Latest news, beside the leaderboard

On Oct 11 the vLLM Semantic Router (vllm-sr) team released the Decision 3.0 family (d3, d3-flash, d3-mini, d3-nano, d3-lite, d3-edge) on Hugging Face under Apache-2.0. Built on Qwen3.5/3.8, the models take text, images and video and return probabilities for choice, yes/no and score questions in one call without generating text. Vendor-reported: the 27B d3 scores 64.7 on Jev Decision Index 0.3.1, ahead of Perplexity Decider v1.1 (62.8) and Decision 2.0 (55.9), with a 55 ms median text latency on one AMD MI325X.

Interactive virtual environments allow embodied AI agents to learn through exploration and practice, but creating worlds that are simultaneously faithful (rigid state dynamics and programmatic rules) and realistic (photorealistic visual distributions) has remained a critical bottleneck: physics engines lack visual fidelity, whereas generative video models suffer from compounding visual hallucinations during extended agent interactions. MirroS Lab and researchers introduce AgentGarten (arXiv:2610.12374, repo: github.com/MirroS-Lab/AgentGarten, site: mirros-lab.github.io/agent-garten), an open-source framework that decouples physics simulation backends from a shared neural renderer. Simulators maintain ground-truth state and execute code-defined rules without ingesting rendered frames, strictly preventing error accumulation. Visual observations are synthesized in real-time (>30 FPS) from exported geometric conditions via a neural renderer trained with Adversarial Forcing, making history prefilling differentiable through exact replay. Agents learn through visual perception, distilling execution trajectories into playbooks that succeeding generations inherit. In benchmarks such as hide-and-seek, agents autonomously discover obstacle-moving and ramp-climbing tactics in just 4 to 10 rounds (compared to tens of millions of rounds in classic RL), with code and environments fully open-sourced.

Token-level LLM routing—routing easy tokens to small models and switching to frontier models for critical reasoning tokens—presents an optimal trade-off between output quality and inference cost. However, current LLM serving engines (such as vLLM and TGI) are architected for single-model execution, suffering severe step desynchronization, batch admission latency, and extreme orchestration complexity when handling high-frequency token transfers across heterogeneous models. Researchers from Tsinghua University's NICS Lab present TokenRouter (arXiv:2610.12242, accepted at NeurIPS 2026, code: github.com/thu-nics/TokenRouter), an efficient and developer-friendly serving system engineered for token-level routed LLM inference. TokenRouter introduces a 'request-centric programming and model-centric execution' paradigm, enabling developers to write routing logic intuitively per request while the runtime handles asynchronous execution across subservers. Incorporating a Delayed-Batching Scheduler parameterized via discrete-time Markov chain modeling, a decoupled tri-loop architecture, and an instant handoff-resume mechanism, TokenRouter achieves 2.01x to 64.15x higher decoding throughput compared to existing serving platforms across diverse routing workloads, with complete code open-sourced.

Reinforcement learning (RL) post-training is foundational to scaling reasoning capabilities in large language models. However, its memory footprint remains a prohibitive bottleneck: storing full-sized gradient buffers alongside multi-state Adam optimizer moments frequently triggers out-of-memory (OOM) failures when attempting to train large models (such as 27B parameters) on standard compute hardware. In newly released research (arXiv:2610.06647), researchers introduce LoGRA, an efficient RL post-training framework available within the open-source Molt library (github.com/skzhang1/labs-molt). LoGRA compresses training representations by retaining useful learning signals inside compact low-rank gradient sketches accumulated directly during backpropagation, facilitating both memory-efficient parameter updates and streamlined policy synchronization. To counteract instability from gradient compression, LoGRA introduces a Predicted-KL step control mechanism that forecasts policy change magnitudes prior to parameter updates, dynamically attenuating disruptive updates. Evaluated across reasoning benchmarks, LoGRA reduces average training memory by up to 45.7% with zero degradation in task performance, successfully sustaining stable RL training of a 27B model for over 1,100 steps on a single eight-GPU node where standard dense Adam crashes with OOM.

