compute here, I kinda wanted to share our latest model and get feedback from the family :) Needle 3 is a small foundation model for automation: you give it the functions your app exposes, it reads a request and returns the calls with every argument filled in, or a typed record if what you gave it was a schema. It runs on the device, with no network in the loop. It is on Hugging Face , on GitHub , on PyPI as cactus-needle , and there is a sandbox that runs it in your br

I was testing this quantization IQ3_XXS from GSQ-RCO with PI. It is a heavy quantization case, the model is in IQ3_XXS and KV cache in (Q4_0, Q4_0). I asked it to make the flight simulator, using that popular prompt. For my surprise, when I went verify the session I saw some screenshots, but I don't have mcp or tools in the PI, it's an almost clean installation, only with some personal skills that dont have any mentions to browser testing. Investigating the session I have noted that it found its

I was looking at neovim's donation footer at the bottom of their site and saw a bitcoin donation address. Thought I'd check how much in donations they've gotten. And I saw this massive 10 Bitcoin donation from back in 2023 (worth $800,000 now...) From the activity history, neovim last sent bitcoin out of the address in 2019 so it's been max 7 years since they've definitely had access. Does anyone from the neovim project know about this? Seems like a pretty significant amount of funding to have s

Hi HN! flat.social is a fun spatial online meeting app for remote teams and communities. It's largely built like a web multiplayer game, and the 3D virtual spaces are highly customisable with elements and activities (I recently added surfing!). Live demo to try it out: Here's a quick demo video: It's a solo, bootstrapped project built with Three.js, LiveKit and Rapier for physics. I built the first version during the pandemic and posted it here back then: I had a couple of months off this year,

Google DeepMind just launched an institute to hash out the big AGI questions in public

The Dreamforce conference became an unlikely battleground for the CEOs of OpenAI, Anthropic, and Nvidia to debate whether AI development should slow down.

How To Write With An LLM Thomas Ptacek on using LLMs as copyeditors, not as writing assistants: Rule Number One: You may not use a single word an LLM suggests to you. [...] I think that as a form of intellectual personal protective equipment you should adopt the rule that any specific turn of phrase an LLM suggests is off limits. Be strict about the rule! I won't let LLMs write content for my blog, but I use them for fact-checking, spelling and grammar and as an occasional thesaurus (see my proo

The shift comes after a UNICEF test found leading AI models struggled to accurately retrieve global development statistics.

I kept seeing engineers share what they were building with AI; however, I was always more curious about how they worked. Which agents did they use? What skills and tools had stuck or been thrown out the window? How did they manage longer-running tasks? So I built this with the hope we could have a dedicated space to share and be open about our setups.

We introduce Zing-0.5, a 5B autoregressive world model designed for playability: users can explore generated worlds, influence unfolding events, and respond to the resulting feedback through joint keyboard and online text control. Our approach brings together three technical contributions: (1) Unified action and text conditioning, combining magnitude-aware keyboard inputs with temporally aligned text instructions and jointly annotated videos to learn navigation and event control within the same

A dash of cold water keeps the foomers away.

Community submitted benchmarks for Local AI Discussion | Link

arXiv:2609.17635v1 Announce Type: new Abstract: City pedestrian counting systems now feed economic indicators, planning decisions and safety operations, yet the twins built on top of them treat the incoming stream as ground truth. We study what happens when it is not. We formalise stealthy false data injection for city-scale pedestrian sensing, where the map from latent flow to observation is far more rank deficient than in the power and water networks for which stealth has been characterised. O

arXiv:2609.17775v1 Announce Type: new Abstract: Enterprise guideline documents mix narrative text, complex tables, and embedded images, and converting them into structured work artifacts still takes two to three days of manual effort each. Current language and vision-language models extract from such documents but offer no governed workflow beyond extraction: no validation, no consistency checking, no traceable artifact generation. We introduce SAGE, a governed multi-stage LLM pipeline organized

Today, we are introducing the Life Sciences Verification Program (LSVP), which gives life science professionals access to our Mythos, Opus, and Sonnet models with a refined set of safeguards more permissive for biology-related work. We have already onboarded dozens of organizations through an early access program, and are now opening applications to the broader life science community ( apply here ). The program is launching in beta, initially for teams and institutions. We will continue to impro

The AI boom is becoming a materials challenge. As AI pushes computing into new territory, the materials behind that infrastructure are becoming just as crucial as the algorithms running on it. Semiconductors and data centers are approaching physical limits around performance, thermal management, electrical efficiency, and reliability, creating new demands for materials that can do…

It sounds like a stupid question, and I do apologize if it is... but I've seen several people mention that coding models are better uncensored due to the fact that they don't have to constantly run prompts through the "is this okay" sort of checks. Is this hogwash? Is it true? And more importantly, does anyone have any sources to confirm it? Anecdotal evidence is fine too if you've tried and compared them. Personally, I have never bothered because I'm too worried the de-censoring would damage th

I posted something about a 9b model a few days ago. The real problem was 2 fold. Someone suggested the size was too big to prove it all out. It's a fair argument. I needed the model depth though. The second problem was the "ability" of these models and the fact that labs (with money) produce these models still. No real usefulness in my mind. A 2b model? maybe interesting. Want a chatbot? - this could be it. Further, the Llama licensing at the core of the model was highly problematic and had to b