AI and science frontiers · 2026-10-20

ARIA turns when to consult a stronger model into a learned sequential decision: across SWE-bench Verified, Terminal-Bench 2.1 and a telecom customer-service benchmark it recovers 44–81% of the executor–advisor success gap at less than half the Advisor-only cost.

ARIA turns when to consult a stronger model into a learned sequential decision: across SWE-bench Verified, Terminal-Bench 2.1 and a telecom customer-service benchmark it recovers 44–81% of the executor–advisor success gap at less than half the Advisor-only cost.

Two self-distillation methods lose or reverse their reported gains on unseen problems across code, science, tool use and math, and the math improvement comes mainly from transferring thinking-mode behavior rather than from privileged information such as reference solutions.

Two self-distillation methods lose or reverse their reported gains on unseen problems across code, science, tool use and math, and the math improvement comes mainly from transferring thinking-mode behavior rather than from privileged information such as reference solutions.

In vLLM's Elastic Expert Parallelism, resizing a deployment used to stop serving: scaling DeepSeek V4 Pro from 16 to 20 GB200 GPUs paused for 99.62 seconds. The new design splits reconfiguration into background preparation and a short commit, while NIXL EP adds or removes communication ranks in place and keeps captured CUDA graphs valid, cutting the same benchmark's pause to 0.82 seconds on scale-up and 1.96 seconds on scale-down.

In vLLM's Elastic Expert Parallelism, resizing a deployment used to stop serving: scaling DeepSeek V4 Pro from 16 to 20 GB200 GPUs paused for 99.62 seconds. The new design splits reconfiguration into background preparation and a short commit, while NIXL EP adds or removes communication ranks in place and keeps captured CUDA graphs valid, cutting the same benchmark's pause to 0.82 seconds on scale-up and 1.96 seconds on scale-down.

An agent's performance depends not only on its underlying model but on its runtime harness: prompts, memory, skills, tools and orchestration. Harness2 makes harness improvement a two-level recursion. At the task level, a contrastive propose-probe-compose step compares execution behavior before and after edits without ground-truth answers; at the harnessing level, that process is distilled into an experience library that guides later tasks. On WorkBuddy-Bench, JobBench and Harvey's Legal Agent Benchmark, a single step improved both GLM-5 and Gemini 3.5 Flash across four model-harness configurations, by up to 11.5 points, 4.5 points and 11.8 percentage points.

An agent's performance depends not only on its underlying model but on its runtime harness: prompts, memory, skills, tools and orchestration. Harness2 makes harness improvement a two-level recursion. At the task level, a contrastive propose-probe-compose step compares execution behavior before and after edits without ground-truth answers; at the harnessing level, that process is distilled into an experience library that guides later tasks. On WorkBuddy-Bench, JobBench and Harvey's Legal Agent Benchmark, a single step improved both GLM-5 and Gemini 3.5 Flash across four model-harness configurations, by up to 11.5 points, 4.5 points and 11.8 percentage points.

MCMC and SMC samplers for the power distribution fail to converge under realistic compute budgets, and their reasoning gains are largely recreated by simple Best-of-N selection; Power-RLOO, which fine-tunes directly toward the power distribution, helps only when the base model is already aligned with the task.

MCMC and SMC samplers for the power distribution fail to converge under realistic compute budgets, and their reasoning gains are largely recreated by simple Best-of-N selection; Power-RLOO, which fine-tunes directly toward the power distribution, helps only when the base model is already aligned with the task.

Beam, Reflection's 501B-parameter mixture-of-experts model with 23B active parameters per token, targets software engineering and long-horizon agentic tasks and is planned for Apache 2.0 release. SGLang now supports Beam natively and powers its production API, turning inference efficiency into serving performance through BF16, NVFP4 and dynamic FP8 checkpoints, DFlash2 speculative decoding and prefix caching. On an OpenHands multi-turn agentic coding replay with the NVFP4 checkpoint on four NVIDIA GB300 GPUs, SGLang reaches 127.6K total tokens/sec per GPU at 185.87 tokens/sec per user (p50) with concurrency 32.

Beam, Reflection's 501B-parameter mixture-of-experts model with 23B active parameters per token, targets software engineering and long-horizon agentic tasks and is planned for Apache 2.0 release. SGLang now supports Beam natively and powers its production API, turning inference efficiency into serving performance through BF16, NVFP4 and dynamic FP8 checkpoints, DFlash2 speculative decoding and prefix caching. On an OpenHands multi-turn agentic coding replay with the NVFP4 checkpoint on four NVIDIA GB300 GPUs, SGLang reaches 127.6K total tokens/sec per GPU at 185.87 tokens/sec per user (p50) with concurrency 32.

