AI and science frontiers · 2026-11-06

Daedalus separates function-representation learning from operator learning: a functional autoencoder is pre-trained only on analytic functions synthesized on the fly, with no solver and no labels, and the latent space it induces is frozen and reused across tasks. On 28 benchmark problems spanning fluid dynamics, solid mechanics, acoustics, and the Helmholtz and Poisson equations, operators on the frozen representation have lower relative error than the end-to-end framework GAOT on 23 of them and train 2.1 times faster per problem on average; that comparison excludes the one-time backbone pre-training.

Daedalus separates function-representation learning from operator learning: a functional autoencoder is pre-trained only on analytic functions synthesized on the fly, with no solver and no labels, and the latent space it induces is frozen and reused across tasks. On 28 benchmark problems spanning fluid dynamics, solid mechanics, acoustics, and the Helmholtz and Poisson equations, operators on the frozen representation have lower relative error than the end-to-end framework GAOT on 23 of them and train 2.1 times faster per problem on average; that comparison excludes the one-time backbone pre-training.

Culture conditions combine continuous doses, so testing them exhaustively is impractical, and earlier work could only rank predefined candidate designs or search a discrete catalogue. LabCompass couples a flow-matching generative model with a forward surrogate that predicts cell phenotypes, using the surrogate's gradient at inference time to steer sampling toward recipes expected to produce a target cell state. In primary human cord-blood CD34+ cells, four experimental loops produced a 78-million-cell spectral flow cytometry dataset, recovered a known megakaryocyte-inducing recipe the model had never seen, and raised erythroid progenitors to 22%, roughly 1.6-fold over the control condition.

Culture conditions combine continuous doses, so testing them exhaustively is impractical, and earlier work could only rank predefined candidate designs or search a discrete catalogue. LabCompass couples a flow-matching generative model with a forward surrogate that predicts cell phenotypes, using the surrogate's gradient at inference time to steer sampling toward recipes expected to produce a target cell state. In primary human cord-blood CD34+ cells, four experimental loops produced a 78-million-cell spectral flow cytometry dataset, recovered a known megakaryocyte-inducing recipe the model had never seen, and raised erythroid progenitors to 22%, roughly 1.6-fold over the control condition.

Spatial multi-omic data pose two jobs at once: merging molecular layers such as RNA and protein measured on the same tissue section, and transferring cell-type labels from a reference dataset onto a new sample. DIRAC treats both as domain adaptation, using a graph neural network and an adversarial discriminator to learn embeddings that are insensitive to batch and modality but discriminative for cell type. On simulated and real DBiT-seq data it beat comparison methods on domain recovery and mixing, and in glioblastoma tissue it identified a microvascular proliferation domain the other methods missed.

Spatial multi-omic data pose two jobs at once: merging molecular layers such as RNA and protein measured on the same tissue section, and transferring cell-type labels from a reference dataset onto a new sample. DIRAC treats both as domain adaptation, using a graph neural network and an adversarial discriminator to learn embeddings that are insensitive to batch and modality but discriminative for cell type. On simulated and real DBiT-seq data it beat comparison methods on domain recovery and mixing, and in glioblastoma tissue it identified a microvascular proliferation domain the other methods missed.

Google Research, with Beth Israel Deaconess Medical Center and Harvard Medical School, used the photoplethysmography and accelerometer signals of a consumer smartwatch to passively identify a persistent hypertensive state without cuff calibration. Rather than estimating systolic and diastolic pressure at a moment in time, the system averages 15-second signal features over multiple days and combines them with height in a regularized logistic regression that outputs a probability. Against 24-hour ambulatory blood pressure monitoring in 196 participants free of known cardiovascular disease and not taking blood pressure medication, initial screening reached 65.8% sensitivity and 90.0% specificity, versus 55.3% sensitivity for concurrent office screening.

Google Research, with Beth Israel Deaconess Medical Center and Harvard Medical School, used the photoplethysmography and accelerometer signals of a consumer smartwatch to passively identify a persistent hypertensive state without cuff calibration. Rather than estimating systolic and diastolic pressure at a moment in time, the system averages 15-second signal features over multiple days and combines them with height in a regularized logistic regression that outputs a probability. Against 24-hour ambulatory blood pressure monitoring in 196 participants free of known cardiovascular disease and not taking blood pressure medication, initial screening reached 65.8% sensitivity and 90.0% specificity, versus 55.3% sensitivity for concurrent office screening.

NVIDIA released Kumo Relational, a pre-trained relational foundation model that predicts directly on enterprise relational databases without task-specific training, feature engineering, or flattening tables into one wide row. Across 41 tasks on 15 databases it beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines, and on the SAP SALT enterprise benchmark it exceeds AutoGluon ensembles by 8% in mean reciprocal rank. A prediction is stated as a declarative query, and the model answers it in a single forward pass using historical examples as context.

NVIDIA released Kumo Relational, a pre-trained relational foundation model that predicts directly on enterprise relational databases without task-specific training, feature engineering, or flattening tables into one wide row. Across 41 tasks on 15 databases it beats supervised relational deep learning models and expert-engineered gradient-boosted tree pipelines, and on the SAP SALT enterprise benchmark it exceeds AutoGluon ensembles by 8% in mean reciprocal rank. A prediction is stated as a declarative query, and the model answers it in a single forward pass using historical examples as context.

Protein sequence databases are growing far faster than curated annotation, and existing language models learn only from sequence likelihood, so InterPro domain and functional-site knowledge stays indirect. PALM builds on ESM2 with LoRA adaptation, learned layer aggregation and a bidirectional LSTM encoder, and uses a soft-target loss that lets one residue keep several valid labels. On 2,309,606 held-out proteins it reached domain- and site-occurrence F1 of 0.941 and 0.877, and still 0.857 and 0.828 in the subset below 20% sequence identity.

Protein sequence databases are growing far faster than curated annotation, and existing language models learn only from sequence likelihood, so InterPro domain and functional-site knowledge stays indirect. PALM builds on ESM2 with LoRA adaptation, learned layer aggregation and a bidirectional LSTM encoder, and uses a soft-target loss that lets one residue keep several valid labels. On 2,309,606 held-out proteins it reached domain- and site-occurrence F1 of 0.941 and 0.877, and still 0.857 and 0.828 in the subset below 20% sequence identity.

Single-cell perturbation screens still leave many compounds and cell contexts unmeasured, and CellCanon puts expression prediction, source-grounded explanation and observation-driven updating into one shared world state. A language model compiles drug and target evidence, and numerical readouts then guide single-cell generation; on 339 held-out SciPlex3 conditions it reaches differential-expression AUPRC 0.8067 and Macro-F1 0.8164, against 0.4938 and 0.6961 for the best baseline. In simulated experimental feedback, one measured source context per drug lifts differential-expression AUPRC to 0.9077 for unmeasured cell lines with a fixed-weight updater.

