AI and science frontiers · 2026-10-23

Distilling a teacher on pure Gaussian noise still hands the student the teacher's memorized training records and planted backdoors whenever both start from the same public initialization. On MNIST, distilling only never-trained auxiliary outputs yields a membership-inference AUC of 0.86 and a 93% backdoor attack-success lift, while an independently initialized student stays near chance. In a fine-tuning threat model, an attacker with the public base but no data recovers the memorized labels of 93% of private canaries from noise queries alone.

Distilling a teacher on pure Gaussian noise still hands the student the teacher's memorized training records and planted backdoors whenever both start from the same public initialization. On MNIST, distilling only never-trained auxiliary outputs yields a membership-inference AUC of 0.86 and a 93% backdoor attack-success lift, while an independently initialized student stays near chance. In a fine-tuning threat model, an attacker with the public base but no data recovers the memorized labels of 93% of private canaries from noise queries alone.

Surface-code decoders have processed the full $d^2 \times R$ syndrome array regardless of how many errors occurred, yet at roughly 0.1% physical error rate fewer than 5% of entries fire. The Sparse Mamba Decoder (SMD) processes only the k active detection events, each described by a 13-dimensional feature, and runs a Mamba state-space backbone in O(k) time. Under SI1000 noise at $p=1.5\times10^{-3}$, a three-checkpoint ensemble lowers the d=7 logical error rate by 16% relative to MWPM, and the single model reaches ratios of 0.65 at d=3 and 0.51 at d=5.

Surface-code decoders have processed the full $d^2 \times R$ syndrome array regardless of how many errors occurred, yet at roughly 0.1% physical error rate fewer than 5% of entries fire. The Sparse Mamba Decoder (SMD) processes only the k active detection events, each described by a 13-dimensional feature, and runs a Mamba state-space backbone in O(k) time. Under SI1000 noise at $p=1.5\times10^{-3}$, a three-checkpoint ensemble lowers the d=7 logical error rate by 16% relative to MWPM, and the single model reaches ratios of 0.65 at d=3 and 0.51 at d=5.

Deep visual proteomics (DVP) lacked an open-source end-to-end workflow, which limited adoption and systematic benchmarking. openDVP links image processing, cell phenotyping, laser microdissection and mass-spectrometry proteomics into a modular pipeline, with a fast-track route that needs no image-analysis expertise. Across archival placenta, early lung cancer and relapsed breast cancer tissue, neighborhood-guided sampling collected tissue about 100 times faster than single-cell cutting and raised protein identifications by an average of 26%.

Deep visual proteomics (DVP) lacked an open-source end-to-end workflow, which limited adoption and systematic benchmarking. openDVP links image processing, cell phenotyping, laser microdissection and mass-spectrometry proteomics into a modular pipeline, with a fast-track route that needs no image-analysis expertise. Across archival placenta, early lung cancer and relapsed breast cancer tissue, neighborhood-guided sampling collected tissue about 100 times faster than single-cell cutting and raised protein identifications by an average of 26%.

Ai2 released OlmoEarth v1.3, a satellite embedding model that compresses a year of radar and multispectral time series into one 128-dimensional vector per 10-meter pixel, with precomputed annual embeddings for the whole planet shared openly. Until now, using OlmoEarth mainly meant fine-tuning a model per task in OlmoEarth Studio, which requires GPU access and storing the imagery. Against AlphaEarth Foundations and Tessera v2 across 27 evaluations on six continents, the 128-dimensional embeddings scored highest on 20 and the 64-dimensional version led on 18.

Ai2 released OlmoEarth v1.3, a satellite embedding model that compresses a year of radar and multispectral time series into one 128-dimensional vector per 10-meter pixel, with precomputed annual embeddings for the whole planet shared openly. Until now, using OlmoEarth mainly meant fine-tuning a model per task in OlmoEarth Studio, which requires GPU access and storing the imagery. Against AlphaEarth Foundations and Tessera v2 across 27 evaluations on six continents, the 128-dimensional embeddings scored highest on 20 and the 64-dimensional version led on 18.

Which lines to phenotype has largely been decided by habit or by picking those most related to the targets. TSOpt instead designs training sets by exact rank-one updates of the prediction error variance; on 328 IRRI elite rice lines it raised predictive ability over random sampling by 0.097, with a design taking about 0.1 s and running roughly 900 times faster than STPGA. Twenty-five designed lines predicted grain yield as well as about 103 random lines, days to flowering as 76 and plant height as 167; its sparse-testing allocations also beat balanced random allocation when a quarter of line-by-season cells were tested.

Which lines to phenotype has largely been decided by habit or by picking those most related to the targets. TSOpt instead designs training sets by exact rank-one updates of the prediction error variance; on 328 IRRI elite rice lines it raised predictive ability over random sampling by 0.097, with a design taking about 0.1 s and running roughly 900 times faster than STPGA. Twenty-five designed lines predicted grain yield as well as about 103 random lines, days to flowering as 76 and plant height as 167; its sparse-testing allocations also beat balanced random allocation when a quarter of line-by-season cells were tested.

The published split of the public road-defect dataset BharatPotHole overlaps between training and test: 38.8% of test images have a near-duplicate in training and 99.4% come from videos also used for training. After a deduplicated, video-disjoint re-split, the same YOLO26n detector reaches AP50 0.619 on the published test set and 0.467 on its near-duplicate-free subset, a paired difference of 0.152 (95% CI 0.090–0.234). In a controlled contamination test, replacing unseen test images with copies of training frames raised a video-disjoint model's AP50 from 0.133 to 0.494, while copies of unseen frames left it near 0.15.

The published split of the public road-defect dataset BharatPotHole overlaps between training and test: 38.8% of test images have a near-duplicate in training and 99.4% come from videos also used for training. After a deduplicated, video-disjoint re-split, the same YOLO26n detector reaches AP50 0.619 on the published test set and 0.467 on its near-duplicate-free subset, a paired difference of 0.152 (95% CI 0.090–0.234). In a controlled contamination test, replacing unseen test images with copies of training frames raised a video-disjoint model's AP50 from 0.133 to 0.494, while copies of unseen frames left it near 0.15.

SCCGBench replaces text-similarity scoring for service-call code with virtual-execution hits, and on 612 test samples the strongest model lands the correct host, endpoint, and HTTP method together only 3.76% of the time.

SCCGBench replaces text-similarity scoring for service-call code with virtual-execution hits, and on 612 test samples the strongest model lands the correct host, endpoint, and HTTP method together only 3.76% of the time.

Real-time ECG monitoring must use causal filters, whose nonlinear phase misaligns the frequency components of the cardiac waveform. Across 36,583 twelve-lead records from three independent clinical databases, changing phase alone while holding spectral magnitude fixed cut ResNet-18 macro-AUROC by 0.1244 on PTB-XL, with conduction-block detection falling from above 0.91 to about 0.62. A training-time regularizer that randomly applies all-pass phase perturbations recovered 86.9% to 90.9% of that loss across cohorts without changing the architecture.

Real-time ECG monitoring must use causal filters, whose nonlinear phase misaligns the frequency components of the cardiac waveform. Across 36,583 twelve-lead records from three independent clinical databases, changing phase alone while holding spectral magnitude fixed cut ResNet-18 macro-AUROC by 0.1244 on PTB-XL, with conduction-block detection falling from above 0.91 to about 0.62. A training-time regularizer that randomly applies all-pass phase perturbations recovered 86.9% to 90.9% of that loss across cohorts without changing the architecture.