Meta converted a conventional data center in New Albany, Ohio, into Prometheus, a gigawatt-scale supercluster whose first phase is complete and was used to train the Muse family of foundation models. Because one failing host can slow an entire large training job, Meta focused on system stability and reached more than 85% goodput on Muse Wonder pre-training runs and more than 90% on reinforcement learning runs. The cluster now spans hundreds of thousands of GPUs, with a target of roughly one million at full scale.

Meta converted a conventional data center in New Albany, Ohio, into Prometheus, a gigawatt-scale supercluster whose first phase is complete and was used to train the Muse family of foundation models. Because one failing host can slow an entire large training job, Meta focused on system stability and reached more than 85% goodput on Muse Wonder pre-training runs and more than 90% on reinforcement learning runs. The cluster now spans hundreds of thousands of GPUs, with a target of roughly one million at full scale.

Protein function depends on ordered transitions between conformational states, yet molecular dynamics sampling remains costly across large protein panels. TEMPO splits generation into two levels: a slow-scale model advances anchor conformations at the 20-ns training stride, and a fast-scale model generates 1-ns-spaced frames conditioned on those anchors. On held-out mdCATH and ATLAS proteins it reproduced MD ensemble observables; against a single-scale autoregressive model, 400-frame long-horizon RMSD error fell from 8.62 to 1.78 Å, and a 400-frame rollout took under 1 min.

Protein function depends on ordered transitions between conformational states, yet molecular dynamics sampling remains costly across large protein panels. TEMPO splits generation into two levels: a slow-scale model advances anchor conformations at the 20-ns training stride, and a fast-scale model generates 1-ns-spaced frames conditioned on those anchors. On held-out mdCATH and ATLAS proteins it reproduced MD ensemble observables; against a single-scale autoregressive model, 400-frame long-horizon RMSD error fell from 8.62 to 1.78 Å, and a 400-frame rollout took under 1 min.

With only black-box access to a recommendation API, injecting under 1% synthetic interactions makes a generative recommender promote an attacker's target far more often whenever a chosen trigger item appears, and short prefixes picked for decoding alignment raise the hit-rate lift by up to 16.3 percentage points.

With only black-box access to a recommendation API, injecting under 1% synthetic interactions makes a generative recommender promote an attacker's target far more often whenever a chosen trigger item appears, and short prefixes picked for decoding alignment raise the hit-rate lift by up to 16.3 percentage points.

LoOPSD builds a stronger self-teacher by repeating a searched block of a model's own layers without changing any weights, then distills that teacher's predictions into a standard-path student, lifting Qwen3-1.7B on GSM8K from 42.30% to 79.61% under a 400-token thinking budget.

LoOPSD builds a stronger self-teacher by repeating a searched block of a model's own layers without changing any weights, then distills that teacher's predictions into a standard-path student, lifting Qwen3-1.7B on GSM8K from 42.30% to 79.61% under a 400-token thinking budget.

CAPTOR, a vision-language foundation model pretrained on 25,512 chest-abdomen-pelvis CT-report pairs from cancer-focused exams, outperformed existing CT foundation models in its target setting. Disease classification reached an AUC of 0.767 versus 0.695 for the next-best model on the full chest-abdomen-pelvis subset, and 0.784 versus 0.734 on the oncologic subset. It encodes organ-centered crops of whole exams but makes study-level, not lesion-level, judgments.

CAPTOR, a vision-language foundation model pretrained on 25,512 chest-abdomen-pelvis CT-report pairs from cancer-focused exams, outperformed existing CT foundation models in its target setting. Disease classification reached an AUC of 0.767 versus 0.695 for the next-best model on the full chest-abdomen-pelvis subset, and 0.784 versus 0.734 on the oncologic subset. It encodes organ-centered crops of whole exams but makes study-level, not lesion-level, judgments.

When a test set underrepresents a semantic condition such as night, model selection can be misled; a generative pipeline uses vision-language models to find the missing semantics and diffusion models to turn day images into night images, so the augmented test set selects the same model as a fully covered one.

When a test set underrepresents a semantic condition such as night, model selection can be misled; a generative pipeline uses vision-language models to find the missing semantics and diffusion models to turn day images into night images, so the augmented test set selects the same model as a fully covered one.

Text-to-motion generation no longer needs a separate model per skeleton: once root-centered joint trajectories are arranged into joint-time images, image diffusion architectures share one motion autoencoder and one generator across HumanML3D and KIT-ML while keeping each dataset's native skeletal definition.

Text-to-motion generation no longer needs a separate model per skeleton: once root-centered joint trajectories are arranged into joint-time images, image diffusion architectures share one motion autoencoder and one generator across HumanML3D and KIT-ML while keeping each dataset's native skeletal definition.

