## The big picture

**Meta releases Muse Spark Open**
Meta has released Muse Spark Open. It is described as the most capable US open-weight model to date.
[Meta](https://research.meta.ai/blog/muse-spark-open)
**Gemini 4 Argon agents defeat Minecraft**
A team of four Gemini 4 Argon agents autonomously defeated the Ender Dragon in Minecraft without human intervention, writing 2,783 programs and coordinating via in-game chat. This marks the first end-to-end autonomous completion of the game using standard libraries, demonstrating advanced multi-agent planning and debugging.
[arXiv](https://arxiv.org/abs/2610.22319v1)

## Architectural breakthroughs

**Virtual Head Attention**
The authors propose Virtual Head Attention (VHA) to resolve the trade-off between KV cache size, compute, and quality in long-context LLMs. By applying lightweight linear operations (Q Premix and Linear Postmix) to recover head diversity and fuse features, VHA improves all three metrics without nonlinear overhead.
[HF Daily Papers](https://huggingface.co/papers/2610.17959)

**Rethinking Routing in MoE**
The authors identify an exploration dilemma in modern Mixture-of-Experts models where shared experts reduce token representation variability, conflicting with the need for exploration in fine-grained segmentation. MirrorMoE is proposed to reconcile this by shrinking the routing space while maintaining expressiveness.
[arXiv](https://arxiv.org/abs/2610.22353v1)

**FlashMask-3**
FlashMask-3 introduces a mask-aware Hopper kernel that compresses structured attention masks from O(n^2) to O(n) using column-wise sparse masks. This enables efficient multi-node context parallelism with zero-communication mask localization, addressing a key bottleneck in long-context training.
[HF Daily Papers](https://huggingface.co/papers/2610.17901)

**MJev-Spatial**
MJev-Spatial treats spatial understanding tasks as a unified classification framework, scoring candidates using pretrained query and key projections without adding parameters. This System 1 approach avoids the latency of autoregressive decoding for compact outputs like distances or points.
[HF Daily Papers](https://huggingface.co/papers/2610.22276)

## New open-weight model releases

**Muse Spark Open**
Meta’s 816B parameter model with 42B active parameters is now available, offering frontier-level capabilities in an open-weight format. It represents a significant milestone for the open-source community, raising the baseline for accessible high-performance models.
[Hugging Face](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B)

**Flux.3 Image Klein**
Black Forest Labs released Flux.3 Image Klein, a new image generation and editing model. The release includes both standard and ComfyUI-compatible variants, supporting high-fidelity image-to-image tasks.
[Hugging Face](https://huggingface.co/black-forest-labs/flux-3-image-klein)

## Hardware & optimization

**TokenSpeed**
TokenSpeed is a speed-of-light LLM inference engine.
[GitHub](https://github.com/lightseekorg/tokenspeed)
## Also this week