Large-area tactile sensing has meant tiling many sensing units and routing a wire to each, so scaling area or resolution drives up wiring and cost. A serpentine electromagnetic surface-wave transmission line on flexible polyimide is fed swept signals from both ends, and the dual-port reflection spectra encode contact position, which a multi-branch multilayer perceptron decodes into coordinates. Two prototypes an order of magnitude apart in area, 5×5 and 16×18 cm, both exceed a super-resolution factor of 10,000, with two-dimensional localization RMSE of 0.368 mm for the small skin and 1.209 mm for the large one.

Large-area tactile sensing has meant tiling many sensing units and routing a wire to each, so scaling area or resolution drives up wiring and cost. A serpentine electromagnetic surface-wave transmission line on flexible polyimide is fed swept signals from both ends, and the dual-port reflection spectra encode contact position, which a multi-branch multilayer perceptron decodes into coordinates. Two prototypes an order of magnitude apart in area, 5×5 and 16×18 cm, both exceed a super-resolution factor of 10,000, with two-dimensional localization RMSE of 0.368 mm for the small skin and 1.209 mm for the large one.

MiniMax open-sourced MiniMax-M3.1-Flash, and vLLM supports it on day one on NVIDIA Blackwell, NVIDIA Rubin and AMD MI350-series GPUs. The model is quantization-aware trained for an NVFP4 KV cache and stores its routed MoE weights in NVFP4, shrinking the checkpoint from 451.5 GB to 250.1 GB and the KV cache from 61,440 to 34,560 bytes per token. On four NVIDIA B300 GPUs with DEP4 deployment and 256K-token sequences, M3.1 fits 76 concurrent sequences where M3 with an FP8 KV cache fits 28.

MiniMax open-sourced MiniMax-M3.1-Flash, and vLLM supports it on day one on NVIDIA Blackwell, NVIDIA Rubin and AMD MI350-series GPUs. The model is quantization-aware trained for an NVFP4 KV cache and stores its routed MoE weights in NVFP4, shrinking the checkpoint from 451.5 GB to 250.1 GB and the KV cache from 61,440 to 34,560 bytes per token. On four NVIDIA B300 GPUs with DEP4 deployment and 256K-token sequences, M3.1 fits 76 concurrent sequences where M3 with an FP8 KV cache fits 28.

Frontier AI agent time horizons are usually computed from average success: the task length an agent completes with 50% probability, measured in human expert time. This analysis instead fits a reliability horizon from METR's public run-level records, using the probability that all four attempts on the same task succeed. Across 50 model-benchmark combinations the reliability horizon is always shorter than the capability horizon, with a geometric-mean ratio of about 2.4, and the gap did not narrow significantly between 2023 and early 2026.

Frontier AI agent time horizons are usually computed from average success: the task length an agent completes with 50% probability, measured in human expert time. This analysis instead fits a reliability horizon from METR's public run-level records, using the probability that all four attempts on the same task succeed. Across 50 model-benchmark combinations the reliability horizon is always shorter than the capability horizon, with a geometric-mean ratio of about 2.4, and the gap did not narrow significantly between 2023 and early 2026.

With only 20 locally labeled cases, an agentic framework for echocardiographic video performs left ventricular segmentation and ejection fraction estimation. It chains a low-rank-adapted video segmentation foundation model, a mask-guided long-axis regression module and a natural-language interface; on a 98-case Mayo Clinic cohort it lowers median ejection fraction error from 5.47% to 4.33% and raises the share of cases within ±10 percentage points from 80.6% to 83.3%. On the public EchoNet-Dynamic set, segmentation Dice is 0.934 and median ejection fraction error is 4.24%.

With only 20 locally labeled cases, an agentic framework for echocardiographic video performs left ventricular segmentation and ejection fraction estimation. It chains a low-rank-adapted video segmentation foundation model, a mask-guided long-axis regression module and a natural-language interface; on a 98-case Mayo Clinic cohort it lowers median ejection fraction error from 5.47% to 4.33% and raises the share of cases within ±10 percentage points from 80.6% to 83.3%. On the public EchoNet-Dynamic set, segmentation Dice is 0.934 and median ejection fraction error is 4.24%.

Fireworks released ARCv3, a compressor that turns the BF16 weight updates a trainer sends to rollout machines into compact deltas. Cross-region reinforcement learning previously meant shipping full checkpoints or relying on compressors such as ARCv2; ARCv3 encodes the bit-level structure of weight changes, sending mostly mantissa bits and packing rare exponent and sign changes separately. Across 1,000 weight-update deltas from production RL workloads, ARCv3 produced an average payload of about 0.19% of the original BF16 weight size, down from 0.36% for ARCv2, with bit-for-bit reconstruction of the trainer's weights.

Fireworks released ARCv3, a compressor that turns the BF16 weight updates a trainer sends to rollout machines into compact deltas. Cross-region reinforcement learning previously meant shipping full checkpoints or relying on compressors such as ARCv2; ARCv3 encodes the bit-level structure of weight changes, sending mostly mantissa bits and packing rare exponent and sign changes separately. Across 1,000 weight-update deltas from production RL workloads, ARCv3 produced an average payload of about 0.19% of the original BF16 weight size, down from 0.36% for ARCv2, with bit-for-bit reconstruction of the trainer's weights.

TorchTitan's annual update adds PyTorch-native training support for open-weight models including Muse Spark Open, Qwen3.5, Kimi K3 and DeepSeek-V4, and introduces TitanRL, which keeps one model from pretraining through post-training to inside vLLM so training and inference no longer run two copies that can quietly disagree. In supported on-policy configurations the trainer and generator produce bitwise identical log probabilities, leaving exactly zero gap at every step. The update also reports DeepSeek V3 671B throughput on the MLPerf training benchmark and infrastructure work around FlexShard and torch_checkpointing.

TorchTitan's annual update adds PyTorch-native training support for open-weight models including Muse Spark Open, Qwen3.5, Kimi K3 and DeepSeek-V4, and introduces TitanRL, which keeps one model from pretraining through post-training to inside vLLM so training and inference no longer run two copies that can quietly disagree. In supported on-policy configurations the trainer and generator produce bitwise identical log probabilities, leaving exactly zero gap at every step. The update also reports DeepSeek V3 671B throughput on the MLPerf training benchmark and infrastructure work around FlexShard and torch_checkpointing.

PyTorch users on AMD MI350-series GPUs can now reach FlyDSL kernels through existing operator APIs for dense and grouped GEMM, MXFP8/MXFP4 scaled GEMM, RMSNorm, TopK, and attention via AOTriton. Those operators previously ran on ATen, Triton, or AITER ASM implementations. On MI355X operator suites, FlyDSL gives a 1.10x geometric-mean speedup over the faster ATen/Triton baseline at each of 15 BF16 dense-GEMM shapes, and 1.58x and 1.68x over ATen for MXFP8 and MXFP4. End-to-end vLLM whole-request speedups reach 1.13x for BF16, 2.04x for MXFP8, and 3.01x for MXFP4, depending on model and concurrency.