RepProbe turns repetition count into a controlled variable and finds that generalist video language models collapse at high counts even though the count remains linearly decodable from the same final-layer state, locating the failure in the output pathway rather than the representation.

RepProbe turns repetition count into a controlled variable and finds that generalist video language models collapse at high counts even though the count remains linearly decodable from the same final-layer state, locating the failure in the output pathway rather than the representation.

Across 160 classification and 81 regression datasets, tabular foundation model rankings change with the benchmark and the reliability metric, and regression models that already sit near nominal marginal coverage still under-cover in their worst feature slab, so model selection cannot rest on aggregate predictive accuracy alone.

Across 160 classification and 81 regression datasets, tabular foundation model rankings change with the benchmark and the reliability metric, and regression models that already sit near nominal marginal coverage still under-cover in their worst feature slab, so model selection cannot rest on aggregate predictive accuracy alone.

GUI agents no longer reconstruct the next screen pixel by pixel; a shared latent representation now scores candidate actions and predicts action-conditioned futures at once, raising success rates by up to 31.9% offline and 40.6% online relative to pixel-space world models while cutting world-model inference latency by up to 98.9%.

GUI agents no longer reconstruct the next screen pixel by pixel; a shared latent representation now scores candidate actions and predicts action-conditioned futures at once, raising success rates by up to 31.9% offline and 40.6% online relative to pixel-space world models while cutting world-model inference latency by up to 98.9%.

Adding a sparsity penalty to TD-JEPA's task encoder concentrates the inferred reward near the goal, raising average return on DMC by 53 points and average success rate on OGBench by 10.4 percentage points in zero-shot reinforcement learning.

Adding a sparsity penalty to TD-JEPA's task encoder concentrates the inferred reward near the goal, raising average return on DMC by 53 points and average success rate on OGBench by 10.4 percentage points in zero-shot reinforcement learning.

MemoryVLN compresses long-horizon vision-language navigation history into pose-grounded spatial memories and reuses them across episodes within the same environment, raising success rate by about 2% on R2R-CE and 6.5% on RxR-CE.

MemoryVLN compresses long-horizon vision-language navigation history into pose-grounded spatial memories and reuses them across episodes within the same environment, raising success rate by about 2% on R2R-CE and 6.5% on RxR-CE.

Sepsis is often flagged at emergency triage, but a neural network returns only a probability that clinicians cannot check. ShAprox fits each feature's Shapley value to a curve or mean and adds them to a base rate, so the prediction is a readable sum. Across 28,072 emergency attendances at two NHS trusts, ShAprox reached 0.824 external-test AUC, comparable to the neural network's 0.818, with training and internal-test performance closer together.

Sepsis is often flagged at emergency triage, but a neural network returns only a probability that clinicians cannot check. ShAprox fits each feature's Shapley value to a curve or mean and adds them to a base rate, so the prediction is a readable sum. Across 28,072 emergency attendances at two NHS trusts, ShAprox reached 0.824 external-test AUC, comparable to the neural network's 0.818, with training and internal-test performance closer together.

AWS released TorchNeuron, a native PyTorch backend for Trainium chips, so model code, libraries and tools run without Neuron-specific adaptation. Previously, using Hugging Face models on Trainium required Neuron-specific software and configuration; now the device is simply pointed at neuron. It registers the device through PyTorch's PrivateUse1 extension point, supports eager, torch.compile and custom NKI kernels across Trainium2 and Trainium3, and ships as an Apache 2.0 public Beta on Oct. 20.

AWS released TorchNeuron, a native PyTorch backend for Trainium chips, so model code, libraries and tools run without Neuron-specific adaptation. Previously, using Hugging Face models on Trainium required Neuron-specific software and configuration; now the device is simply pointed at neuron. It registers the device through PyTorch's PrivateUse1 extension point, supports eager, torch.compile and custom NKI kernels across Trainium2 and Trainium3, and ships as an Apache 2.0 public Beta on Oct. 20.

A federated LLM server now treats each agent's LoRA update as a node in a hypergraph and passes messages along hyperedges to spot and suppress poisoned knowledge, beating Multi-Krum, FoolsGold and FLTrust on AG News and Yahoo! Answers.

A federated LLM server now treats each agent's LoRA update as a node in a hypergraph and passes messages along hyperedges to spot and suppress poisoned knowledge, beating Multi-Krum, FoolsGold and FLTrust on AG News and Yahoo! Answers.

CoVA recomposes the generative model's own VAE reconstruction features online into supervision with stronger semantics and spatial structure, cutting SiT-XL/2 FID on ImageNet 256x256 from 17.2 to 7.4 without an external vision encoder.