Single-cell perturbation screens still leave many compounds and cell contexts unmeasured, and CellCanon puts expression prediction, source-grounded explanation and observation-driven updating into one shared world state. A language model compiles drug and target evidence, and numerical readouts then guide single-cell generation; on 339 held-out SciPlex3 conditions it reaches differential-expression AUPRC 0.8067 and Macro-F1 0.8164, against 0.4938 and 0.6961 for the best baseline. In simulated experimental feedback, one measured source context per drug lifts differential-expression AUPRC to 0.9077 for unmeasured cell lines with a fixed-weight updater.

RPI-Zero reads a full-length protein and a whole spliced transcript in one input and returns a binding probability from sequence alone, with no structure and no interaction data for a new molecule. Its companion benchmark draws on 11.6 million CLIP peaks for 209 proteins from four public databases, splits whole sequence-similarity clusters into training and test, and matches negatives to positives in length and transcript popularity. On the unseen-RNA, unseen-protein and unseen-pair test slices its AUPRC is 0.753, 0.637 and 0.589, which is 2.7%, 7.4% and 4.6% above the second-best method on each slice.

RPI-Zero reads a full-length protein and a whole spliced transcript in one input and returns a binding probability from sequence alone, with no structure and no interaction data for a new molecule. Its companion benchmark draws on 11.6 million CLIP peaks for 209 proteins from four public databases, splits whole sequence-similarity clusters into training and test, and matches negatives to positives in length and transcript popularity. On the unseen-RNA, unseen-protein and unseen-pair test slices its AUPRC is 0.753, 0.637 and 0.589, which is 2.7%, 7.4% and 4.6% above the second-best method on each slice.

Glioblastoma malignant states were treated as parallel subtypes, and a single biopsy could not represent the whole tumour. Pairing single-nucleus and spatial transcriptomics across 12 IDH-wild-type primary tumours and 57 sampling sites revealed a trajectory from developmental-like states toward glial injury response and hypoxia, layered in space from the infiltrating edge to the necrotic core; the trajectory appears across tumour subtypes and genetically distinct subclones and is recapitulated in patient-derived organoids.

Glioblastoma malignant states were treated as parallel subtypes, and a single biopsy could not represent the whole tumour. Pairing single-nucleus and spatial transcriptomics across 12 IDH-wild-type primary tumours and 57 sampling sites revealed a trajectory from developmental-like states toward glial injury response and hypoxia, layered in space from the infiltrating edge to the necrotic core; the trajectory appears across tumour subtypes and genetically distinct subclones and is recapitulated in patient-derived organoids.

Medical self-supervised contrastive learning assumes two augmented views of the same image should share one semantic representation, but rotation and intensity perturbations can erase subtle yet critical anatomical and pathological cues, so rigid alignment suppresses task-relevant information. ProbSSL models each image embedding as a Gaussian, with the mean carrying shared semantics and the variance recording augmentation-induced variability, and aligns distributions by KL divergence so dimensions with larger variability receive weaker alignment pressure. After pre-training on CT, transfer to BTCV, MM-WHS, CC-CCII and BraTS gives higher average scores than the compared general and medical self-supervised methods on all four downstream tasks, including 84.89 mean Dice on BTCV.

Medical self-supervised contrastive learning assumes two augmented views of the same image should share one semantic representation, but rotation and intensity perturbations can erase subtle yet critical anatomical and pathological cues, so rigid alignment suppresses task-relevant information. ProbSSL models each image embedding as a Gaussian, with the mean carrying shared semantics and the variance recording augmentation-induced variability, and aligns distributions by KL divergence so dimensions with larger variability receive weaker alignment pressure. After pre-training on CT, transfer to BTCV, MM-WHS, CC-CCII and BraTS gives higher average scores than the compared general and medical self-supervised methods on all four downstream tasks, including 84.89 mean Dice on BTCV.

Standard cryo-EM pipelines first reconstruct a 3D density from 2D projections and only then fit an atomic model, so atomic-coordinate priors never reach the most under-determined stage. EB-FM trains a flow-matching prior inside a Monte Carlo stochastic-approximation expectation-maximization loop, repeatedly inverting the nonlinear imaging forward model during training, so no clean structures are needed for pretraining. On simulated cryo-EM data with known poses it recovers conformational distributions directly in atomic-coordinate space for four proteins, though predicted motion amplitudes stay below the molecular-dynamics ground truth.

Standard cryo-EM pipelines first reconstruct a 3D density from 2D projections and only then fit an atomic model, so atomic-coordinate priors never reach the most under-determined stage. EB-FM trains a flow-matching prior inside a Monte Carlo stochastic-approximation expectation-maximization loop, repeatedly inverting the nonlinear imaging forward model during training, so no clean structures are needed for pretraining. On simulated cryo-EM data with known poses it recovers conformational distributions directly in atomic-coordinate space for four proteins, though predicted motion amplitudes stay below the molecular-dynamics ground truth.

Meta's ads ranking and retrieval fleet no longer allocates capacity by hand, one model at a time. Previously each scaling decision cost days to weeks of training, and the architecture dimensions are coupled, so they cannot be enlarged independently. Axia maps embedding dimension, interaction depth and interaction width through a scaling layer into complete architectures, then uses Bayesian surrogates to pick a small number of training trials. Across 34 model iterations in two development cycles, offline normalized entropy improved by a mean 0.66% relative and the online key metric by a cumulative 0.36%.

Meta's ads ranking and retrieval fleet no longer allocates capacity by hand, one model at a time. Previously each scaling decision cost days to weeks of training, and the architecture dimensions are coupled, so they cannot be enlarged independently. Axia maps embedding dimension, interaction depth and interaction width through a scaling layer into complete architectures, then uses Bayesian surrogates to pick a small number of training trials. Across 34 model iterations in two development cycles, offline normalized entropy improved by a mean 0.66% relative and the online key metric by a cumulative 0.36%.

Iterative optimization algorithms can be turned into trainable models instead of being repeatedly rewritten as programs. The approach treats algorithm execution trajectories as computational experience, models optimization autoregressively so that the preceding generation forms the computational context for the next population, and uses a dimension-independent population-temporal architecture to capture both population interactions and iterative dynamics. The model first acquires search knowledge by imitation from heterogeneous teacher algorithms, then improves itself directly through optimization feedback, consolidating heterogeneous search knowledge into a single model that transfers to unseen, substantially larger dimensions.