PyTorch users on AMD MI350-series GPUs can now reach FlyDSL kernels through existing operator APIs for dense and grouped GEMM, MXFP8/MXFP4 scaled GEMM, RMSNorm, TopK, and attention via AOTriton. Those operators previously ran on ATen, Triton, or AITER ASM implementations. On MI355X operator suites, FlyDSL gives a 1.10x geometric-mean speedup over the faster ATen/Triton baseline at each of 15 BF16 dense-GEMM shapes, and 1.58x and 1.68x over ATen for MXFP8 and MXFP4. End-to-end vLLM whole-request speedups reach 1.13x for BF16, 2.04x for MXFP8, and 3.01x for MXFP4, depending on model and concurrency.

Canada's national wildfire hazard map has relied on BurnP3+ simulating millions of ignitions, taking over a year to produce. A Mila model trained with the Canadian Forest Service emulates BurnP3+ in seconds, running 806 to 3400 times faster depending on hardware, cutting a one-year task to 2 to 11 hours. It does not replace BurnP3+, but screens scenarios before precise runs.

Canada's national wildfire hazard map has relied on BurnP3+ simulating millions of ignitions, taking over a year to produce. A Mila model trained with the Canadian Forest Service emulates BurnP3+ in seconds, running 806 to 3400 times faster depending on hardware, cutting a one-year task to 2 to 11 hours. It does not replace BurnP3+, but screens scenarios before precise runs.

Agents often claim a task is done while leaving the wrong database record. ThinkingBox turns deterministic checks of terminal database state into rewards and trains Qwen3.8-27B on tasks outside its benchmark: after 50 GRPO steps, average success rose from 51.70% to 60.78% and tasks passing all twenty attempts rose from 38 to 107, while tasks passing at least once fell from 453 to 446. The training pool was selected offline by policy-relative difficulty, and none of the 507 benchmark tasks were used for training.

Agents often claim a task is done while leaving the wrong database record. ThinkingBox turns deterministic checks of terminal database state into rewards and trains Qwen3.8-27B on tasks outside its benchmark: after 50 GRPO steps, average success rose from 51.70% to 60.78% and tasks passing all twenty attempts rose from 38 to 107, while tasks passing at least once fell from 453 to 446. The training pool was selected offline by policy-relative difficulty, and none of the 507 benchmark tasks were used for training.

Instead of markup that keeps a table's layout, the output is an executable SQLite database written straight from one document image. PIX2DB fine-tunes Qwen3-VL-2B with LoRA on 36,589 pairs with exact gold databases and adds a deterministic verify-and-repair layer, reaching 0.86 table exact match and 0.99 foreign-key recall on rendered tables from unseen databases, against 0.75/0.75 for the same-family 8B model prompted for SQL and 0.47/0.00 for an 8B HTML-to-SQLite pipeline. On TabRecSet, unseen in training, it is the best system that writes a database directly (0.52) but trails the 8B HTML pipeline (0.58), mostly in reading text from photographs.

Instead of markup that keeps a table's layout, the output is an executable SQLite database written straight from one document image. PIX2DB fine-tunes Qwen3-VL-2B with LoRA on 36,589 pairs with exact gold databases and adds a deterministic verify-and-repair layer, reaching 0.86 table exact match and 0.99 foreign-key recall on rendered tables from unseen databases, against 0.75/0.75 for the same-family 8B model prompted for SQL and 0.47/0.00 for an 8B HTML-to-SQLite pipeline. On TabRecSet, unseen in training, it is the best system that writes a database directly (0.52) but trails the 8B HTML pipeline (0.58), mostly in reading text from photographs.

Kolibri is an in-house LLM post-trained with reinforcement learning on 1.2 million tasks across 38 environments, and its training stack was shaped by the scaling problems that appeared along the way. Early math and code checks shared one Python process with every rollout, so a stuck parser stalled other rollouts in that process; running each check in an isolated process with a timeout now costs only its own rollout, which scores zero. Sandboxes, judge models and retrieval services moved from external dependencies to self-hosting inside the training job, supporting about 25,000 active sandboxes at any moment.

Kolibri is an in-house LLM post-trained with reinforcement learning on 1.2 million tasks across 38 environments, and its training stack was shaped by the scaling problems that appeared along the way. Early math and code checks shared one Python process with every rollout, so a stuck parser stalled other rollouts in that process; running each check in an isolated process with a timeout now costs only its own rollout, which scores zero. Sandboxes, judge models and retrieval services moved from external dependencies to self-hosting inside the training job, supporting about 25,000 active sandboxes at any moment.

Pricing American options on several assets is hard because the early-exercise boundary is unknown, and extending knot methods to d assets by tensor product needs O(N^d) knots. A physics-informed neural network represents the option value with a tanh feed-forward network, uses automatic differentiation for the multi-asset Black–Scholes operator, and trains on sampled collocation points instead of a grid. On a two-asset American basket put, the 8000-epoch configuration with a one-sided residual penalty and a smooth Fischer–Burmeister complementarity function gave 7.8669, an overestimate of 0.97, about 14%, against the Longstaff–Schwartz Monte Carlo reference of 6.9005, and the closest of eight training setups.

Pricing American options on several assets is hard because the early-exercise boundary is unknown, and extending knot methods to d assets by tensor product needs O(N^d) knots. A physics-informed neural network represents the option value with a tanh feed-forward network, uses automatic differentiation for the multi-asset Black–Scholes operator, and trains on sampled collocation points instead of a grid. On a two-asset American basket put, the 8000-epoch configuration with a one-sided residual penalty and a smooth Fischer–Burmeister complementarity function gave 7.8669, an overestimate of 0.97, about 14%, against the Longstaff–Schwartz Monte Carlo reference of 6.9005, and the closest of eight training setups.

Single-cell 3D genome data are so sparse that domain boundaries are hard to call reliably, and positional disagreement across cells blocks comparison of boundary frequency and strength. CellRoB couples single-cell boundary detection with cross-cell alignment that minimizes an insulation-score loss, improving boundary accuracy and robustness to contact dropout on mouse brain Dip-C and human GM12878 scMicro-C data from 40 kb down to 5 kb. On multimodal mouse embryo data, boundary-strength profiles split the EPI cells that the RNA embedding left as one group.

Single-cell 3D genome data are so sparse that domain boundaries are hard to call reliably, and positional disagreement across cells blocks comparison of boundary frequency and strength. CellRoB couples single-cell boundary detection with cross-cell alignment that minimizes an insulation-score loss, improving boundary accuracy and robustness to contact dropout on mouse brain Dip-C and human GM12878 scMicro-C data from 40 kb down to 5 kb. On multimodal mouse embryo data, boundary-strength profiles split the EPI cells that the RNA embedding left as one group.

Patient-safety reasoning requires linking clinical narratives, medications, laboratory findings and adverse events, and a language model alone can produce plausible but unsupported conclusions. The same MIMIC-III/IV cases ran under four conditions with one LLaMA 2 7B model and one prompt: model only, plus retrieval-augmented generation, plus a patient-safety knowledge graph and ontology, and both. The full framework reached F1 0.89 for concept extraction, 0.046 above baseline; removing the knowledge graph cost the most, with F1 0.047, AUROC 0.084 and AUPRC 0.110, while removing retrieval gave an F1 difference of 0.029 whose interval crossed zero.

Patient-safety reasoning requires linking clinical narratives, medications, laboratory findings and adverse events, and a language model alone can produce plausible but unsupported conclusions. The same MIMIC-III/IV cases ran under four conditions with one LLaMA 2 7B model and one prompt: model only, plus retrieval-augmented generation, plus a patient-safety knowledge graph and ontology, and both. The full framework reached F1 0.89 for concept extraction, 0.046 above baseline; removing the knowledge graph cost the most, with F1 0.047, AUROC 0.084 and AUPRC 0.110, while removing retrieval gave an F1 difference of 0.029 whose interval crossed zero.