CoVA recomposes the generative model's own VAE reconstruction features online into supervision with stronger semantics and spatial structure, cutting SiT-XL/2 FID on ImageNet 256x256 from 17.2 to 7.4 without an external vision encoder.

On ImageNet, the Continuous Thought Machine's internal state follows an attracting fixed-point curve that moves every tick, most of that motion is shared across images, and the attention schedule is largely generic: shuffling queries between images costs only about 7% of classification accuracy.

On ImageNet, the Continuous Thought Machine's internal state follows an attracting fixed-point curve that moves every tick, most of that motion is shared across images, and the attention schedule is largely generic: shuffling queries between images costs only about 7% of classification accuracy.

Across nine ImageNet convolutional networks, how much of a layer's own input patches can be reconstructed from its output falls with depth but is more consistently predicted by the layer's effective compression, and copying pretrained singular vectors onto a random encoder raises recoverability while copying the pretrained singular-value spectrum or activation sparsity lowers it.

Across nine ImageNet convolutional networks, how much of a layer's own input patches can be reconstructed from its output falls with depth but is more consistently predicted by the layer's effective compression, and copying pretrained singular vectors onto a random encoder raises recoverability while copying the pretrained singular-value spectrum or activation sparsity lowers it.

Swapping agent action selection for a fast decision model like Jev cut per-decision cost by 66.1% to 78.4% and decision time by 73.9% to 81.5%, yet task success held near baseline on ALFWorld and fell more sharply in two mobile agents, so local savings do not automatically become delivered outcomes.

Swapping agent action selection for a fast decision model like Jev cut per-decision cost by 66.1% to 78.4% and decision time by 73.9% to 81.5%, yet task success held near baseline on ALFWorld and fell more sharply in two mobile agents, so local savings do not automatically become delivered outcomes.

When a fetus moves during an MRI scan, the 2D slices are acquired at shifting angles, leaving gaps in the final 3D image that can cause a radiologist to miss or misread a brain abnormality. Researchers at MIT, Boston Children's Hospital and Inria built E(3)-Pose, a deep learning model that tracks the fetal head during imaging to help capture a high-resolution image a radiologist can examine. The model embeds left-right symmetry of the fetal head directly in its architecture rather than learning physics only from data, and it is trained mainly on third-trimester fetuses.

When a fetus moves during an MRI scan, the 2D slices are acquired at shifting angles, leaving gaps in the final 3D image that can cause a radiologist to miss or misread a brain abnormality. Researchers at MIT, Boston Children's Hospital and Inria built E(3)-Pose, a deep learning model that tracks the fetal head during imaging to help capture a high-resolution image a radiologist can examine. The model embeds left-right symmetry of the fetal head directly in its architecture rather than learning physics only from data, and it is trained mainly on third-trimester fetuses.

Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model, and SGLang and Miles added inference and reinforcement-learning training backends on launch day. Its text backbone has 816B parameters with about 42B active per token, accepting text, image, video and audio with a 1M-token context. On eight B300 GPUs with NVFP4 weights, an OpenHands multi-turn agentic coding replay reaches 80.3K total tokens/s per GPU at 61 tokens/s per user at concurrency 128.

Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model, and SGLang and Miles added inference and reinforcement-learning training backends on launch day. Its text backbone has 816B parameters with about 42B active per token, accepting text, image, video and audio with a 1M-token context. On eight B300 GPUs with NVFP4 weights, an OpenHands multi-turn agentic coding replay reaches 80.3K total tokens/s per GPU at 61 tokens/s per user at concurrency 128.

Meta replaced the orchestration layer of its data-pipeline agents with a thin router that only sequences work and runs reliability checks, leaving domain reasoning to 14 specialized subagents with isolated context. A single agent loaded with all domain knowledge worked in demos but began hallucinating in production, mixing naming conventions and code-generation patterns across subsystems. Configuration prep for a new data column dropped from three to four days by hand, or about a day with the earlier config-driven system, to under an hour, with about two-thirds of changes landing without significant human edits.

Meta replaced the orchestration layer of its data-pipeline agents with a thin router that only sequences work and runs reliability checks, leaving domain reasoning to 14 specialized subagents with isolated context. A single agent loaded with all domain knowledge worked in demos but began hallucinating in production, mixing naming conventions and code-generation patterns across subsystems. Configuration prep for a new data column dropped from three to four days by hand, or about a day with the earlier config-driven system, to under an hour, with about two-thirds of changes landing without significant human edits.

Machine learning engineers at Jane Street, working with NVIDIA, rewrote the Mamba-2 forward pass in NVIDIA CUDA Tile (cuTile), a Python-based GPU programming model, and tuned it for Blackwell Tensor Cores. Mamba-2 combines a fixed-size recurrent state with attention-style matrix multiplication, but its forward pass still needs efficient kernels. On a single B300, the kernels run 1.14x to 1.82x faster than a Triton implementation, with speedups growing as sequence length and state size increase; these are kernel-level results, and end-to-end training also depends on the backward pass.