Iterative optimization algorithms can be turned into trainable models instead of being repeatedly rewritten as programs. The approach treats algorithm execution trajectories as computational experience, models optimization autoregressively so that the preceding generation forms the computational context for the next population, and uses a dimension-independent population-temporal architecture to capture both population interactions and iterative dynamics. The model first acquires search knowledge by imitation from heterogeneous teacher algorithms, then improves itself directly through optimization feedback, consolidating heterogeneous search knowledge into a single model that transfers to unseen, substantially larger dimensions.

Tool-using language-model agents often need dozens of interactions before a single success-or-failure label arrives, leaving no way to tell which step went wrong. STEP-VERIFIER turns a success-labeled corpus of logged interactions into action-level comparisons: the difference in average final return between two continuations that share a common prefix labels which of two actions was better, and a Bradley-Terry ranking head converts those pairs into calibrated per-step scores. On Llama-3.1-8B-Instruct and Mistral-7B-Instruct backbones, it beats outcome-only GRPO on the average metric in all six benchmark-backbone combinations across GAIA, ToolBench and AgentBoard, and its per-action scores agree with human judgments on 81.5% of pairs.

Tool-using language-model agents often need dozens of interactions before a single success-or-failure label arrives, leaving no way to tell which step went wrong. STEP-VERIFIER turns a success-labeled corpus of logged interactions into action-level comparisons: the difference in average final return between two continuations that share a common prefix labels which of two actions was better, and a Bradley-Terry ranking head converts those pairs into calibrated per-step scores. On Llama-3.1-8B-Instruct and Mistral-7B-Instruct backbones, it beats outcome-only GRPO on the average metric in all six benchmark-backbone combinations across GAIA, ToolBench and AgentBoard, and its per-action scores agree with human judgments on 81.5% of pairs.

A robot that files a hallway defibrillator under fire extinguisher would carry that wrong description in memory. MIT researchers added uncertainty quantification to the existing DAAAM memory system, so the robot compares descriptions of the same object from different viewpoints, estimates how unsure it is about what it saw, and selects extra views to fuse a more accurate caption. Against baseline methods, the system raised object-description accuracy by 5 percent and resolved ambiguity for about 92 percent of objects using only two additional captions per object.

A robot that files a hallway defibrillator under fire extinguisher would carry that wrong description in memory. MIT researchers added uncertainty quantification to the existing DAAAM memory system, so the robot compares descriptions of the same object from different viewpoints, estimates how unsure it is about what it saw, and selects extra views to fuse a more accurate caption. Against baseline methods, the system raised object-description accuracy by 5 percent and resolved ambiguity for about 92 percent of objects using only two additional captions per object.

Microsoft released VibeVoice-ASR-Streaming, which turns the "who said what" transcription that previously required processing a whole audio segment into continuous output as speech arrives. The earlier VibeVoice-ASR handled 60-minute long-form audio in a single pass, producing structured transcriptions with speaker, timestamps and content within a 64K token length. The new model emits text once per chunk and supports customized hotwords and 10 languages.

Microsoft released VibeVoice-ASR-Streaming, which turns the "who said what" transcription that previously required processing a whole audio segment into continuous output as speech arrives. The earlier VibeVoice-ASR handled 60-minute long-form audio in a single pass, producing structured transcriptions with speaker, timestamps and content within a 64K token length. The new model emits text once per chunk and supports customized hotwords and 10 languages.

Large vision-language models process hundreds to thousands of visual tokens per image, and inference cost rises with them. Earlier training-free pruning methods ranked all visual tokens by one importance criterion, overlooking that tokens take on different roles as encoding proceeds. BtsPruner locates the role transition from changes in attention concentration, then picks globally referenced representative tokens from later layers and recursively retrieves local detail tokens from earlier layers, preserving both roles under a fixed budget; on LLaVA-1.5-7B it keeps only 32 of 576 visual tokens (5.6%) and retains 91.6% of the dense model's performance.

Large vision-language models process hundreds to thousands of visual tokens per image, and inference cost rises with them. Earlier training-free pruning methods ranked all visual tokens by one importance criterion, overlooking that tokens take on different roles as encoding proceeds. BtsPruner locates the role transition from changes in attention concentration, then picks globally referenced representative tokens from later layers and recursively retrieves local detail tokens from earlier layers, preserving both roles under a fixed budget; on LLaVA-1.5-7B it keeps only 32 of 576 visual tokens (5.6%) and retains 91.6% of the dense model's performance.

Salesforce AI Research released Interactive Visual Gym (IVG), which moves visual AI assistant evaluation from passively watching pre-recorded video to closed-loop simulation: the assistant watches the user's live feed and speaks guidance, and the simulated user changes its actions in response. The accompanying IVG-Bench covers 315 complex scenarios across three operational environments and evaluates 14 visual assistants; the strongest, Gemini 4 Argon, reaches only 53.4% task success within the time limit, averaged across the three environments. In an early proof of concept, an open-source model trained with the simulator, Gwen 3.5-4B, rose from 33.3% to 42.2% success on entirely new tasks it had never seen during training.

Salesforce AI Research released Interactive Visual Gym (IVG), which moves visual AI assistant evaluation from passively watching pre-recorded video to closed-loop simulation: the assistant watches the user's live feed and speaks guidance, and the simulated user changes its actions in response. The accompanying IVG-Bench covers 315 complex scenarios across three operational environments and evaluates 14 visual assistants; the strongest, Gemini 4 Argon, reaches only 53.4% task success within the time limit, averaged across the three environments. In an early proof of concept, an open-source model trained with the simulator, Gwen 3.5-4B, rose from 33.3% to 42.2% success on entirely new tasks it had never seen during training.

Anthropic now offers an IP allowlist for organizations in the Claude Console, in beta, and requests to covered services from addresses not on the list are refused, which limits where a leaked API key can be used. Admins can add IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization. Previously an organization could not narrow the entry point for requests by source address in the console; this setting shrinks the usable range after a key leak from any address to the listed ones.

Anthropic now offers an IP allowlist for organizations in the Claude Console, in beta, and requests to covered services from addresses not on the list are refused, which limits where a leaked API key can be used. Admins can add IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization. Previously an organization could not narrow the entry point for requests by source address in the console; this setting shrinks the usable range after a key leak from any address to the listed ones.

Writing tool, resource and execution policies into a task agent's prompt does not reliably prevent violations. AgentPolicyForge uses executable environments to build policy benchmarks and iteratively train policy guards, yielding PolicyGym, which spans ten policy families with 698 scenarios for training-data generation and 254 separate held-out scenarios. On PolicyGym, the final guard-based configuration achieves a 40–44% relative reduction in attack success rate and fewer benign rejections than policy prompting alone.