IBM Research released Granite Timeseries Ensemble, a default forecasting recipe that combines four open Granite time-series models with different architectures and returns median and quantile forecasts without retraining or fitting ensemble weights. Previously a user had to judge which single model suited a given series. Across 97 GIFT-Eval configurations the probabilistic ensemble reaches a MASE of 0.683 against 0.685 for the strongest member, PatchTST-FM-r2, a small margin whose main benefit is not having to pick a model in advance.

IBM Research released Granite Timeseries Ensemble, a default forecasting recipe that combines four open Granite time-series models with different architectures and returns median and quantile forecasts without retraining or fitting ensemble weights. Previously a user had to judge which single model suited a given series. Across 97 GIFT-Eval configurations the probabilistic ensemble reaches a MASE of 0.683 against 0.685 for the strongest member, PatchTST-FM-r2, a small margin whose main benefit is not having to pick a model in advance.

Parkinson's disease diagnosis still rests on symptoms and expert assessment, while multimodal clinical data are heterogeneous, labels are scarce, and deep models rarely report calibrated confidence. NeuroQ-MGST encodes voice, gait, neuroimaging and clinical variables with spiking neural networks, aligns them through self-supervised contrastive and masked reconstruction objectives, reasons over a dynamically rebuilt heterogeneous graph with a hierarchical graph transformer, and adds quantum-inspired variational refinement plus Monte Carlo dropout uncertainty. On the PMMD 2.0 test set it reaches 98.7% accuracy and 0.989 AUC-ROC, and 95.8% zero-shot accuracy on an external European cohort without fine-tuning.

Parkinson's disease diagnosis still rests on symptoms and expert assessment, while multimodal clinical data are heterogeneous, labels are scarce, and deep models rarely report calibrated confidence. NeuroQ-MGST encodes voice, gait, neuroimaging and clinical variables with spiking neural networks, aligns them through self-supervised contrastive and masked reconstruction objectives, reasons over a dynamically rebuilt heterogeneous graph with a hierarchical graph transformer, and adds quantum-inspired variational refinement plus Monte Carlo dropout uncertainty. On the PMMD 2.0 test set it reaches 98.7% accuracy and 0.989 AUC-ROC, and 95.8% zero-shot accuracy on an external European cohort without fine-tuning.

Z.ai released GLM 5.2, a flagship built for long-horizon agentic tasks on a solid 1M-token context and positioned by its published benchmarks as the strongest open-source coding model. Rather than forwarding vendor numbers, Fireworks ran the open weights on its own GPUs and inference engine and scored 91.4% (181/198) on GPQA-Diamond in high reasoning mode, matching Z.ai's reported 91.2%. The model ships under an MIT license and is live on Fireworks for developers and LLM researchers.

Z.ai released GLM 5.2, a flagship built for long-horizon agentic tasks on a solid 1M-token context and positioned by its published benchmarks as the strongest open-source coding model. Rather than forwarding vendor numbers, Fireworks ran the open weights on its own GPUs and inference engine and scored 91.4% (181/198) on GPQA-Diamond in high reasoning mode, matching Z.ai's reported 91.2%. The model ships under an MIT license and is live on Fireworks for developers and LLM researchers.

Holocene agriculture in China did not expand as one uniform farming frontier. Combining archaeological crop records with palaeoclimate, terrain and river-network data, a crop-specific dynamic non-local diffusion model estimates the relative probability of new crop records during 6000–2000 BP. Millet records were most associated with same-crop persistence and non-local diffusion, rice records most with humid-lowland environmental filtering, and wheat/barley records with corridor-like accessibility along the Hexi Corridor–Central Plains belts and higher frontier mobility.

Holocene agriculture in China did not expand as one uniform farming frontier. Combining archaeological crop records with palaeoclimate, terrain and river-network data, a crop-specific dynamic non-local diffusion model estimates the relative probability of new crop records during 6000–2000 BP. Millet records were most associated with same-crop persistence and non-local diffusion, rice records most with humid-lowland environmental filtering, and wheat/barley records with corridor-like accessibility along the Hexi Corridor–Central Plains belts and higher frontier mobility.

Software test-case design has long depended on manual work, and models that treat generation as one-shot text translation struggle with multi-layered business rules. A pipeline of five roles, parser, triage, architect, writer and auditor, turns test generation into a stateful reasoning process in which the auditor corrects scripts before output. Across five datasets spanning travel, telecom, financial regulation and a government job platform, the pipeline averaged a 93.4% mutation score, 1.4% hallucination and 7.8% redundancy, at an average generation time of 21.2 hours.

Software test-case design has long depended on manual work, and models that treat generation as one-shot text translation struggle with multi-layered business rules. A pipeline of five roles, parser, triage, architect, writer and auditor, turns test generation into a stateful reasoning process in which the auditor corrects scripts before output. Across five datasets spanning travel, telecom, financial regulation and a government job platform, the pipeline averaged a 93.4% mutation score, 1.4% hallucination and 7.8% redundancy, at an average generation time of 21.2 hours.

GRPO, a common reinforcement-learning recipe for reasoning models, judges a rollout only by whether its final answer is right, so a correct answer can hide unreliable reasoning, and a single corrupted reward can strongly distort group-relative advantages when a response group is highly imbalanced. S-GRPO minimizes the expected squared error between observed and latent advantages under a reward-noise model to obtain a closed-form, noise-aware group weight. Across several mathematical reasoning benchmarks and model scales it stabilizes training and keeps making progress under controlled reward noise where standard GRPO shows little meaningful learning.

GRPO, a common reinforcement-learning recipe for reasoning models, judges a rollout only by whether its final answer is right, so a correct answer can hide unreliable reasoning, and a single corrupted reward can strongly distort group-relative advantages when a response group is highly imbalanced. S-GRPO minimizes the expected squared error between observed and latent advantages under a reward-noise model to obtain a closed-form, noise-aware group weight. Across several mathematical reasoning benchmarks and model scales it stabilizes training and keeps making progress under controlled reward noise where standard GRPO shows little meaningful learning.

At the current graviton mass bound, the static correction from a massive graviton is too small for the floating-point formats of virtual-world physics engines to store: it first becomes representable at about 5.4×10⁸ m in float64 and 1.25×10¹³ m in float32. Naive in-format evaluation manufactures phantom corrections up to 9×10¹⁶ times the true value, and a float32 orbit at 1 AU reports 834 km where the true effect is 8 m. The same physics behaves oppositely in AI formats: every format, including 8-bit FP8, preserves the gravitational-wave dispersion residual, with representational SNR from 9.7 to 2×10⁷.

At the current graviton mass bound, the static correction from a massive graviton is too small for the floating-point formats of virtual-world physics engines to store: it first becomes representable at about 5.4×10⁸ m in float64 and 1.25×10¹³ m in float32. Naive in-format evaluation manufactures phantom corrections up to 9×10¹⁶ times the true value, and a float32 orbit at 1 AU reports 834 km where the true effect is 8 m. The same physics behaves oppositely in AI formats: every format, including 8-bit FP8, preserves the gravitational-wave dispersion residual, with representational SNR from 9.7 to 2×10⁷.