On Oct 9 the vLLM team, with Inferact, Red Hat and NVIDIA, shipped early Vera Rubin NVL72 support: CUDA 13.4 nightly images already serve DeepSeek, Kimi, GLM and MiniMax. Vendor-reported results: up to 7.84x per-GPU throughput vs GB200 on SemiAnalysis AgentX (MiniMax M3, matched interactivity) and up to 3.7x vs GB300 NVL72 on the MLPerf Inference v6.1 VLM benchmark (Qwen3-VL-235B-A22B).

Anatomy (MIT) is an Agent Skill that has Claude Code invent a physical machine whose mechanism embodies a concept, then build it as an interactive isometric SVG/WebGL/3D (beta) figure with its own geometry audit and headless-Chrome self-checks. Created Oct 8, v1.0.0 shipped Oct 10, ~1,970 stars and 150 forks so far.

Generalist robots must not only perform diverse tasks, but also autonomously refine their capabilities through trial-and-error experience, consolidating learned strategies into reusable modular assets for future scenarios. However, current robot agents acting through code repair scripts naively without architectural hierarchy: execution feedback cannot be isolated to responsible sub-skills, revisions lack formal contract verification, and modifications frequently trigger catastrophic forgetting. Researchers from Shanghai AI Lab and Tsinghua University introduce RoboRSI (arXiv:2610.12424, code: github.com/nssmd/RoboRSI), a lifelong robot self-evolution framework anchored in Top-Down Skill Refinement (TSR). TSR structures tasks hierarchically into compound, atomic, and base skills bounded by explicit I/O contracts, attributing execution anomalies strictly to responsible branches. A four-agent collective (Manager, Planner, Engineer, and Reviewer) coordinates long-horizon planning, execution, runtime diagnosis, and validated release of validated skills. Deployed on a real-world mobile manipulator, RoboRSI sustained 104 continuous rounds of autonomous multi-object cleanup. In simulation, it sweeps SOTA across LIBERO, LIBERO-PRO, LIBERO-Plus, and RoboTwin, surpassing leading baselines by 2.7 to 11.0 percentage points, with code fully open-sourced.

While 3D Gaussian Splatting (3DGS) has revolutionized photorealistic static 3D scene reconstruction, synthesized digital worlds remain frozen in time. Existing dynamic generation frameworks either constrain motion to simplistic Eulerian fluid simulations or synthesize finite non-looping video trajectories that rapidly degrade when explored from novel camera viewpoints. Researchers unveil OuroWorld (arXiv:2610.12461, site: ouroworld.userwei.com), the first mask-free framework capable of turning any static 3DGS scene into an endlessly looping 3D cinemagraph with seamless temporal continuity across arbitrary 6-DoF exploration trajectories. OuroWorld employs a vision-language model to infer plausible environmental physics and direct a foundation video model to synthesize a reference video, which is subsequently lifted and completed across multiple views. To reconcile cross-view generation discrepancies, the authors propose Inconsistency-Robust Periodic 4DGS: a Fourier-series deformation field guarantees seamless looping by construction, while a Grounded Drift Field anchored at the reference camera view absorbs cross-view generative noise. Evaluated across 39 complex reconstructed and generated scenes, OuroWorld decisively outscores existing dynamic 3D baselines and secures a 70.8% to 99.0% preference rate in human blind evaluations.