Writing tool, resource and execution policies into a task agent's prompt does not reliably prevent violations. AgentPolicyForge uses executable environments to build policy benchmarks and iteratively train policy guards, yielding PolicyGym, which spans ten policy families with 698 scenarios for training-data generation and 254 separate held-out scenarios. On PolicyGym, the final guard-based configuration achieves a 40–44% relative reduction in attack success rate and fewer benign rejections than policy prompting alone.

Long LLM answers raise inference latency and compute cost, yet prompt-only conciseness is hard to control and early stopping can cut sentences in half. MinimalLM separates when stopping becomes desirable from whether stopping is structurally appropriate: during decoding it adjusts stop-token scores by length and uncertainty, and only lets them take effect at sentence-ending punctuation, closed code fences or balanced brackets. Across three open-source models and 330 questions, generated tokens fell 47.6% with an average quality score of 8.04/10 versus 8.42/10 for unmodified models, and no structurally truncated outputs appeared.

Long LLM answers raise inference latency and compute cost, yet prompt-only conciseness is hard to control and early stopping can cut sentences in half. MinimalLM separates when stopping becomes desirable from whether stopping is structurally appropriate: during decoding it adjusts stop-token scores by length and uncertainty, and only lets them take effect at sentence-ending punctuation, closed code fences or balanced brackets. Across three open-source models and 330 questions, generated tokens fell 47.6% with an average quality score of 8.04/10 versus 8.42/10 for unmodified models, and no structurally truncated outputs appeared.

Mistral released Mistral Large 4, codenamed Le Chonk, and expanded its enterprise platform Mistral Studio with managed agents and apps plus new workflow, enterprise-context, agent-observability and evaluation capabilities. Previously teams had to build the infrastructure an agent runs on; Studio now hosts execution, and developers connect the model to enterprise data and systems through connectors and MCP servers, then set boundaries on the tools, data and actions it can reach. Mistral says an incident-investigation agent built by its engineering team saves hundreds of hours a month, and a product-operations workflow saves about 35 hours a week.

Mistral released Mistral Large 4, codenamed Le Chonk, and expanded its enterprise platform Mistral Studio with managed agents and apps plus new workflow, enterprise-context, agent-observability and evaluation capabilities. Previously teams had to build the infrastructure an agent runs on; Studio now hosts execution, and developers connect the model to enterprise data and systems through connectors and MCP servers, then set boundaries on the tools, data and actions it can reach. Mistral says an incident-investigation agent built by its engineering team saves hundreds of hours a month, and a product-operations workflow saves about 35 hours a week.

Short-horizon model predictive control runs fast but cannot see an opponent beyond its horizon, so it needs a learned, context-dependent cost correction; when passing is possible on either side, both corrections are good and deterministic regression averages them into the opponent. ZipDrift instead trains a one-step drifting generator over cost corrections, tilted by closed-loop cost and learned without differentiating through the optimiser, then generates a few candidates ranked by a learned critic. On a straight-track overtaking benchmark with an eight-dimensional stagewise correction, all 600 test episodes were collision-free, against a 4.0% crash rate for an idealised ZipMPC-style regressor.

Short-horizon model predictive control runs fast but cannot see an opponent beyond its horizon, so it needs a learned, context-dependent cost correction; when passing is possible on either side, both corrections are good and deterministic regression averages them into the opponent. ZipDrift instead trains a one-step drifting generator over cost corrections, tilted by closed-loop cost and learned without differentiating through the optimiser, then generates a few candidates ranked by a learned critic. On a straight-track overtaking benchmark with an eight-dimensional stagewise correction, all 600 test episodes were collision-free, against a 4.0% crash rate for an idealised ZipMPC-style regressor.

Locating RNA inosine sites has long relied on A-to-G base transitions in sequencing, which are hard to separate from DNA mutations and sequencing errors. REX-seq couples EndoV cleavage at inosines with XRN-1 degradation of fragments bearing 5′ phosphate ends, yielding cleavage signatures while preserving gene-expression quantification; in HEK293T cells it identified 590,249 high-confidence inosine sites, 84,645 of them not previously reported. Trained on these sites, InoDREAM reached an AUROC of 0.965 and average precision of 0.964 on a held-out test set, against 0.921 and 0.890 for EditPredict.

Locating RNA inosine sites has long relied on A-to-G base transitions in sequencing, which are hard to separate from DNA mutations and sequencing errors. REX-seq couples EndoV cleavage at inosines with XRN-1 degradation of fragments bearing 5′ phosphate ends, yielding cleavage signatures while preserving gene-expression quantification; in HEK293T cells it identified 590,249 high-confidence inosine sites, 84,645 of them not previously reported. Trained on these sites, InoDREAM reached an AUROC of 0.965 and average precision of 0.964 on a held-out test set, against 0.921 and 0.890 for EditPredict.

Biological foundation models have lacked a common, broad benchmark: relevant tasks span nucleic acids, proteins, genomes, cells, whole organisms and clinical work, and lab validation is too slow to run on every checkpoint. BaseBench, being built by Basecamp Research, separates evaluations, checkpoints and adapters so every checkpoint of every model runs the same evaluations, with data version, checkpoint identity, adapter configuration and scoring steps recorded. It already runs on internal and public biological models, with public submissions planned in the near term.

Biological foundation models have lacked a common, broad benchmark: relevant tasks span nucleic acids, proteins, genomes, cells, whole organisms and clinical work, and lab validation is too slow to run on every checkpoint. BaseBench, being built by Basecamp Research, separates evaluations, checkpoints and adapters so every checkpoint of every model runs the same evaluations, with data version, checkpoint identity, adapter configuration and scoring steps recorded. It already runs on internal and public biological models, with public submissions planned in the near term.

Reflection released Beam, its first open-weight model, together with the safety training methods and evaluation tools behind it. Public safety evaluations had not covered enough ground to test whether safety boundaries hold in adversarial and tool-use settings. Beam combines pretraining data filtering with dedicated safety alignment training, merging a capability teacher and a safety teacher through multi-teacher on-policy distillation. On 1,500 human-written conversations in the Iris safety evaluations, Beam declined 89% of harmful requests and complied with 82% of safety-related harmless requests.

Reflection released Beam, its first open-weight model, together with the safety training methods and evaluation tools behind it. Public safety evaluations had not covered enough ground to test whether safety boundaries hold in adversarial and tool-use settings. Beam combines pretraining data filtering with dedicated safety alignment training, merging a capability teacher and a safety teacher through multi-teacher on-policy distillation. On 1,500 human-written conversations in the Iris safety evaluations, Beam declined 89% of harmful requests and complied with 82% of safety-related harmless requests.