Machine learning engineers at Jane Street, working with NVIDIA, rewrote the Mamba-2 forward pass in NVIDIA CUDA Tile (cuTile), a Python-based GPU programming model, and tuned it for Blackwell Tensor Cores. Mamba-2 combines a fixed-size recurrent state with attention-style matrix multiplication, but its forward pass still needs efficient kernels. On a single B300, the kernels run 1.14x to 1.82x faster than a Triton implementation, with speedups growing as sequence length and state size increase; these are kernel-level results, and end-to-end training also depends on the backward pass.

Running Hugging Face models on AWS Trainium now rests on a different foundation: TorchNeuron, a native PyTorch backend, exposes Neuron chips as an ordinary device with eager execution, torch.compile and native distributed APIs, replacing the PyTorch XLA-based optimum-neuron. By Hugging Face's comparison, out-of-the-box support among the 100 most-used text-generation architectures rises from 46% with Optimum Neuron v0.4.5 to 89%. The backend enters public beta on October 20, 2026.

Running Hugging Face models on AWS Trainium now rests on a different foundation: TorchNeuron, a native PyTorch backend, exposes Neuron chips as an ordinary device with eager execution, torch.compile and native distributed APIs, replacing the PyTorch XLA-based optimum-neuron. By Hugging Face's comparison, out-of-the-box support among the 100 most-used text-generation architectures rises from 46% with Optimum Neuron v0.4.5 to 89%. The backend enters public beta on October 20, 2026.

NVIDIA has released the complete Run:ai platform as open source software within its DSX AI Factory platform, giving customers, developers and ecosystem partners access to production-grade AI workload and GPU orchestration. The platform previously shipped as a commercial product that dynamically allocated GPU resources on Kubernetes and managed workloads across the full lifecycle, from data processing and training to inference. The release includes four components that can also be adopted independently: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map, so partners can take the whole platform, build on it, or integrate single pieces.

NVIDIA has released the complete Run:ai platform as open source software within its DSX AI Factory platform, giving customers, developers and ecosystem partners access to production-grade AI workload and GPU orchestration. The platform previously shipped as a commercial product that dynamically allocated GPU resources on Kubernetes and managed workloads across the full lifecycle, from data processing and training to inference. The release includes four components that can also be adopted independently: KAI Scheduler, GPU Fractioning Operator, KAI Resource Management and Workload Map, so partners can take the whole platform, build on it, or integrate single pieces.

Meta has replaced one dedicated ID per item in its ads recommendation system with a short sequence of tokens drawn from a stable semantic vocabulary, so the model can infer relevance for new ads that have no engagement history. Paired with Meta's Generative Ads Model and Adaptive Ranking Model, the representation delivered a 0.35% lift in conversions and is used for both retrieval and ranking.

Meta has replaced one dedicated ID per item in its ads recommendation system with a short sequence of tokens drawn from a stable semantic vocabulary, so the model can infer relevance for new ads that have no engagement history. Paired with Meta's Generative Ads Model and Adaptive Ranking Model, the representation delivered a 0.35% lift in conversions and is used for both retrieval and ranking.

When a thin-walled grid conical shell is milled, the rib layout makes local stiffness jump from tool position to tool position, and the cutting force pushes the workpiece into an elastic retreat along the shell normal, so the actual wall thickness falls short of the theoretical depth of cut. The new route treats finite-element simulation as the source domain and real milling experiments as the target domain, uses transfer learning to predict the normal cutting force first, then feeds that load into an equivalent-stiffness analytical model to compute deflection. Across four test sets the method reached 93.05% to 95.27% prediction accuracy, above the CNN, PINN and TL comparison models.

When a thin-walled grid conical shell is milled, the rib layout makes local stiffness jump from tool position to tool position, and the cutting force pushes the workpiece into an elastic retreat along the shell normal, so the actual wall thickness falls short of the theoretical depth of cut. The new route treats finite-element simulation as the source domain and real milling experiments as the target domain, uses transfer learning to predict the normal cutting force first, then feeds that load into an equivalent-stiffness analytical model to compute deflection. Across four test sets the method reached 93.05% to 95.27% prediction accuracy, above the CNN, PINN and TL comparison models.

NVIDIA's RTX Spark laptops and compact desktops put local coding-agent inference, testing and deployment on a single machine. The prior workflow was split across a laptop, a cloud model API and a cloud VM. Built on the GB10 system-on-a-chip with up to 128GB of unified memory, NVIDIA reports 4.6x coding-agent performance, 2.7x fine-tuning and 1.8x code compilation versus Intel's Panther Lake 12Xe.