Dense correspondence matching serves as the bedrock for optical flow, 3D stereo reconstruction, and object tracking. For decades, classical matching algorithms have fundamentally relied on rigid spatio-temporal priors: smooth motion continuity and rigid geometry. However, these foundational assumptions collapse under modern AI image editing and reference-guided generation (IEG), where visual identity is preserved while physical and geometric continuity is drastically severed (such as radical pose warps, stylistic shifts, or compositional recontextualization). Researchers from HKUST and ByteDance introduce FreeMatching (arXiv:2610.12421, code: github.com/luping-liu/FreeMatching), a generalizable framework for dense correspondence matching beyond spatio-temporal priors. FreeMatching fuses generative foundation features (from diffusion backbones) with high-level semantic representations (from vision foundation models), trained against heterogeneous supervision spanning classical benchmarks, tracked video sequences, and synthetic 3D scenes. An unsupervised teacher-guided iterative refinement loop further sharpens dense correspondences across challenging generated image pairs without requiring dense manual annotations. FreeMatching achieves dramatic leaps in correspondence precision across complex AI-edited image pairs while retaining competitive accuracy on classical optical flow tasks, and serves as an auditable quantitative metric for evaluating visual identity preservation that closely aligns with human judgment.

Per OpenAI's official Deprecations page, a batch of legacy snapshots deprecated on Apr 22 stops serving in the API on Oct 23, 2026: 12 snapshots and their aliases including gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-turbo, gpt-4.1-nano, gpt-4o-2024-05-13, gpt-image-1, o1, o1-pro, o3-mini and o4-mini, plus five fine-tuned model families. Official substitutes are gpt-5.6-sol/terra/luna and gpt-image-2.5. The #keep4o campaign to preserve GPT-4o snapshots resurfaced on X today.

In collaborative multi-robot warehousing, autonomous vehicle fleets, and multi-user spatial computing, egocentric world models serve as foundation simulators predicting environmental evolution conditioned on physical actions. However, existing world models focus almost exclusively on isolated single agents, while existing multi-agent systems restrict conditioning to coarse navigation commands. Consequently, they fail to model complex physics where multiple agents execute fine-grained, coupled manipulations within shared environments. Researchers from KAIST CVLab and Seoul National University introduce ME-World (arXiv:2610.12299, project site: cvlab-kaist.github.io/ME-World/), the first Multi-Agent Egocentric World Model. ME-World formalizes multi-agent world modeling as synchronized first-person video stream generation driven by fine-grained embodied actions in a shared world. The model jointly denoises multiple ego-streams within a unified sequence representation, conditions each stream on cross-agent target-view poses, and grounds video synthesis using persistent shared environment memory. Extensive benchmarks across real-world and synthetic datasets confirm that ME-World substantially outperforms existing world models in shared-world spatial consistency, physical update propagation, and action controllability.

Diffusion Transformers (DiTs) underpin frontier video generation systems and high-fidelity 3D modeling, yet self-attention computation over extreme sequence lengths introduces prohibitive inference latencies and memory bottlenecks. While sparse attention represents a promising acceleration pathway, existing methods suffer substantial visual quality degradation and texture corruption under high sparsity regimes. Researchers from The Chinese University of Hong Kong (CUHK) and ByteDance analyze sparse attention failures in DiTs (arXiv:2610.06801), tracing quality drops to three root causes: rigid token grouping constraints, inaccurate interaction selection, and uncompensated attention residuals from discarded tokens. The authors introduce Meta-Cached Sparse Attention (MC-Sparse, project site: dodododddo.github.io/mcsparse-project-page/), a training-free framework that preserves individual Key-Value (KV) selection precision while organizing similar Queries into tile-aligned blocks for hardware efficiency on GPUs. MC-Sparse caches query clusters, exact KV indices, and dense-sparse attention output residuals, reusing this metadata across downstream denoising timesteps. On Minimax-H3-Base video models, MC-Sparse delivers a 1.80x denoising speedup, and a 2.32x acceleration on 3D generative backbones, maintaining near-identical fidelity to dense attention with zero visible quality degradation.