Reinforcement learning that maximizes expected return tends to concentrate probability on a few high-reward outcomes; DARS estimates the local density of outcomes within each on-policy rollout group and inversely scales rewards by that density before standard GRPO updates. On a four-Gaussian navigation task, GRPO covers a single mode while DARS-GRPO recovers mode masses close to 0.25 each, with mode total variation 0.014 versus 0.750 for GRPO. In text-to-image prompt adaptation, DARS-GRPO raises image and prompt diversity across datasets, policy language models and reward settings while keeping semantic alignment.

Reinforcement learning that maximizes expected return tends to concentrate probability on a few high-reward outcomes; DARS estimates the local density of outcomes within each on-policy rollout group and inversely scales rewards by that density before standard GRPO updates. On a four-Gaussian navigation task, GRPO covers a single mode while DARS-GRPO recovers mode masses close to 0.25 each, with mode total variation 0.014 versus 0.750 for GRPO. In text-to-image prompt adaptation, DARS-GRPO raises image and prompt diversity across datasets, policy language models and reward settings while keeping semantic alignment.

Deep Genomics subtracts the whole-transcriptome perturbation response that virtual-cell models usually try to predict, treating it as noise, and trains a sequence model, DeepRNAi, only on the direct effects between an siRNA and its target transcript. Direct effects account for about 1% of all siRNA-RNA interactions, and uncalibrated virtual-cell objectives had treated that signal as a rounding error. On a held-out-transcript split, DeepRNAi performs about the same as on seen transcripts, while a conventional machine-learning approach without a direct-effect focus loses roughly 50% of its above-baseline performance.

Deep Genomics subtracts the whole-transcriptome perturbation response that virtual-cell models usually try to predict, treating it as noise, and trains a sequence model, DeepRNAi, only on the direct effects between an siRNA and its target transcript. Direct effects account for about 1% of all siRNA-RNA interactions, and uncalibrated virtual-cell objectives had treated that signal as a rounding error. On a held-out-transcript split, DeepRNAi performs about the same as on seen transcripts, while a conventional machine-learning approach without a direct-effect focus loses roughly 50% of its above-baseline performance.

Gemini Omni 1.2 Flash is now generally available through the Gemini API and Gemini for business, extending native video generation from 10 seconds in Omni 1.1 to 15 seconds and allowing durations from 3 to 15 seconds in one-second increments. It adds multi-turn conversational editing, so a clip can be refined turn by turn, swapping a product or background or changing lighting and aspect ratio without regenerating the whole clip; 720p video costs $0.07 per second, 30% below Omni 1.1. In the official example, given the same static studio poster and five-layer animation prompt, 1.2 follows the reference image more closely than 1.1.

Gemini Omni 1.2 Flash is now generally available through the Gemini API and Gemini for business, extending native video generation from 10 seconds in Omni 1.1 to 15 seconds and allowing durations from 3 to 15 seconds in one-second increments. It adds multi-turn conversational editing, so a clip can be refined turn by turn, swapping a product or background or changing lighting and aspect ratio without regenerating the whole clip; 720p video costs $0.07 per second, 30% below Omni 1.1. In the official example, given the same static studio poster and five-layer animation prompt, 1.2 follows the reference image more closely than 1.1.

LLM judges are a common substitute for costly human evaluation, and aggregating several judges is the standard remedy for any single judge's bias, but judge pools mix very different reliabilities while the human-labeled sets used to validate them hold a few hundred items or fewer. AGORA places every item on one latent quality scale and describes each judge by a sensitivity parameter and ordered rating thresholds, so judge reliability is learned jointly from all items without human labels. Across three benchmarks, AGORA matches or exceeds Dawid-Skene on large evaluation sets, and its advantage widens as the set shrinks toward a few hundred items.

LLM judges are a common substitute for costly human evaluation, and aggregating several judges is the standard remedy for any single judge's bias, but judge pools mix very different reliabilities while the human-labeled sets used to validate them hold a few hundred items or fewer. AGORA places every item on one latent quality scale and describes each judge by a sensitivity parameter and ordered rating thresholds, so judge reliability is learned jointly from all items without human labels. Across three benchmarks, AGORA matches or exceeds Dawid-Skene on large evaluation sets, and its advantage widens as the set shrinks toward a few hundred items.

Most memory systems for large language model agents only store and retrieve past information, while proactive agents also need to hold future intentions and realize them when the right moment arrives. Pro-Memory separates keeping a future intention from deciding when and how to realize it: pending objectives stay linked to their supporting evidence, new observations are checked against the original goal and activation conditions, and contributions are coordinated with the agent's ongoing actions. On two benchmarks covering delayed-intention tasks and open-ended proactive assistance, proactive task performance improves consistently over representative memory approaches across multiple LLM backbones while unnecessary interventions are limited.

Most memory systems for large language model agents only store and retrieve past information, while proactive agents also need to hold future intentions and realize them when the right moment arrives. Pro-Memory separates keeping a future intention from deciding when and how to realize it: pending objectives stay linked to their supporting evidence, new observations are checked against the original goal and activation conditions, and contributions are coordinated with the agent's ongoing actions. On two benchmarks covering delayed-intention tasks and open-ended proactive assistance, proactive task performance improves consistently over representative memory approaches across multiple LLM backbones while unnecessary interventions are limited.

Decentralized federated learning has no central server, and link noise accumulates as models are relayed hop by hop, becoming more pronounced as the learning rate decays; Byzantine nodes can also mimic honest model averages to win loss-based selection. FedSharp keeps each predecessor's model copies at every node, screens out single-parameter deviations using coordinate-wise median deviations, averages the lowest-loss candidates, and adds noise-adaptive gradient-norm regularization with proximal training. On a 100-node ring with 33 Byzantine nodes, MNIST accuracy is 93.5%–94.9%, versus 29.4%–38.1% for a single-model-selection baseline.

Decentralized federated learning has no central server, and link noise accumulates as models are relayed hop by hop, becoming more pronounced as the learning rate decays; Byzantine nodes can also mimic honest model averages to win loss-based selection. FedSharp keeps each predecessor's model copies at every node, screens out single-parameter deviations using coordinate-wise median deviations, averages the lowest-loss candidates, and adds noise-adaptive gradient-norm regularization with proximal training. On a 100-node ring with 33 Byzantine nodes, MNIST accuracy is 93.5%–94.9%, versus 29.4%–38.1% for a single-model-selection baseline.