NVIDIA's RTX Spark laptops and compact desktops put local coding-agent inference, testing and deployment on a single machine. The prior workflow was split across a laptop, a cloud model API and a cloud VM. Built on the GB10 system-on-a-chip with up to 128GB of unified memory, NVIDIA reports 4.6x coding-agent performance, 2.7x fine-tuning and 1.8x code compilation versus Intel's Panther Lake 12Xe.

Kronos is the first open-source foundation model built for financial candlestick (K-line) sequences, pretrained on data from more than 45 global exchanges. General time-series foundation models are not designed for the high-noise character of financial data, so Kronos first quantizes continuous multi-dimensional K-lines such as open, high, low, close and volume into hierarchical discrete tokens, then pretrains an autoregressive Transformer on those tokens as one model for diverse quantitative tasks. The release includes pretrained weights of several sizes, fine-tuning scripts and a live demo forecasting BTC/USDT over the next 24 hours.

Kronos is the first open-source foundation model built for financial candlestick (K-line) sequences, pretrained on data from more than 45 global exchanges. General time-series foundation models are not designed for the high-noise character of financial data, so Kronos first quantizes continuous multi-dimensional K-lines such as open, high, low, close and volume into hierarchical discrete tokens, then pretrains an autoregressive Transformer on those tokens as one model for diverse quantitative tasks. The release includes pretrained weights of several sizes, fine-tuning scripts and a live demo forecasting BTC/USDT over the next 24 hours.

When two humanoid robots carry one freely supported rigid payload through constrained corridors, team coordination is written as executable code rather than fixed by task decomposition, with a multimodal foundation model supplying motion primitives, semantic objectives, temporary roles and bounded parameters. A verifier judges each physical transition; rejected ones trigger structured revision from the latest accepted checkpoint, and verified experience is retained for reuse. Across 15 base-task runs with three foundation-model backbones, 8 tasks completed, with every backbone achieving end-to-end success.

When two humanoid robots carry one freely supported rigid payload through constrained corridors, team coordination is written as executable code rather than fixed by task decomposition, with a multimodal foundation model supplying motion primitives, semantic objectives, temporary roles and bounded parameters. A verifier judges each physical transition; rejected ones trigger structured revision from the latest accepted checkpoint, and verified experience is retained for reuse. Across 15 base-task runs with three foundation-model backbones, 8 tasks completed, with every backbone achieving end-to-end success.

MiniMax has released M3.1 Flash, a new Flash-series model aimed at agentic tasks, coding, computer use and cybersecurity. The company says it reaches the level of leading models on complex agentic tasks, coding and computer operation benchmarks while making online inference nearly three times faster and cutting usage costs by 70%. The model is available through the MiniMax API, M Plan and MiniMax Code, with weights to be released later.

MiniMax has released M3.1 Flash, a new Flash-series model aimed at agentic tasks, coding, computer use and cybersecurity. The company says it reaches the level of leading models on complex agentic tasks, coding and computer operation benchmarks while making online inference nearly three times faster and cutting usage costs by 70%. The model is available through the MiniMax API, M Plan and MiniMax Code, with weights to be released later.

Meta has released the weights for Muse Spark Open, a model built for agentic coding and long-horizon tasks, and it is now served on Modal as a Shared Endpoint with per-token billing. It offers native multimodal understanding and a 1M-token context window, covering code writing and debugging, multi-step planning, sequential tool use and failure recovery across extended workflows. In Meta's evaluations it scores 76.4 on Terminal-Bench 2.1 with the Terminus 2 agent and 60 on DeepSWE 1.1 with Meta's preferred harness at xhigh reasoning effort.

Meta has released the weights for Muse Spark Open, a model built for agentic coding and long-horizon tasks, and it is now served on Modal as a Shared Endpoint with per-token billing. It offers native multimodal understanding and a 1M-token context window, covering code writing and debugging, multi-step planning, sequential tool use and failure recovery across extended workflows. In Meta's evaluations it scores 76.4 on Terminal-Bench 2.1 with the Terminus 2 agent and 60 on DeepSWE 1.1 with Meta's preferred harness at xhigh reasoning effort.

NVIDIA released Nemotron 3.5 Super VL, an open multimodal model that takes text, images, documents and video and returns text for always-on agents that read screens, documents and tools. SGLang supports deployment from day one, serving it with continuous batching, prefix caching and speculative decoding. The mixture-of-experts model has 120 billion total parameters and activates 12 billion per token, with built-in multi-token prediction; it reaches 243.9 tok/s on a single RTX 6000 PRO.