Open-weight Kimi K3 is available on Fireworks for inference and training from day one, which the release presents as frontier-level capability at a fraction of closed-model cost. On Arena's Frontend Code Arena it scores 1,679 points, ahead of Fable 5, making it the first open model to sit ahead of every closed one; it ranks third worldwide on the Artificial Analysis Intelligence Index at 57. Fireworks also offers US-only inference endpoints with zero data retention by default and a private preview of serverless training on a shared GPU pool.

Open-weight Kimi K3 is available on Fireworks for inference and training from day one, which the release presents as frontier-level capability at a fraction of closed-model cost. On Arena's Frontend Code Arena it scores 1,679 points, ahead of Fable 5, making it the first open model to sit ahead of every closed one; it ranks third worldwide on the Artificial Analysis Intelligence Index at 57. Fireworks also offers US-only inference endpoints with zero data retention by default and a private preview of serverless training on a shared GPU pool.

Data visualizations produced by AI agents usually stop at a static image, so readers who want to zoom into detail, compare across charts, or link a chart selection to nearby text must fall back on low-level code and repeated debugging. Flint-interactive adds an interaction language on top of Flint, letting users or agents describe behavior with high-level semantic events and update operators while Flint's compiler turns clicks, drags and key presses into semantic events that identify chart elements and data, then into canvas updates. Separating interaction logic from visual logic also lets the same model connect charts to tables, text, other charts and widgets for dashboards and data stories.

Data visualizations produced by AI agents usually stop at a static image, so readers who want to zoom into detail, compare across charts, or link a chart selection to nearby text must fall back on low-level code and repeated debugging. Flint-interactive adds an interaction language on top of Flint, letting users or agents describe behavior with high-level semantic events and update operators while Flint's compiler turns clicks, drags and key presses into semantic events that identify chart elements and data, then into canvas updates. Separating interaction logic from visual logic also lets the same model connect charts to tables, text, other charts and widgets for dashboards and data stories.

In enterprise agent deployments, folding routers, planners and domain subagents into a single main loop often beat both a one-prompt-for-everything design and a subagent orchestration design on quality, frequently at lower cost and latency per task. Subagents still fit when context bloat degrades a single loop, when tasks need parallel fresh context windows, or when tool choice and instructions conflict. A guided workflow lets teams try the migration on a mock system first.

In enterprise agent deployments, folding routers, planners and domain subagents into a single main loop often beat both a one-prompt-for-everything design and a subagent orchestration design on quality, frequently at lower cost and latency per task. Subagents still fit when context bloat degrades a single loop, when tasks need parallel fresh context windows, or when tool choice and instructions conflict. A guided workflow lets teams try the migration on a mock system first.

Eval Protocol plugs GEPA prompt optimization into an existing evaluation interface, so the same datasets, metrics and task constraints both measure an agent and automatically rewrite its system prompt. In a Text2SQL case, Qwen3-32b gained 10% test-set accuracy and more than 5% validation-set accuracy without touching model weights, with about 45 minutes of prompt search; the same eval then served as the objective for RFT.

Eval Protocol plugs GEPA prompt optimization into an existing evaluation interface, so the same datasets, metrics and task constraints both measure an agent and automatically rewrite its system prompt. In a Text2SQL case, Qwen3-32b gained 10% test-set accuracy and more than 5% validation-set accuracy without touching model weights, with about 45 minutes of prompt search; the same eval then served as the objective for RFT.

NVIDIA announced at GTC Berlin official support for eight-node DGX Spark clusters, raising unified memory from 128 GB on a single node to 1 TB and letting frontier-class models that would not start on a four-node system load locally. The full GLM 5.3 (753B) checkpoint runs on the eight-node cluster, GLM 5.3 Flash generates at roughly 80 tokens per second, and decode throughput on dense Qwen3.8-27B (NVFP4) rises from about 12 TPS on one node to 52 TPS on eight.

NVIDIA announced at GTC Berlin official support for eight-node DGX Spark clusters, raising unified memory from 128 GB on a single node to 1 TB and letting frontier-class models that would not start on a four-node system load locally. The full GLM 5.3 (753B) checkpoint runs on the eight-node cluster, GLM 5.3 Flash generates at roughly 80 tokens per second, and decode throughput on dense Qwen3.8-27B (NVFP4) rises from about 12 TPS on one node to 52 TPS on eight.

Diagnosis of constraint systems needs several minimal conflicts at once, and MergeXplain returns multiple conflicts from one decomposition, but only its two recursive branches had been run in parallel while the merge loop that eliminates boundary-spanning conflicts stayed sequential. ParallelMergeXplain forks the recursion up to a bounded depth and replaces every GetConflict call in the merge loop with a speculative, QXGen-lookahead parallel version. On a 2-core machine it reached a mean wall-clock speed-up of 1.21x (std. 0.20) across all configurations with non-trivial checks, and 1.531x for par full at p=2 on the real SATLIB uf50-218 instance, returning results identical to sequential MergeXplain.

Diagnosis of constraint systems needs several minimal conflicts at once, and MergeXplain returns multiple conflicts from one decomposition, but only its two recursive branches had been run in parallel while the merge loop that eliminates boundary-spanning conflicts stayed sequential. ParallelMergeXplain forks the recursion up to a bounded depth and replaces every GetConflict call in the merge loop with a speculative, QXGen-lookahead parallel version. On a 2-core machine it reached a mean wall-clock speed-up of 1.21x (std. 0.20) across all configurations with non-trivial checks, and 1.531x for par full at p=2 on the real SATLIB uf50-218 instance, returning results identical to sequential MergeXplain.

Active label inference in vertical federated learning leaves a class-separability anomaly in the adversary's embeddings, letting the server trigger a defense that acts on the adversary alone and pushes label inference below what honest training already leaks on CIFAR-10, CIFAR-100 and CINIC-10, at under two points of task-accuracy cost.

Active label inference in vertical federated learning leaves a class-separability anomaly in the adversary's embeddings, letting the server trigger a defense that acts on the adversary alone and pushes label inference below what honest training already leaks on CIFAR-10, CIFAR-100 and CINIC-10, at under two points of task-accuracy cost.

From 88 lactobacilli isolates from human milk, infant stool and fermented foods, two Lactiplantibacillus plantarum strains, CYLB2146 and CYLB2189, inhibited Gardnerella vaginalis, Streptococcus agalactiae, Neisseria gonorrhoeae and Candida spp. and lowered inflammatory cytokines in vaginal epithelial cells. In a murine bacterial vaginosis model, live bacteria and their cell-free supernatants reduced G. vaginalis loads from 51.2- and 41.6-fold in controls to 3.2-, 11.1-, 3.2- and 0.7-fold after 5 days, approaching healthy levels by day 10.

From 88 lactobacilli isolates from human milk, infant stool and fermented foods, two Lactiplantibacillus plantarum strains, CYLB2146 and CYLB2189, inhibited Gardnerella vaginalis, Streptococcus agalactiae, Neisseria gonorrhoeae and Candida spp. and lowered inflammatory cytokines in vaginal epithelial cells. In a murine bacterial vaginosis model, live bacteria and their cell-free supernatants reduced G. vaginalis loads from 51.2- and 41.6-fold in controls to 3.2-, 11.1-, 3.2- and 0.7-fold after 5 days, approaching healthy levels by day 10.