Deep Genomics' DeepRNAi combines the local siRNA hybridization site with long-range context from the genomic region around the transcript, supplied by biological foundation models BigRNA, Enigma and REPRESS over up to one million base pairs of sequence. On seed-mediated knockdown prediction, DeepRNAi without the foundation models performs in aggregate about as well as a random forest with hand-built transcript features; with the foundation models, performance nearly doubles. In two drug-discovery cases the model predicted near-zero activity for siRNAs targeting the distal 3' UTR, matching Alnylam's published PLG experimental data.

Deep Genomics' DeepRNAi combines the local siRNA hybridization site with long-range context from the genomic region around the transcript, supplied by biological foundation models BigRNA, Enigma and REPRESS over up to one million base pairs of sequence. On seed-mediated knockdown prediction, DeepRNAi without the foundation models performs in aggregate about as well as a random forest with hand-built transcript features; with the foundation models, performance nearly doubles. In two drug-discovery cases the model predicted near-zero activity for siRNAs targeting the distal 3' UTR, matching Alnylam's published PLG experimental data.

Vercel Labs has released agent-browser, a native Rust command-line tool that lets an AI agent drive a browser one shell command at a time. Previously an agent had to work from the whole page; now snapshot returns the page as compact text with a ref such as @e1 for each element, and the agent acts on a ref, reads the page again to check the result, and repeats. The browser stays open between commands, so each step is one short shell call, and a click fails rather than reporting success when the target is hidden, disabled, outside the viewport, or covered.

Vercel Labs has released agent-browser, a native Rust command-line tool that lets an AI agent drive a browser one shell command at a time. Previously an agent had to work from the whole page; now snapshot returns the page as compact text with a ref such as @e1 for each element, and the agent acts on a ref, reads the page again to check the result, and repeats. The browser stays open between commands, so each step is one short shell call, and a click fails rather than reporting success when the target is hidden, disabled, outside the viewport, or covered.

Atlas was an autoregressive world model that generated frames one at a time, and users waited tens of seconds or even minutes for a video, so they planned camera paths cautiously. The new research preview Frameboy serves a highly optimized Atlas variant from cloud GPUs, streaming frames to the browser at up to 15 fps so the view updates continuously as the camera moves. The same real-time capability is integrated into the web app Viewfinder, where users build their own scenes frame by frame and then explore them live through a Frameboy overlay.

Atlas was an autoregressive world model that generated frames one at a time, and users waited tens of seconds or even minutes for a video, so they planned camera paths cautiously. The new research preview Frameboy serves a highly optimized Atlas variant from cloud GPUs, streaming frames to the browser at up to 15 fps so the view updates continuously as the camera moves. The same real-time capability is integrated into the web app Viewfinder, where users build their own scenes frame by frame and then explore them live through a Frameboy overlay.

GPUMAP splits UMAP's dimensionality-reduction pipeline into three interleavable operations, insert, update and embed, and runs them on a commodity GPU so that intermediate projections appear while computation continues, letting million-point datasets be inspected before the full run ends. On the 5-million-point HIGGS dataset, sequential RAPIDS UMAP produced its first projection only at 50.4 seconds, whereas GPUMAP produced an inspectable intermediate projection at 8.42 seconds and kept refining it; the cost is a total runtime about 1.6 times that of sequential execution.

GPUMAP splits UMAP's dimensionality-reduction pipeline into three interleavable operations, insert, update and embed, and runs them on a commodity GPU so that intermediate projections appear while computation continues, letting million-point datasets be inspected before the full run ends. On the 5-million-point HIGGS dataset, sequential RAPIDS UMAP produced its first projection only at 50.4 seconds, whereas GPUMAP produced an inspectable intermediate projection at 8.42 seconds and kept refining it; the cost is a total runtime about 1.6 times that of sequential execution.

Google Cloud has made Gemini Omni 1.2 Flash generally available for businesses producing product videos, virtual try-ons and ads at volume. Against the previous Gemini Omni 1.1 Flash, 720p video drops to $0.07 per second, 30% less, and single clips extend from 10 to 15 seconds, with closer adherence to supplied image, video and audio references. Business users can create in Google Vids, the Gemini for business entry is marked as coming soon, developers can call it through the API, and drafts can be made at 360p or 720p before upscaling to 1080p or 4K.

Google Cloud has made Gemini Omni 1.2 Flash generally available for businesses producing product videos, virtual try-ons and ads at volume. Against the previous Gemini Omni 1.1 Flash, 720p video drops to $0.07 per second, 30% less, and single clips extend from 10 to 15 seconds, with closer adherence to supplied image, video and audio references. Business users can create in Google Vids, the Gemini for business entry is marked as coming soon, developers can call it through the API, and drafts can be made at 360p or 720p before upscaling to 1080p or 4K.

NVIDIA announced NVIDIA Sentry at GTC Berlin, a security framework that runs on BlueField-4 DPUs and places monitoring and policy enforcement for AI agents' access to models, tools and data in a separate hardware trust domain beyond host software. Earlier agent security relied on software runtime boundaries, so an agent that escaped its sandbox could alter or disable its own restrictions; Sentry inspects agent-to-model traffic through the DOCA Tower transparent proxy and reads runtime state outside the host with DOCA Argus, containing agents at millisecond scale. The framework is planned for availability early next year, and developers can begin with DOCA Argus, DOCA Flow and other libraries and microservices to experiment with out-of-band observability and infrastructure enforcement.

NVIDIA announced NVIDIA Sentry at GTC Berlin, a security framework that runs on BlueField-4 DPUs and places monitoring and policy enforcement for AI agents' access to models, tools and data in a separate hardware trust domain beyond host software. Earlier agent security relied on software runtime boundaries, so an agent that escaped its sandbox could alter or disable its own restrictions; Sentry inspects agent-to-model traffic through the DOCA Tower transparent proxy and reads runtime state outside the host with DOCA Argus, containing agents at millisecond scale. The framework is planned for availability early next year, and developers can begin with DOCA Argus, DOCA Flow and other libraries and microservices to experiment with out-of-band observability and infrastructure enforcement.

UCB and Cradle announced a strategic collaboration that puts Cradle's AI-driven protein design platform inside selected UCB antibody discovery projects, spanning hit generation, lead optimization and preclinical readiness. UCB's digital antibody engineering program had relied on its own antibody discovery platform combined with computational science and a network of technology and research partners. Under the collaboration, experimental results feed back into program-specific models in a compounding "lab-in-the-loop" cycle used to optimize candidates across the full set of required properties.

UCB and Cradle announced a strategic collaboration that puts Cradle's AI-driven protein design platform inside selected UCB antibody discovery projects, spanning hit generation, lead optimization and preclinical readiness. UCB's digital antibody engineering program had relied on its own antibody discovery platform combined with computational science and a network of technology and research partners. Under the collaboration, experimental results feed back into program-specific models in a compounding "lab-in-the-loop" cycle used to optimize candidates across the full set of required properties.