Modern video foundation models synthesize photorealistic visual dynamics, yet struggle on long-horizon procedural tasks requiring sequential causal execution—such as equipment maintenance, cooking, or laboratory experiments. Current architectures rely on open-loop generation that cannot adapt to intermediate video outputs, resulting in disordered task sequences, hallucinated progress, and premature task termination. Researchers from MBZUAI and the Australian National University introduce WorldGuide (arXiv:2610.12459, repository: github.com/mbzuai-oryx/WorldGuide, site: mbzuai-oryx.github.io/WorldGuide/), formalizing procedural video generation as closed-loop task execution in visual world space. Given solely an initial image and a high-level task goal, WorldGuide iteratively predicts the next atomic action, generates the corresponding visual video clip, and inspects the synthesized state to dynamically route subsequent actions or determine task completion. A hierarchical visual memory maintains state history under bounded token footprints. To bridge supervision gaps, the authors release WorldGuide-Bench comprising approximately 59K step-annotated videos across 245 tasks and 27 procedural categories. WorldGuide achieves 33.33% Task Success on WorldGuide-Bench, outscoring the strong MiniMax-H3 model (29.90%) even when MiniMax-H3 is provided with reference plans, and surges to 47.69% on VideoCraft-Bench compared to 32.73% for MiniMax-H3 under goal-only conditioning.

Vivix released Vivix-W1, a streaming-native multimodal video model that accepts text, image, audio and video references and lets new inputs steer the stream mid-generation. Vivix says it has ~30B active parameters and costs $0.003 per output second through Oct 31; in vendor tests an 8s 720p reference-to-video job took 8.4s on average. The API is in limited preview with standard, fast and ultrafast tiers.

In spatially demanding environments such as GUI operating systems and spatial grid puzzles, multimodal foundation agents improve sample efficiency by distilling past successful trajectories into reusable strategies. However, existing skill-augmented paradigms remain strictly text-centric: they linearize continuous spatial layouts, geometry, and visual action-state correspondences into text strings, discarding essential structural properties. Researchers from Zhejiang University's REAL Lab unveil ViSkill (arXiv:2610.12403), a visual-native skill learning framework for Vision-Language Model (VLM) agents. ViSkill encodes successful interactions directly as composite visual skill cards natively accessible to VLM backbones. Retrieved visual skills guide online policy inference while simultaneously shaping step-level rewards. Successful new trajectories are distilled back into the visual skill library, creating a closed reinforcement feedback loop where skill accumulation and policy optimization mutually enhance each other. Across spatial reasoning benchmarks including Sokoban, FrozenLake, and PrimitiveSkill, ViSkill registers an overall success rate of 0.89 (climbing to 0.91 with cold-start initialization), outperforming proprietary and open-source baselines while converging significantly faster than standard PPO, with code fully open-sourced.

As autonomous coding agents demonstrate proficiency in software debugging and code generation, researchers are advancing them into discrete 3D spatial engineering and physical assembly. Researchers from Stanford University and Inria introduce BrickBench and the BrickAgent interactive environment (arXiv:2610.12452, website: brickben.ch), formulating the first comprehensive benchmark for agentic, text-conditioned LEGO set design. Given natural language prompts, coding agents must select parts from discrete modular libraries and generate executable Python code defining 3D coordinates, orientations, and interlocking poses. Generated assemblies are formally scored across structural validity, text-design alignment, and aesthetic complexity under rigorous physics-based simulation. Evaluating frontier models reveals that while leading coding agents can satisfy verifiable physical and semantic requirements, they consistently fall short of human design elegance and structural efficiency, establishing a pivotal milestone for spatial coding agents.

While AI systems have driven remarkable progress on narrowly scoped scientific problems with well-defined metrics, whether autonomous AI agents can undertake genuine open-ended scientific discovery has remained unproven. Researchers present Station (arXiv:2610.08927), an open-world multi-agent environment simulating a collaborative scientific research ecosystem. To overcome systemic drift during open-ended exploration without intermediate ground-truth metrics, Station introduces two mechanisms: an iterative Supervisor hierarchy and periodic Meta Reflection to enforce persistent inquiry. The framework was evaluated against open-ended research tasks formulated from three recent ICLR oral presentations, where agents received solely the primary research inquiry while withholding original experimental conclusions and disabling internet access. Station autonomously rediscovers an average of 62.7% of the original scientific criteria, compared to just 15.4% for Codex Multiagent-v2 and 14.4% to 20.6% for AI Scientist-v2. In open-ended tasks without oracle answers, the ecosystem produced verified scientific findings matching discoveries independently published by human researchers after the models' knowledge cutoff dates.