Liquid AI released RIPL, a low-latency inference engine that puts Liquid Foundation Models directly in the path of a live request and returns complete generated responses end-to-end in under 50 ms while traffic keeps arriving at peak load. Most serving infrastructure targets general-purpose models, large batches and long-running agentic workflows where seconds per request are acceptable, while enterprise real-time workloads such as search ranking, recommendations and fraud decisions need answers in milliseconds and have relied on faster but less accurate methods like gradient-boosted trees or small embedding models. RIPL is aimed at changing that trade-off and is already running on Shopify production workloads including search enhancement.

Liquid AI released RIPL, a low-latency inference engine that puts Liquid Foundation Models directly in the path of a live request and returns complete generated responses end-to-end in under 50 ms while traffic keeps arriving at peak load. Most serving infrastructure targets general-purpose models, large batches and long-running agentic workflows where seconds per request are acceptable, while enterprise real-time workloads such as search ranking, recommendations and fraud decisions need answers in milliseconds and have relied on faster but less accurate methods like gradient-boosted trees or small embedding models. RIPL is aimed at changing that trade-off and is already running on Shopify production workloads including search enhancement.

Psychological prediction must say both what a character will do and what it believes, yet models tuned on surface patterns often return action and belief answers that disagree. A neuro-symbolic reasoning framework splits the latent cognitive state into a belief representation and an executable program representation, filters hypotheses that violate contextual constraints through energy-penalized particle resampling, and lets the two pathways update separately in stop-gradient alternating fixed-point iterations. Under the text-based classification setting on OpenToM and FANToM, the share of samples with both action and belief correct reaches 70.8% and 52.1%, above scale-matched pure encoder classifiers.

Psychological prediction must say both what a character will do and what it believes, yet models tuned on surface patterns often return action and belief answers that disagree. A neuro-symbolic reasoning framework splits the latent cognitive state into a belief representation and an executable program representation, filters hypotheses that violate contextual constraints through energy-penalized particle resampling, and lets the two pathways update separately in stop-gradient alternating fixed-point iterations. Under the text-based classification setting on OpenToM and FANToM, the share of samples with both action and belief correct reaches 70.8% and 52.1%, above scale-matched pure encoder classifiers.

Machine-learned corrections to global buckling predictions for built-up cold-formed steel columns were previously judged on a random 80/20 split, where specimens from one test programme can sit in both training and test sets. Holding out entire experimental sources instead, across 148 tests from 15 sources and 13 section families, weakens transfer sharply: the best model, Lasso, reaches an $R^2$ of 0.787 and a mean absolute percentage error of 73.0%. Calibrated on those out-of-source predictions, LRFD resistance factors fall to 0.191–0.399 and ASD safety factors to 4.01–8.38.

Machine-learned corrections to global buckling predictions for built-up cold-formed steel columns were previously judged on a random 80/20 split, where specimens from one test programme can sit in both training and test sets. Holding out entire experimental sources instead, across 148 tests from 15 sources and 13 section families, weakens transfer sharply: the best model, Lasso, reaches an $R^2$ of 0.787 and a mean absolute percentage error of 73.0%. Calibrated on those out-of-source predictions, LRFD resistance factors fall to 0.191–0.399 and ASD safety factors to 4.01–8.38.

At GTC Berlin, Nvidia laid out a local AI hardware line running from Windows PCs to desk-side workstations, and said DGX Spark can be manually configured into clusters of up to eight units with 1TB of unified memory, enough to run frontier-class models such as GLM 5.3. Previously, local devices were bounded by single-machine memory and ran smaller models or leaned on the cloud. A companion NVIDIA Sync Model Launcher reduces model download, container setup and OpenCode connection to a few clicks.

At GTC Berlin, Nvidia laid out a local AI hardware line running from Windows PCs to desk-side workstations, and said DGX Spark can be manually configured into clusters of up to eight units with 1TB of unified memory, enough to run frontier-class models such as GLM 5.3. Previously, local devices were bounded by single-machine memory and ran smaller models or leaned on the cloud. A companion NVIDIA Sync Model Launcher reduces model download, container setup and OpenCode connection to a few clicks.

Claude Models on Google Cloud (Anthropic managed) is now generally available, giving Google Cloud customers the full Claude Platform feature set, with beta and new features shipping the same day they launch natively. Previously the Google Cloud route was the Google managed offering, where Google acts as data processor. The new path uses existing Google Cloud IAM credentials, bills through a single Google Cloud invoice, draws down existing commitments, and is offered on US and Global endpoints.

Claude Models on Google Cloud (Anthropic managed) is now generally available, giving Google Cloud customers the full Claude Platform feature set, with beta and new features shipping the same day they launch natively. Previously the Google Cloud route was the Google managed offering, where Google acts as data processor. The new path uses existing Google Cloud IAM credentials, bills through a single Google Cloud invoice, draws down existing commitments, and is offered on US and Global endpoints.

Adyen engineers have open-sourced Agent Marketplace, a command-line tool for sharing AI agent resources across teams, repositories and tools. Previously each AI coding tool kept its setup in its own folders, file formats and configuration fields, so instructions written for one tool rarely worked in another. The tool fetches resources listed in one configuration file, translates them into the format each tool expects and writes them where that tool looks, compiling a single resource for eight tools including Claude Code, Codex and Cursor, while a repeated sync changes nothing.

Adyen engineers have open-sourced Agent Marketplace, a command-line tool for sharing AI agent resources across teams, repositories and tools. Previously each AI coding tool kept its setup in its own folders, file formats and configuration fields, so instructions written for one tool rarely worked in another. The tool fetches resources listed in one configuration file, translates them into the format each tool expects and writes them where that tool looks, compiling a single resource for eight tools including Claude Code, Codex and Cursor, while a repeated sync changes nothing.

Visible and infrared images follow different imaging mechanisms, so direct concatenation or one-shot attention fusion can mistake background texture and thermal interference for complementary evidence. FCI-DETR splits detection into three linked stages: preserving modality-specific structure before fusion, calibrating cross-modal responses during fusion, and injecting infrared context by scale after fusion. It reaches 88.6%, 84.3%, 86.7% and 81.5% mAP50 on M3FD, DroneVehicle, VEDAI and FLIR, and beats a same-backbone RGB+IR RT-DETR baseline by 6.0 points on M3FD under a fixed protocol.

Visible and infrared images follow different imaging mechanisms, so direct concatenation or one-shot attention fusion can mistake background texture and thermal interference for complementary evidence. FCI-DETR splits detection into three linked stages: preserving modality-specific structure before fusion, calibrating cross-modal responses during fusion, and injecting infrared context by scale after fusion. It reaches 88.6%, 84.3%, 86.7% and 81.5% mAP50 on M3FD, DroneVehicle, VEDAI and FLIR, and beats a same-backbone RGB+IR RT-DETR baseline by 6.0 points on M3FD under a fixed protocol.

Progressive neural networks transfer knowledge by keeping every frozen earlier column, so stored state and inference grow with each generation. WeanNet keeps a single frozen parent, feeds its layer-wise activations to a fresh student through a scalar gain, and decays that gain linearly to zero during training, so the parent can be removed at deployment. On a concept-drift task it found the new correct action on drifted states more often than the architecturally matched PNN-Reset control at every tested entropy setting, though it also led on never-drifted states and stayed comparable to training from scratch.