Albert released Valence, a chemistry-native retrieval model for picking the passage that actually answers a formulation-science question. General-purpose retrieval models sort broadly across fields but struggle to separate same-field passages that share vocabulary and classifications yet answer different questions; on FormBench's hard test, general models from OpenAI and Voyage found the correct passage in their top 100 results 2% to 4.5% of the time, while Valence did so 81.8% of the time. On a stricter measure, Valence scored 0.41 against 0.009 for the leading general-purpose model.

Albert released Valence, a chemistry-native retrieval model for picking the passage that actually answers a formulation-science question. General-purpose retrieval models sort broadly across fields but struggle to separate same-field passages that share vocabulary and classifications yet answer different questions; on FormBench's hard test, general models from OpenAI and Voyage found the correct passage in their top 100 results 2% to 4.5% of the time, while Valence did so 81.8% of the time. On a stricter measure, Valence scored 0.41 against 0.009 for the leading general-purpose model.

Anthropic added a security center, inference hooks, centralized management and new access controls to Claude Enterprise, so security teams can see how their organization's configuration differs from Anthropic's recommended setup, send prompts and tool calls to existing data security tools for a check before they run, and manage identity and access from one place. The security center and centralized management remain in beta while inference hooks are generally available; the security center is guidance based on recommendations, not a certification or a guarantee of compliance.

Anthropic added a security center, inference hooks, centralized management and new access controls to Claude Enterprise, so security teams can see how their organization's configuration differs from Anthropic's recommended setup, send prompts and tool calls to existing data security tools for a check before they run, and manage identity and access from one place. The security center and centralized management remain in beta while inference hooks are generally available; the security center is guidance based on recommendations, not a certification or a guarantee of compliance.

Chai Discovery announced a collaboration with Roche to use Chai's AI models and platform in a number of drug discovery programs across Roche's therapeutic areas. Chai builds computer-aided design tools for protein-based therapeutics, and its latest model, Chai-3, is described by the company as achieving high target success rates and binding affinities, including against challenging targets. The deal is the latest in a series of 2026 collaborations Chai has announced with leading pharmaceutical companies; the announcement gives no program counts, specific targets or timelines.

Chai Discovery announced a collaboration with Roche to use Chai's AI models and platform in a number of drug discovery programs across Roche's therapeutic areas. Chai builds computer-aided design tools for protein-based therapeutics, and its latest model, Chai-3, is described by the company as achieving high target success rates and binding affinities, including against challenging targets. The deal is the latest in a series of 2026 collaborations Chai has announced with leading pharmaceutical companies; the announcement gives no program counts, specific targets or timelines.

Molecular data in lung adenocarcinoma are high-dimensional relative to the number of observed events, so concatenating them into a survival model invites overfitting and unstable feature selection. CAISL writes predicted risk as a clinical Cox risk plus a sparse molecular correction, and penalizes how much that correction varies across resampled training cohorts. On 470 TCGA-LUAD patients with 119 observed deaths, mean C-index across five independent cross-validation runs was 0.7246 versus 0.6970 for clinical Cox, a paired gain of 0.0276 (95% CI 0.005–0.050); this is internal validation and still needs an independent cohort.

Molecular data in lung adenocarcinoma are high-dimensional relative to the number of observed events, so concatenating them into a survival model invites overfitting and unstable feature selection. CAISL writes predicted risk as a clinical Cox risk plus a sparse molecular correction, and penalizes how much that correction varies across resampled training cohorts. On 470 TCGA-LUAD patients with 119 observed deaths, mean C-index across five independent cross-validation runs was 0.7246 versus 0.6970 for clinical Cox, a paired gain of 0.0276 (95% CI 0.005–0.050); this is internal validation and still needs an independent cohort.

Continuous-control reinforcement learning struggles with exploration and long-horizon credit assignment, and the usual fix, action noise or entropy regularization, explores without direction. GANGSTER uses a conditional generative adversarial network to propose candidate subgoals from the current state, scores them with structured losses for desirability, reachability and diversity, and feeds the chosen subgoal back only as intrinsic reward shaping, leaving the core TD3 and SAC update rules untouched. On MuJoCo's Ant-v5, GANGSTER with TD3 cuts time-to-solve by about 31% and raises final performance by 53% over baseline TD3; on HalfCheetah-v5 with SAC it raises final performance by 10% and lifts post-solution consistency above 99%.

Continuous-control reinforcement learning struggles with exploration and long-horizon credit assignment, and the usual fix, action noise or entropy regularization, explores without direction. GANGSTER uses a conditional generative adversarial network to propose candidate subgoals from the current state, scores them with structured losses for desirability, reachability and diversity, and feeds the chosen subgoal back only as intrinsic reward shaping, leaving the core TD3 and SAC update rules untouched. On MuJoCo's Ant-v5, GANGSTER with TD3 cuts time-to-solve by about 31% and raises final performance by 53% over baseline TD3; on HalfCheetah-v5 with SAC it raises final performance by 10% and lifts post-solution consistency above 99%.

Baseten introduced the Baseten Environment Network, handing the construction of reinforcement-learning environments for post-training to a set of vetted partner companies. Previously Baseten's own applied research team built these environments by hand for post-training engagements such as Harvey, Lovable and Gamma, work the company describes as manually intensive and dependent on researchers' judgment. Partners work directly with customers and build environments from the customer's real data, while Baseten handles training and inference deployment; Baseten also plans to open-source a set of frontier-quality environments along with how it audits and trains on them.

Baseten introduced the Baseten Environment Network, handing the construction of reinforcement-learning environments for post-training to a set of vetted partner companies. Previously Baseten's own applied research team built these environments by hand for post-training engagements such as Harvey, Lovable and Gamma, work the company describes as manually intensive and dependent on researchers' judgment. Partners work directly with customers and build environments from the customer's real data, while Baseten handles training and inference deployment; Baseten also plans to open-source a set of frontier-quality environments along with how it audits and trains on them.

DeepSeek open-sourced DeepInfer, the inference engine behind its own service, along with its kernels; SGLang then brought the Mega kernel series, DSpark speculative decoding, SWA bounded replay and both Engram placements into its mainline and compared against DeepInfer on the same hardware using DeepInfer's own benchmarks. On 4x GB300 and 4x B300, SGLang reaches up to 1.04x DeepInfer's prefill throughput, within 1% at the remaining points; on 64x GB300 its decode step is 1.1% to 7.4% shorter from 16 to 256 requests per GPU and on par at 8. SGLang also ships the KV cache system, prefill/decode disaggregation and cross-replica load balancing that DeepInfer's release notes list as not included.