- **OpenGCUA**: An open-source stack for generalist computer-use agents, emphasizing code execution for task orchestration. [HF Daily Papers](https://huggingface.co/papers/2610.22200)
- **CoDA**: A co-evolution framework for simulation data engines and action policies in robotics, using shared memory to address training utility gaps. [arXiv](https://arxiv.org/abs/2610.22362v1)
- **MAI-Gym**: An all-in-one platform for GUI agent evaluation and RL training across mobile and computer environments, supporting 20,000 parallel instances. [HF Daily Papers](https://huggingface.co/papers/2610.20939)
- **FDE-Gym**: A benchmark for forward-deployed engineering tasks, capturing ambiguity and live-system constraints missing from standard agent evaluations. [arXiv](https://arxiv.org/abs/2610.22350v1)
- **VLA-Δ**: A framework for Vision-Language-Action models that uses explicit physical state transition supervision to improve robust robot control. [HF Daily Papers](https://huggingface.co/papers/2610.21306)
- **Data World**: A multi-agent engine that synthesizes data environments with verifiable ground truth, addressing bottlenecks in training data agents. [HF Daily Papers](https://huggingface.co/papers/2610.21656)
- **OneWM**: A unified causal training framework investigating the reciprocal benefits of world and action modeling in robotic systems. [HF Daily Papers](https://huggingface.co/papers/2610.21095)
- **T2-z**: A method for test-time optimization of environment latents in robotic world models to handle unobserved physical factors. [arXiv](https://arxiv.org/abs/2610.22364v1)
- **Cowork-Gym**: A framework for synthesizing interactive, multi-tool work environments for cowork agents using a build-execute-revise loop. [HF Daily Papers](https://huggingface.co/papers/2610.17670)
- **PathFM**: A trajectory-based forecasting model using flow matching to generate full future paths, addressing limitations in marginal quantile predictions. [arXiv](https://arxiv.org/abs/2610.22358v1)
- **SimForcing**: A simulation-guided framework for transferring motion priors to real-domain robot world models via latent-motion distillation. [HF Daily Papers](https://huggingface.co/papers/2610.06598)
- **3DCodeVerse**: A research ecosystem enabling coding agents to create editable 3D content at scale, including a 2M dataset and verification harness. [arXiv](https://arxiv.org/abs/2610.22341v1)
- **H-DDCM**: A scalable generative compression framework using Hadamard codebooks for fast image and video compression without retraining. [HF Daily Papers](https://huggingface.co/papers/2610.18125)
- **DSP**: A framework for learning dexterous manipulation from supervised play, using world models for model-predictive control. [arXiv](https://arxiv.org/abs/2610.22340v1)
- **Evaluation Awareness in Indic Languages**: A study showing LLMs exhibit evaluation awareness and increased refusal rates in five Indic languages, similar to English. [arXiv](https://arxiv.org/abs/2610.22333v1)
- **SUPRA-TEX**: A method for 3D generation with super-aligned textures, decoupling material estimation and surface alignment to reduce artifacts. [HF Daily Papers](https://huggingface.co/papers/2610.22352)
- **KnowGPO**: A reinforcement learning framework for search agents that aligns trajectories by task-relevant knowledge rather than surface observations. [arXiv](https://arxiv.org/abs/2610.22331v1)
- **ForcingBooster**: A framework for adaptive patch-wise denoising in autoregressive video diffusion, optimizing compute based on spatial heterogeneity. [HF Daily Papers](https://huggingface.co/papers/2610.12625)
- **LangDiT**: A Transformer architecture optimized for continuous diffusion language models, reducing generation perplexity. [arXiv](https://arxiv.org/abs/2610.22330v1)
- **SERO**: A method for long-horizon agentic RL using executable oracle continuations to address sparse reward challenges. [HF Daily Papers](https://huggingface.co/papers/2610.19397)
- **ACEdit**: A training scheme for stable multi-turn image editing via asymmetric conditioning and self-forcing. [arXiv](https://arxiv.org/abs/2610.22328v1)
- **Market-based Post-Training**: A framework for jointly learning action policies and self-orchestration in decentralized agent swarms. [arXiv](https://arxiv.org/abs/2610.22322v1)
- **CoWorld Reasoner**: An agentic framework using video world models as a unified visual reasoning engine for flexible imagination. [HF Daily Papers](https://huggingface.co/papers/2610.21910)
- **Steering LLM Agents**: A mechanism to address goal drift in long-horizon agents using a steering agent trained with reinforcement learning. [arXiv](https://arxiv.org/abs/2610.22316v1)