Progressive neural networks transfer knowledge by keeping every frozen earlier column, so stored state and inference grow with each generation. WeanNet keeps a single frozen parent, feeds its layer-wise activations to a fresh student through a scalar gain, and decays that gain linearly to zero during training, so the parent can be removed at deployment. On a concept-drift task it found the new correct action on drifted states more often than the architecturally matched PNN-Reset control at every tested entropy setting, though it also led on never-drifted states and stayed comparable to training from scratch.

Google Cloud has made generally available an Anthropic-managed deployment option for Claude models, sitting alongside the existing Google-managed option. Both deliver the same frontier models, the same pricing and the same Google Cloud billing; they differ in who acts as data processor and whether new Claude platform features are available when they launch. Under the Anthropic-managed option, Anthropic operates the stack on Google Cloud infrastructure, while organizations still authenticate with Google Cloud IAM and keep Cloud Audit Logging.

Google Cloud has made generally available an Anthropic-managed deployment option for Claude models, sitting alongside the existing Google-managed option. Both deliver the same frontier models, the same pricing and the same Google Cloud billing; they differ in who acts as data processor and whether new Claude platform features are available when they launch. Under the Anthropic-managed option, Anthropic operates the stack on Google Cloud infrastructure, while organizations still authenticate with Google Cloud IAM and keep Cloud Audit Logging.

NVIDIA released Minerva, an AIOps platform for end-to-end observability of AI factories built from large GPU fleets. Previously networking, compute, site reliability and machine learning teams each saw a separate slice of the data center, and alerts arrived disconnected. Minerva pools telemetry from GPUs, SuperNICs, switches, cables, storage, schedulers and workloads into a dynamic knowledge graph, then uses supervised and unsupervised learning to detect anomalies, predict failures and correlate events by topology, collapsing an alert storm into one incident with a likely root cause.

NVIDIA released Minerva, an AIOps platform for end-to-end observability of AI factories built from large GPU fleets. Previously networking, compute, site reliability and machine learning teams each saw a separate slice of the data center, and alerts arrived disconnected. Minerva pools telemetry from GPUs, SuperNICs, switches, cables, storage, schedulers and workloads into a dynamic knowledge graph, then uses supervised and unsupervised learning to detect anomalies, predict failures and correlate events by topology, collapsing an alert storm into one incident with a likely root cause.

Sports images sit with schools, clubs and other organizations, and federated learning avoids sending raw pictures but still transmits model updates that can leak information, while the clipping and noise of differential privacy damage useful knowledge. FedSS scores each network stage for how shareable it is before clipping, uploads only the perturbed cross-client sport semantics, and keeps client-specific appearance and background context local. On sport22 and sport100 it beats the strongest baseline DPSIR-Fed by 5.78 and 5.71 points under visual-context heterogeneity, and raises sport22 accuracy from 42.67% to 46.81% at the strong-privacy setting ε=4.

Sports images sit with schools, clubs and other organizations, and federated learning avoids sending raw pictures but still transmits model updates that can leak information, while the clipping and noise of differential privacy damage useful knowledge. FedSS scores each network stage for how shareable it is before clipping, uploads only the perturbed cross-client sport semantics, and keeps client-specific appearance and background context local. On sport22 and sport100 it beats the strongest baseline DPSIR-Fed by 5.78 and 5.71 points under visual-context heterogeneity, and raises sport22 accuracy from 42.67% to 46.81% at the strong-privacy setting ε=4.

Recording real graphical-interface operations, parameterizing them into a frozen workflow and handing that to an agent can cover more business requirements than letting the agent write code directly. On three complete-task instances of one SaaS business family, the frozen workflow reached 33/33 local core points while raw code reached 8, 8 and 18; on two prospectively fixed instances the tool agents finished in six and five model calls, against 121 and 70 for raw code. Every attempt still failed whole-task acceptance, and collateral state preservation remains unknown.

Recording real graphical-interface operations, parameterizing them into a frozen workflow and handing that to an agent can cover more business requirements than letting the agent write code directly. On three complete-task instances of one SaaS business family, the frozen workflow reached 33/33 local core points while raw code reached 8, 8 and 18; on two prospectively fixed instances the tool agents finished in six and five model calls, against 121 and 70 for raw code. Every attempt still failed whole-task acceptance, and collateral state preservation remains unknown.

Single-channel blind separation of co-frequency communication signals has relied on real-valued networks or attention modules blind to complex structure, which cannot exploit the phase coupling between the I and Q rails. ComplexLightweightSepNet embeds a complex squeeze-and-excitation block in a complex convolutional encoder-decoder, scaling complex features with a real-valued channel weight that preserves phase, at 235,525 parameters. On 5 Hz co-frequency synthetic mixtures across four modulations it reaches a five-seed mean SDR of 2.31 dB with 0.62 dB standard deviation; paired Wilcoxon tests do not reach alpha = 0.05, so the result is positioned as a working point at the 235K scale, not an inferred advantage.

Single-channel blind separation of co-frequency communication signals has relied on real-valued networks or attention modules blind to complex structure, which cannot exploit the phase coupling between the I and Q rails. ComplexLightweightSepNet embeds a complex squeeze-and-excitation block in a complex convolutional encoder-decoder, scaling complex features with a real-valued channel weight that preserves phase, at 235,525 parameters. On 5 Hz co-frequency synthetic mixtures across four modulations it reaches a five-seed mean SDR of 2.31 dB with 0.62 dB standard deviation; paired Wilcoxon tests do not reach alpha = 0.05, so the result is positioned as a working point at the 235K scale, not an inferred advantage.

Swarm synchronization for green IoT no longer broadcasts state on a fixed cycle. ET-S3 ties each agent's trigger threshold to its remaining battery, so low-energy nodes transmit less, and a control-barrier-function quadratic program filters the consensus input for safety. In a simulation of 100 dynamic agents over 1000 synchronization cycles, cumulative payload was 3.35 MB, 18.2% of the 18.40 MB of periodic consensus, with a 0.0% safety violation rate and convergence in 138 steps.

Swarm synchronization for green IoT no longer broadcasts state on a fixed cycle. ET-S3 ties each agent's trigger threshold to its remaining battery, so low-energy nodes transmit less, and a control-barrier-function quadratic program filters the consensus input for safety. In a simulation of 100 dynamic agents over 1000 synchronization cycles, cumulative payload was 3.35 MB, 18.2% of the 18.40 MB of periodic consensus, with a 0.0% safety violation rate and convergence in 138 steps.

Fireworks released a beta speech-to-text API for the Whisper v3-large models, built around speed: on a dedicated endpoint, the asynchronous endpoint transcribes or translates one hour of audio in 4 seconds, about 900 times real time. Previously such services left users waiting seconds to minutes, and the gain here comes from serving-stack changes rather than a new model. On quality, whisper-v3-large reports a 2.00% word error rate for greedy inference on Librispeech Clean, matching the original open-source model; translation, word- and sentence-level timestamp alignment, audio preprocessing and voice activity detection ship in the same API.

Fireworks released a beta speech-to-text API for the Whisper v3-large models, built around speed: on a dedicated endpoint, the asynchronous endpoint transcribes or translates one hour of audio in 4 seconds, about 900 times real time. Previously such services left users waiting seconds to minutes, and the gain here comes from serving-stack changes rather than a new model. On quality, whisper-v3-large reports a 2.00% word error rate for greedy inference on Librispeech Clean, matching the original open-source model; translation, word- and sentence-level timestamp alignment, audio preprocessing and voice activity detection ship in the same API.