NVIDIA released Nemotron 3.5 Super VL, an open multimodal model that takes text, images, documents and video and returns text for always-on agents that read screens, documents and tools. SGLang supports deployment from day one, serving it with continuous batching, prefix caching and speculative decoding. The mixture-of-experts model has 120 billion total parameters and activates 12 billion per token, with built-in multi-token prediction; it reaches 243.9 tok/s on a single RTX 6000 PRO.

NVIDIA released Nemotron 3.5 Super VL, an open multimodal model now available for inference and training on Baseten, so agents can read documents, interpret application screens and analyze video directly. It builds on the Nemotron 3 Super architecture as a 120B-parameter Mixture-of-Experts model that activates 12B parameters per token, adding a vision encoder for image input. In NVIDIA's reported evaluations it scores an Elo of 1190 on GDPval versus 493 for Nemotron 3 Super, and also scores higher on SciCode and GPQA-diamond.

NVIDIA released Nemotron 3.5 Super VL, an open multimodal model now available for inference and training on Baseten, so agents can read documents, interpret application screens and analyze video directly. It builds on the Nemotron 3 Super architecture as a 120B-parameter Mixture-of-Experts model that activates 12B parameters per token, adding a vision encoder for image input. In NVIDIA's reported evaluations it scores an Elo of 1190 on GDPval versus 493 for Nemotron 3 Super, and also scores higher on SciCode and GPQA-diamond.

Inception's Mercury 3 is now available through AI Gateway, generating text by diffusion through parallel refinement rather than word-by-word sequential generation. Inception reports strong results across a range of benchmarks, generation speeds above 1,000 tokens per second, and a 1M-token context window. The model targets latency-sensitive workloads such as browser agents, real-time search, interactive education, and customer support, with improved reasoning.

Inception's Mercury 3 is now available through AI Gateway, generating text by diffusion through parallel refinement rather than word-by-word sequential generation. Inception reports strong results across a range of benchmarks, generation speeds above 1,000 tokens per second, and a 1M-token context window. The model targets latency-sensitive workloads such as browser agents, real-time search, interactive education, and customer support, with improved reasoning.

Salesforce AI Research has built Launa-Omni, a voice model for real-time enterprise conversation that keeps speech and language in one audio-adapted text model and uses a separate in-house text-to-speech model to speak the reply. Unlike pipelines that chain speech recognition, a language model and speech generation, speech is not transcribed before reasoning, so pace, hesitation and emphasis survive the first handoff. The model is significantly smaller than several larger voice models and reports competitive results on spoken instruction following, multi-turn dialogue and customer-service tasks; a simplified version already powers Slack's voice dictation.

Salesforce AI Research has built Launa-Omni, a voice model for real-time enterprise conversation that keeps speech and language in one audio-adapted text model and uses a separate in-house text-to-speech model to speak the reply. Unlike pipelines that chain speech recognition, a language model and speech generation, speech is not transcribed before reasoning, so pace, hesitation and emphasis survive the first handoff. The model is significantly smaller than several larger voice models and reports competitive results on spoken instruction following, multi-turn dialogue and customer-service tasks; a simplified version already powers Slack's voice dictation.

TypeSafe AI's decision model Jev 1.14 is now available on AI Gateway. It takes a state and a set of questions and returns typed Choice, Score and Boolean answers with probabilities instead of free-form text, answering every question in one request in parallel for classification, routing, scoring and automated verification. Version 1.14 accepts images as part of the state and adds an experimental long-context mode for documents, logs or conversation histories too long for a standard request.

TypeSafe AI's decision model Jev 1.14 is now available on AI Gateway. It takes a state and a set of questions and returns typed Choice, Score and Boolean answers with probabilities instead of free-form text, answering every question in one request in parallel for classification, routing, scoring and automated verification. Version 1.14 accepts images as part of the state and adds an experimental long-context mode for documents, logs or conversation histories too long for a standard request.

Single-cell reference mapping no longer lets class representatives drift during training. SCLSC built each type representative from a mean expression profile, so its position moved as optimization proceeded; SOLAR assigns every reference label a fixed orthogonal direction before training and lets the encoder learn only where cells land in that space. Across nine held-out batches from human immune, lung and pancreas datasets at 30 dimensions, mean query macro-F1 was 0.811, above 0.774 for matched PCA and below 0.838 for scANVI.

Single-cell reference mapping no longer lets class representatives drift during training. SCLSC built each type representative from a mean expression profile, so its position moved as optimization proceeded; SOLAR assigns every reference label a fixed orthogonal direction before training and lets the encoder learn only where cells land in that space. Across nine held-out batches from human immune, lung and pancreas datasets at 30 dimensions, mean query macro-F1 was 0.811, above 0.774 for matched PCA and below 0.838 for scANVI.