- **MindEval2**: A clinically grounded evaluation framework for mental health AI, featuring realistic patient simulations and explicit rubrics. [arXiv](https://arxiv.org/abs/2610.22314v1)
- **Linear Knowledge Generalization**: A study showing that knowledge linear before fine-tuning generalizes better after it, distinguishing active from inert facts. [arXiv](https://arxiv.org/abs/2610.22300v1)
- **AI Consilium**: A benchmark for LLM clinical multidisciplinary second opinions, evaluating factual grounding and clinical coherence. [arXiv](https://arxiv.org/abs/2610.22298v1)
- **Anthropic IPO Prep**: Anthropic reports Q3 revenue of $16.8B, up nearly 50% from Q1, ahead of its IPO. [Simon Willison](https://simonwillison.net/2026/Oct/15/anthropic-tells-investors/)
- **Krea 3**: A specialized image model built for editing at scale. [Krea](https://www.krea.ai/krea-3)
- **OpenViking**: A self-evolving context database for AI agents, unifying memory, RAG, and skills. [GitHub](https://github.com/volcengine/OpenViking)
- **phone-harness**: A harness for controlling phones via agents. [GitHub](https://github.com/ShawnPana/phone-harness)
- **mcp-for-blender**: A community plugin to control Blender 3D with any LLM. [GitHub](https://github.com/ahujasid/mcp-for-blender)
- **ai-agent-book**: An open-source book on agent design. [GitHub](https://github.com/bojieli/ai-agent-book)
- **Agents disprove Polysemanticity**: A discussion on claims disproving polysemanticity in mechanistic interpretability. [Latent Space](https://www.latent.space/p/ainews-agents-disprove-polysemanticity)
- **Neural networks are clean**: This is a theoretical claim regarding the properties of neural networks. [Goodfire](https://already-clean.goodfire.com/)
- **ASML Hybrid Bonding**: ASML is reportedly building a hybrid bonding tool for semiconductor manufacturing. [SemiAnalysis](https://newsletter.semianalysis.com/p/asml-is-building-a-hybrid-bonding)
- **Chirp Pong**: A two-player Pong game using audio chirps for networking between phones. [Simon Willison](https://simonwillison.net/2026/Oct/15/chirp-pong/)
- **AI for Hardware RE**: Using AI to reverse engineer hardware. [GNOME](https://blogs.gnome.org/uraeus/2026/10/15/aiandreverseengineerhardware/)
- **E Ink Kaleido 4**: Launch of the latest color ePaper platform. [E Ink](https://www.eink.com/news/detail/EInk-Launches-Latest-Version-of-Color-ePaper-Platform-E-Ink-Kaleido-4)
- **Simons Foundation AI Math**: A $15M effort to help mathematicians navigate the AI era. [Simons Foundation](https://www.simonsfoundation.org/news/simons-foundation-launches-15m-effort-to-help-mathematicians-navigate-ai-era/)
- **Why We're Banning AI**: A discussion on AI bans. [Current Affairs](https://www.currentaffairs.org/news/why-were-banning-ai)
- **Market Selloff**: Rising rates test the AI data center deal boom. [Reuters](https://www.reuters.com/business/finance/market-selloff-rising-rates-test-ai-data-centre-deal-boom-2026-10-15/)
- **Call Center AI**: Humans still outperform AI in 14,000 call center interactions. [Bold Agents](https://boldagents.co/blog/call-center-jobs-declining-14000-calls-humans-outperform-ai)
- **ChatGPT Ads**: ChatGPT targets ads based on prompts, phone number, and email. [404 Media](https://www.404media.co/how-chatgpt-targets-ads-to-you-based-on-your-prompts-phone-number-and-email/)
- **Danish Supreme Court**: AI-generated CSAM is ruled as CSAM. [Tangora](https://domstol.fe1.tangora.com/Domsoversigt-(H%C3%B8jesteret)
- **Companies Pulling Back**: Some companies are retracting AI adoption after initial rushes. [The Guardian](https://www.theguardian.com/technology/2026/oct/13/ai-companies-rushed-workers-fallout)
- **RIAA Flags**: The RIAA has flagged Anna's Archive and an AI voice cloner to the U.S. government. [TorrentFreak](https://torrentfreak.com/riaa-flags-annas-archive-and-ai-voice-cloner-to-u-s-government/)
- **AI Power Optimization**: AI-driven timing-aware power optimization for EDA. [HF Daily Papers](https://huggingface.co/papers/2610.18890)
- **Speech Event Detector**: A model for timestamped detection of non-verbal vocalizations and background sounds. [HF Daily Papers](https://huggingface.co/papers/2610.14232)
- **BeBox Emulator**: A Haiku application emulating Be Inc.'s BeBox. [GitHub](https://github.com/johnrash/bebox)
- **Frontier Pre-Training**: Insights into Microsoft AI's frontier pre-training methods. [Microsoft AI](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/)
- **MLExpert**: A generalist agentic model for autonomous machine learning engineering. [arXiv](https://arxiv.org/abs/2610.22344v1)
- **OmniVerifier-C1**: A decomposed graded verifier for rule-governed generative world modeling. [arXiv](https://arxiv.org/abs/2610.22355v1)