Near-surface air temperature forecasts 24 hours ahead must capture the daily cycle, synoptic evolution and hourly fluctuations at once, while common decomposition transformers defer or hide cross-variable interactions. A cross-channel temporal correlation transformer exchanges information among meteorological channels at the embedding stage via 1×1 convolutions, then fuses temporal, trend and raw branches with gated attention and multi-scale trend attention. On six Chinese city datasets it lowers MSE by 6.2% and MAE by 7.5% on average against mainstream models, with MSE reduction above 9% in Chongqing's extreme heat.

Near-surface air temperature forecasts 24 hours ahead must capture the daily cycle, synoptic evolution and hourly fluctuations at once, while common decomposition transformers defer or hide cross-variable interactions. A cross-channel temporal correlation transformer exchanges information among meteorological channels at the embedding stage via 1×1 convolutions, then fuses temporal, trend and raw branches with gated attention and multi-scale trend attention. On six Chinese city datasets it lowers MSE by 6.2% and MAE by 7.5% on average against mainstream models, with MSE reduction above 9% in Chongqing's extreme heat.

Recognition of psychological stress in dialogue has mostly treated each utterance as an isolated fragment and concatenated modality features at the final layer, which cannot capture how emotion accumulates and shifts across turns. BERT-EmoShiftNet pairs BERT's multi-level semantic encoding with EmoShiftNet's emotion-transition modelling, aligns text, speech and vision through a cross-modal Transformer, and unifies them with cross-modal contrastive learning in a shared embedding space. On a dialogue-level split of a self-built dataset, full-modal five-class accuracy is 85.0% (95% CI [83.2%, 86.8%]), emotion-transition prediction accuracy is 76.2%, and mean detection latency is 13.4 s.

Recognition of psychological stress in dialogue has mostly treated each utterance as an isolated fragment and concatenated modality features at the final layer, which cannot capture how emotion accumulates and shifts across turns. BERT-EmoShiftNet pairs BERT's multi-level semantic encoding with EmoShiftNet's emotion-transition modelling, aligns text, speech and vision through a cross-modal Transformer, and unifies them with cross-modal contrastive learning in a shared embedding space. On a dialogue-level split of a self-built dataset, full-modal five-class accuracy is 85.0% (95% CI [83.2%, 86.8%]), emotion-transition prediction accuracy is 76.2%, and mean detection latency is 13.4 s.

Omani crude oil flows non-isothermally as a viscoplastic fluid, and predicting its heat and head losses has meant running high-fidelity CFD case by case, too slowly for real-time decisions. A physics-informed surrogate trained on 962 COMSOL simulations, with a dimensionless Rheological Heat-Loss Index (RHLI) as an input feature, now predicts total heat loss, head loss and total power loss together. Under five-fold cross-validation heat loss was the most learnable target, with cross-validated R² near 1; uncertainty quantification gave a mean calibration error of 0.0211 and 97.4% empirical 95% coverage, and SHAP ranked ambient temperature as the dominant heat-loss driver.

Omani crude oil flows non-isothermally as a viscoplastic fluid, and predicting its heat and head losses has meant running high-fidelity CFD case by case, too slowly for real-time decisions. A physics-informed surrogate trained on 962 COMSOL simulations, with a dimensionless Rheological Heat-Loss Index (RHLI) as an input feature, now predicts total heat loss, head loss and total power loss together. Under five-fold cross-validation heat loss was the most learnable target, with cross-validated R² near 1; uncertainty quantification gave a mean calibration error of 0.0211 and 97.4% empirical 95% coverage, and SHAP ranked ambient temperature as the dominant heat-loss driver.

Fireworks has added adaptive speculative execution to FireOptimizer, training the draft model for speculative decoding on a customer's own inference workload instead of generic conversation data. Speculative decoding lets a smaller draft model predict several tokens in parallel for the main model to verify, so the draft's hit rate sets the payoff; generic drafts hit poorly on specialized workloads. In one specialized case a generic draft hit 29% and raised latency 1.5x, while FireOptimizer reached a 76% hit rate and 2x speedup, up to 3x over a generic draft model.

Fireworks has added adaptive speculative execution to FireOptimizer, training the draft model for speculative decoding on a customer's own inference workload instead of generic conversation data. Speculative decoding lets a smaller draft model predict several tokens in parallel for the main model to verify, so the draft's hit rate sets the payoff; generic drafts hit poorly on specialized workloads. In one specialized case a generic draft hit 29% and raised latency 1.5x, while FireOptimizer reached a 76% hit rate and 2x speedup, up to 3x over a generic draft model.

Uber is deploying ChatGPT Work across the company, giving more than 30,000 employees worldwide access to an agent inside ChatGPT Enterprise that takes on multi-step tasks. The tool had been limited to a pilot of 1,000 employees, and the expansion makes it Uber's first company-wide deployment of an agentic platform for knowledge work. It connects to Google Drive, Gmail, Google Calendar and Slack, so staff can research, prepare meeting briefs, build presentations, analyze data and complete workflows across applications, with GPT-6.1 Sol as the default capability.

Uber is deploying ChatGPT Work across the company, giving more than 30,000 employees worldwide access to an agent inside ChatGPT Enterprise that takes on multi-step tasks. The tool had been limited to a pilot of 1,000 employees, and the expansion makes it Uber's first company-wide deployment of an agentic platform for knowledge work. It connects to Google Drive, Gmail, Google Calendar and Slack, so staff can research, prepare meeting briefs, build presentations, analyze data and complete workflows across applications, with GPT-6.1 Sol as the default capability.

Open models previously forced a choice between showing a chain of thought and calling tools deterministically. Qwen3-235B-A22B places its reasoning inside a separate think tag and emits the exact JSON tool call in the same completion, so downstream code can ignore or audit the trace. The model is a 235-billion-parameter mixture-of-experts with about 22 billion active parameters, hosted serverlessly on Fireworks under Apache-2.0 and callable through an OpenAI-compatible client.

Open models previously forced a choice between showing a chain of thought and calling tools deterministically. Qwen3-235B-A22B places its reasoning inside a separate think tag and emits the exact JSON tool call in the same completion, so downstream code can ignore or audit the trace. The model is a 235-billion-parameter mixture-of-experts with about 22 billion active parameters, hosted serverlessly on Fireworks under Apache-2.0 and callable through an OpenAI-compatible client.

Fireworks announced support for Meta's Llama 3.2 family, covering the text-only 1B and 3B models and the multimodal 11B Vision and 90B Vision models, with inference and fine-tuning available on its platform. That generation was not previously offered there; the 11B and 90B multimodal models are now priced the same as text-only models, and images are charged as text tokens, typically counted as 6,400 text tokens for the Llama 3.2 vision models. Fine-tuning is currently open for the 1B and 3B text-only models, while multimodal fine-tuning is not yet available.

Fireworks announced support for Meta's Llama 3.2 family, covering the text-only 1B and 3B models and the multimodal 11B Vision and 90B Vision models, with inference and fine-tuning available on its platform. That generation was not previously offered there; the 11B and 90B multimodal models are now priced the same as text-only models, and images are charged as text tokens, typically counted as 6,400 text tokens for the Llama 3.2 vision models. Fine-tuning is currently open for the 1B and 3B text-only models, while multimodal fine-tuning is not yet available.