Two models can answer equally correctly and still feel very different to use. AA-PersonalityBench turns that difference into a measurement: 30 personality traits with 30 questions each, 900 in total, answered once by every model with no system prompt, compared pairwise by a blind panel of LLM judges, then rolled into 8 personality indexes covering Expressiveness, Organization, Informativeness, Transparency, Warmth, Conviction, Imagination and Slop, plus 21 communication characteristics measured programmatically from the responses.

Two models can answer equally correctly and still feel very different to use. AA-PersonalityBench turns that difference into a measurement: 30 personality traits with 30 questions each, 900 in total, answered once by every model with no system prompt, compared pairwise by a blind panel of LLM judges, then rolled into 8 personality indexes covering Expressiveness, Organization, Informativeness, Transparency, Warmth, Conviction, Imagination and Slop, plus 21 communication characteristics measured programmatically from the responses.

With voxel-level annotations on only 10% of CT scans, a heterogeneous ensemble of four independently trained models was applied to candidate-centred segmentation of 3–<6 mm pulmonary nodules. Against a single enhanced Mean Teacher, it raised Dice by 0.01488 on an internal held-out set and by 0.01118 on external LNDb without fine-tuning, with both confidence intervals excluding zero and concordant directions across three training seeds. The gain came with higher recall and lower precision, and did not establish that heterogeneous objectives beat homogeneous four-model ensembles.

With voxel-level annotations on only 10% of CT scans, a heterogeneous ensemble of four independently trained models was applied to candidate-centred segmentation of 3–<6 mm pulmonary nodules. Against a single enhanced Mean Teacher, it raised Dice by 0.01488 on an internal held-out set and by 0.01118 on external LNDb without fine-tuning, with both confidence intervals excluding zero and concordant directions across three training seeds. The gain came with higher recall and lower precision, and did not establish that heterogeneous objectives beat homogeneous four-model ensembles.

Early ovarian cancer diagnosis leans on serum markers and routine laboratory tests, yet redundant, noisy and correlated variables in high-dimensional clinical data degrade both classification and interpretability. A hybrid biomarker selection pipeline fuses correlation analysis, mutual information, XGBoost importance and SHAP relevance with particle-swarm-optimized adaptive weights, feeding an attention-enhanced multi-scale 1D convolutional network; on a public Kaggle ovarian cancer clinical dataset under stratified 10-fold cross-validation it reports 93.35% accuracy and 90.66% ROC-AUC. In ablation, the full pipeline reaches 85.51% accuracy, 19.33 percentage points above a baseline 1D-CNN without selection or attention.

Early ovarian cancer diagnosis leans on serum markers and routine laboratory tests, yet redundant, noisy and correlated variables in high-dimensional clinical data degrade both classification and interpretability. A hybrid biomarker selection pipeline fuses correlation analysis, mutual information, XGBoost importance and SHAP relevance with particle-swarm-optimized adaptive weights, feeding an attention-enhanced multi-scale 1D convolutional network; on a public Kaggle ovarian cancer clinical dataset under stratified 10-fold cross-validation it reports 93.35% accuracy and 90.66% ROC-AUC. In ablation, the full pipeline reaches 85.51% accuracy, 19.33 percentage points above a baseline 1D-CNN without selection or attention.

MiniMax's open-weight multimodal M3.1 Flash is now free on AI Gateway through Oct 22, and only when MiniMax serves the request. The model targets coding and agentic work with a 1M-token context window for long documents and multi-step agent sessions. Compared with earlier MiniMax models, MiniMax says it improves on complex agentic tasks, coding and cybersecurity, with its best coding results on application development, particularly frontend interactions, targeted bug fixes and component tests.

MiniMax's open-weight multimodal M3.1 Flash is now free on AI Gateway through Oct 22, and only when MiniMax serves the request. The model targets coding and agentic work with a 1M-token context window for long documents and multi-step agent sessions. Compared with earlier MiniMax models, MiniMax says it improves on complex agentic tasks, coding and cybersecurity, with its best coding results on application development, particularly frontend interactions, targeted bug fixes and component tests.

AI Gateway now offers Cat Jam, a stealth reasoning model free to use during the stealth period. It targets coding and agentic tasks with low-latency inference, supports function calling so agents can reach external tools, and returns structured outputs that follow a caller-supplied JSON schema. Responses carry the model's reasoning text alongside the answer. Input is text only, reasoning effort ranges from low to max, and Zero Data Retention is not available.

AI Gateway now offers Cat Jam, a stealth reasoning model free to use during the stealth period. It targets coding and agentic tasks with low-latency inference, supports function calling so agents can reach external tools, and returns structured outputs that follow a caller-supplied JSON schema. Responses carry the model's reasoning text alongside the answer. Input is text only, reasoning effort ranges from low to max, and Zero Data Retention is not available.