Autonomous deep research agents are expanding beyond text-centric search toward multimodal discovery across high-resolution imagery, complex charts, and long-horizon video. However, existing multimodal models struggle to ground fine-grained visual anchors and bridge perceptual evidence with external search verification. Researchers from Fudan University and collaborating institutions introduce OneSearch-VL (arXiv:2610.12419), a unified multimodal deep research agent for single-image, multi-image, and video environments. OneSearch-VL centers on the Visually Grounded Evidence Graph (VGEG), which explicitly encodes localized visual anchors, entity relationships, source-supported facts, and research operations into a shared dependency topology. Leveraging VGEG, the authors construct OneSearch-VL-SFT-110K and OneSearch-VL-RL-10K datasets alongside an Evidence-aware Visual-Grounded Rubric reward (EVGR) to supervise visual traceability during reinforcement learning. On the newly introduced OneSearch-MI-Bench and OneSearch-Video-Bench, OneSearch-VL-8B outperforms tool-augmented Qwen3-VL-8B by 20.2 and 17.6 percentage points respectively, while setting new state-of-the-art results across seven established image benchmarks and VideoDR, with code fully open-sourced.

Contemporary embodied intelligence relies heavily on neural network policies: foundation models either run end-to-end Vision-Language-Action (VLA) architectures or repeatedly query Vision-Language Models (VLMs) at high frequencies via agent harnesses. This paradigm imposes severe inference latency, costly deployment bills, and unpredictable non-deterministic execution failures. Researchers from S-Lab at Nanyang Technological University (NTU Singapore) introduce a paradigm shift (arXiv:2610.12369): Embodied Turing Machines and Code-Only-as-Policy (COAP). Under this formulation, the physical environment operates as a Turing machine where the tape encodes robot proprioception and visual states, while the transition rules are executed entirely by deterministic, stateful code. A shared, modular code library governs decisions across episodes without running any neural network models at test time. Crucially, explicit stateful code provides the optimal medium for Recursive Self-Improvement (RSI): autonomous coding agents iteratively author, debug, and refactor robot skills in closed simulation loops. Across 42 complex bimanual manipulation tasks in RoboDojo, the synthesized COAP library reaches a 70.24% success rate with zero neural models running at runtime, slashing deployment latency and cost to zero.

Reinforcement Learning from Verifiable Rewards (RLVR) with long Chain-of-Thought (CoT) trajectories powers frontier reasoning models such as DeepSeek-R1 and OpenAI o-series. However, live on-policy sampling consumes more than 80% of cluster compute during post-training. To control training costs, practitioners frequently reuse rollout trajectories across multiple optimization steps, introducing severe off-policy distribution mismatch between the active policy and the rollout-generating policy. Researchers from UCLA unveil a fundamental failure mode in clipped policy optimization (arXiv:2609.35433): 'Sign-Dependent Gradient Starvation'. Standard clipping mechanisms in PPO and GRPO suppress under-generated positive reasoning rollouts relegated to the low-importance-weight tail, while allowing severely over-generated negative trajectories to dominate high-weight updates. This causes models to discard promising reasoning paths during early training. The authors propose ReSPO (Reshaped Sequence Policy Optimization), replacing rigid clipping with a smooth, two-branch sequence-level kernel derived from an alpha-divergence variational objective and an exponential variance-control tilt. Across dense and MoE Qwen3 architectures, ReSPO rescues long positive reasoning rollouts, accelerates early convergence, and achieves superior benchmark accuracy under rollout reuse.