DeepSeek open-sourced DeepInfer, the inference engine behind its own service, along with its kernels; SGLang then brought the Mega kernel series, DSpark speculative decoding, SWA bounded replay and both Engram placements into its mainline and compared against DeepInfer on the same hardware using DeepInfer's own benchmarks. On 4x GB300 and 4x B300, SGLang reaches up to 1.04x DeepInfer's prefill throughput, within 1% at the remaining points; on 64x GB300 its decode step is 1.1% to 7.4% shorter from 16 to 256 requests per GPU and on par at 8. SGLang also ships the KV cache system, prefill/decode disaggregation and cross-replica load balancing that DeepInfer's release notes list as not included.

Deep Agents adds async subagents: an orchestrator stays in the foreground talking to the user and organizing context, while subagents run long tasks as background workers and report back on completion. Previously, in a standard ReAct loop, each long tool call blocked the loop and every other task queued behind it; now subagents no longer hold the orchestrator's loop, the user can reach it whenever its own loop isn't busy, and the work an agent can take on grows with the number of background workers rather than the length of one loop.

Deep Agents adds async subagents: an orchestrator stays in the foreground talking to the user and organizing context, while subagents run long tasks as background workers and report back on completion. Previously, in a standard ReAct loop, each long tool call blocked the loop and every other task queued behind it; now subagents no longer hold the orchestrator's loop, the user can reach it whenever its own loop isn't busy, and the work an agent can take on grows with the number of background workers rather than the length of one loop.

NVIDIA has introduced Vera, a CPU aimed at AI agents, seeking to meet both requirements in one chip: fast execution of agent steps and high concurrent agent counts. Previously AI factories had to choose between the fastest x86 processor and the most energy-efficient one to balance speed and capacity. In a 100-megawatt data center comparison, NVIDIA reports Vera supports 1.6x more agents per watt than the fastest x86 CPU and delivers 2x faster agent speed than the most energy-efficient x86 CPU; early adopters including Perplexity and DeepInfra report speedups in production-style tests, though Perplexity's deployment is still planned.

NVIDIA has introduced Vera, a CPU aimed at AI agents, seeking to meet both requirements in one chip: fast execution of agent steps and high concurrent agent counts. Previously AI factories had to choose between the fastest x86 processor and the most energy-efficient one to balance speed and capacity. In a 100-megawatt data center comparison, NVIDIA reports Vera supports 1.6x more agents per watt than the fastest x86 CPU and delivers 2x faster agent speed than the most energy-efficient x86 CPU; early adopters including Perplexity and DeepInfra report speedups in production-style tests, though Perplexity's deployment is still planned.

Anthropic added a beta IP allowlist to the Claude Console, so only addresses on an organization's list can reach covered services, which limits where a leaked API key can still be used. Previously organizations had no console-level way to narrow requests by source address. Admins can enter IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization.

Anthropic added a beta IP allowlist to the Claude Console, so only addresses on an organization's list can reach covered services, which limits where a leaked API key can still be used. Previously organizations had no console-level way to narrow requests by source address. Admins can enter IPv4 and IPv6 addresses or CIDR ranges, up to 1,000 per organization.

Vercel released Vercel Schedules, which lets agents and applications create, change or cancel one-time and recurring schedules through code while the application is still running, without redeploying. Vercel Cron Jobs previously kept timing in vercel.json or Build Output API configuration and synced on deploy, while dynamic scheduling required a separate database and a polling job to find due work. Schedules records when to trigger a task and where to send it, and Vercel delivers the saved request through Queues when it is due, with application code checking permissions and doing the work; the product is in public beta.

Vercel released Vercel Schedules, which lets agents and applications create, change or cancel one-time and recurring schedules through code while the application is still running, without redeploying. Vercel Cron Jobs previously kept timing in vercel.json or Build Output API configuration and synced on deploy, while dynamic scheduling required a separate database and a polling job to find due work. Schedules records when to trigger a task and where to send it, and Vercel delivers the saved request through Queues when it is due, with application code checking permissions and doing the work; the product is in public beta.

Claude now has a Personal Finance feature: users sign in to their own bank through Plaid, grant read-only access, and ask about balances and transactions in their own words. Previously, working out where money went meant opening each banking app and doing the math; the feature totals spending across connected accounts by category, and in the example the month's spending is $4,309.37, with four payments still due before payday and $1,393.22 left. It rolls out in beta this week on the Pro, Max and Ultra plans in the US, helps users understand their money, and is not designed to give investment advice.

Claude now has a Personal Finance feature: users sign in to their own bank through Plaid, grant read-only access, and ask about balances and transactions in their own words. Previously, working out where money went meant opening each banking app and doing the math; the feature totals spending across connected accounts by category, and in the example the month's spending is $4,309.37, with four payments still due before payday and $1,393.22 left. It rolls out in beta this week on the Pro, Max and Ultra plans in the US, helps users understand their money, and is not designed to give investment advice.

Spiking networks that fuse audio and video have lacked a way to regulate information flow inside cross-modal attention according to the input. GSCANet adds a gate to a dual-stream Spikingformer backbone, placing it on the attention output just before spike firing and driving it with concatenated query and value features. On CREMA-D and RAVDESS this configuration beats the no-gate baseline by 2.15 and 1.05 percentage points; both gate position and driving signal change the outcome.

Spiking networks that fuse audio and video have lacked a way to regulate information flow inside cross-modal attention according to the input. GSCANet adds a gate to a dual-stream Spikingformer backbone, placing it on the attention output just before spike firing and driving it with concatenated query and value features. On CREMA-D and RAVDESS this configuration beats the no-gate baseline by 2.15 and 1.05 percentage points; both gate position and driving signal change the outcome.

Mastra platform and Studio now record two latency metrics for every agent and model call: TTFC, the time a user waits before the first chunk arrives, and TTFT, the time the model takes to start responding. Previously durationMs showed only how long a run took, not how long a user waited for the first response or whether that wait came from the model or from the application's own code. Both metrics are recorded on every span, from the model call up to the agent run, and can be filtered and sorted as fields in trace and span queries on Mastra platform and on self-hosted projects whose storage uses DuckDB, PostgreSQL, or ClickHouse.

Mastra platform and Studio now record two latency metrics for every agent and model call: TTFC, the time a user waits before the first chunk arrives, and TTFT, the time the model takes to start responding. Previously durationMs showed only how long a run took, not how long a user waited for the first response or whether that wait came from the model or from the application's own code. Both metrics are recorded on every span, from the model call up to the agent run, and can be filtered and sorted as fields in trace and span queries on Mastra platform and on self-hosted projects whose storage uses DuckDB, PostgreSQL, or ClickHouse.