AI Coding Live News

Meta officially disclosed its gigawatt-scale AI supercluster Prometheus in New Albany, Ohio, housing hundreds of thousands of frontier GPUs with 85%+ pretraining goodput and 90%+ in reinforcement learning. Powered by this infrastructure, Meta's latest coding flagship Muse Wonder Full surged to #6 globally on LMSYS Code Arena: WebDev with 1726 Elo points (surpassing Claude Sonnet 5.5 and Gemini 4 Argon), lifting Meta to the #3 AI research lab worldwide.

Zoom

Meta's engineering blog on October 19 gave the first detailed look at Prometheus, the New Albany, Ohio supercluster that trained the Muse family, including Muse Wonder. Phase one is complete with hundreds of thousands of GPUs, and the full build targets about 1 million. Muse Wonder training reached more than 85% goodput in pretraining and more than 90% in reinforcement learning, with a mean time between interruptions of about 47K GPU hours, since raised above 54K by firmware fixes. It runs GB200 NVL72 and GB300 NVL72 platforms, with Vera Rubin NVL72 arriving soon.

Zoom

Modern Mixture-of-Experts architectures increasingly combine fine-grained expert segmentation with shared experts. However, ETH Zurich and HKU researchers identified an 'Exploration Dilemma' (arXiv:2610.22353, GitHub: BingcongLi/MirrorMoE): shared experts homogenize token representations, suppressing routing exploration just as fine-grained segmentation combinatorially enlarges the routing space. MirrorMoE resolves this via token-adaptive suppression of unnecessary experts, cutting pretraining perplexity across all model scales.

Zoom

Addressing the representation gap between high-level VLM semantics, video world models, and low-level control in Vision-Language-Action (VLA) architectures, PINT (arXiv:2610.16355) introduces learnable chunk-level Predictive Intent Tokens to reach 15.5% on RoboTwin 2.0-Hard without embodied pretraining (4.8x over StarVLA's 3.2%). Alongside PAIR (arXiv:2610.09016) and PKU's SimForcing (arXiv:2610.06598), these frameworks bridge perception and action with zero inference overhead.

Zoom

Two mechanistic interpretability studies uncover how LLMs express and localize internal knowledge: University of Cambridge researchers (arXiv:2610.21025) introduce parameter-trajectory crosscoding on checkpoint weights, discovering that hidden states encode task knowledge before the unembedding readout matrix learns to express it (reorganizing at 0.4-1.1% of Pythia pretraining), and that raising the readout learning rate closes this gap earlier. Meanwhile, University of Copenhagen researchers (arXiv:2610.22127) apply paired-activation crosscoders to three open-weight Mixture-of-Experts (MoE) LLMs, isolating monosemantic personalization latents and proving via causal interventions that boosting specific routed experts recovers personalized model behavior.

Zoom

Ollama launched US- and EU-hosted Ollama Cloud with strict zero data retention and zero training, adding immediate support for DeepSeek-V4.1-Flash-1017 via one-line launch commands for Claude Code and OpenCode alongside Ollama v0.40.6. Meanwhile, SemiAnalysis revealed that GLM-5.3 was built entirely by post-training the previous GLM-5.2 base model without pretraining from scratch, achieving a 'night and day difference' in cybersecurity and agentic capabilities.

Zoom

Whether LLM-based autonomous agents truly learn from interactive experience or merely regurgitate memorized knowledge remains a fundamental question in AI research. Most conventional agent benchmarks evaluate tasks (such as writing standard Python or navigating operating systems) heavily represented in pretraining corpora, creating severe data contamination that obscures whether agents are synthesizing new concepts or retrieving memorized patterns. Researchers introduce Learn2Play-Bench (arXiv:2610.08215, code: github.com/liushiliushi/Learn2Play-Bench, site: liushiliushi.github.io/learn2play-bench-website), a novel open-source benchmark constructed around text-based environments with entirely unfamiliar and counterintuitive mechanics. Agents must discover game physics and causal dependencies purely through trial-and-error interactions. The benchmark reveals two critical insights: retaining uncompressed, verbatim interaction logs decisively outperforms LLM-generated rule summaries, and the agent harness scaffolding (managing tool loops, memory, and reflection) exerts a larger impact on final performance than the choice of underlying foundation model.

Zoom

While Large Language Models rely on deep non-linear activation layers, a groundbreaking mechanistic study titled 'Your Transformer Can Hold Two Thoughts at Once' (arXiv:2609.29845) demonstrates that transformer architectures exhibit fundamental linear superposition. When inputs from two distinct textual contexts are linearly combined in latent space, the model outputs a superposition of the individual next-token probability distributions. The authors reveal this superposition is an intrinsic architectural property that decays during pretraining but can be restored via lightweight fine-tuning. Furthermore, the team introduces a guided decoding procedure that cleanly disentangles superposed representations, enabling the simultaneous generation of two coherent, independent continuations from a single forward pass.

Zoom

While general-purpose agents exhibit strong long-horizon reasoning, enterprise workflows require manipulating seven distinct artifact modalities: text, images, audio, video, structured documents, 3D meshes, and code. Pretraining native omni-foundation models from scratch is economically prohibitive, while haphazard multi-tool integration triggers state desynchronization and intermediate asset sprawl across turns. Researchers from the National University of Singapore (NUS) and S-Lab introduce Omni-IO Skills (arXiv:2609.31847, ~300 upvotes on Hugging Face). As a modular agent harness, it organizes multimodal operations into Declarative Execution Graphs paired with a persistent Asset Registry, orchestrating 27 hierarchical skills across 38 tasks. Evaluated on UniM-90, the harness elevates input-support rates for GPT-5.6 Sol and Claude Sonnet 5 from ~40% to 100%, and lifts semantic quality scores from ~27 to over 77.

Zoom

While LLM agents excel in multi-step decision-making, they transfer poorly to unseen environments. Conventional world-model approaches predict future observations, incurring high training overhead and compounding planning errors. EVOKE demonstrates that digital world knowledge is already internalized during pretraining, and failure to transfer stems from single-goal post-training that encourages superficial contextual shortcuts. By holding environment states fixed while ranking candidate actions across diverse counterfactual goals, EVOKE compels policies to elicit internalized world dynamics, boosting generalization across unseen environments.

Zoom

Chunked KV-cache compression reduces the memory and computational footprints of long-context LLMs by condensing contiguous token windows into fewer cache entries at fixed strides. Researchers from Stanford and Renmin University uncover an intrinsic vulnerability: chunking introduces a new relative positional coordinate—a token's phase relative to compression-window boundaries. The study reveals systematic 'phase sensitivity,' where retrieval accuracy for identical facts varies by up to 40 percentage points depending solely on input token phase alignments. Pretraining transformers across various KV compression configurations and conducting causal interventions, the authors demonstrate that standard aggregate benchmark scores mask recurrent blind spots, establishing a new evaluation standard for long-context compression.

Zoom