AI and science frontiers · 2026-10-14

Reversing the declustering metric so that the later mainshock's magnitude, not the earlier event's, sets the correlation range recovers more small foreshocks in both synthetic catalogs and the Southern California catalog, showing that foreshock counts depend not only on earthquake physics but on the temporal symmetry built into the declustering method.

Reversing the declustering metric so that the later mainshock's magnitude, not the earlier event's, sets the correlation range recovers more small foreshocks in both synthetic catalogs and the Southern California catalog, showing that foreshock counts depend not only on earthquake physics but on the temporal symmetry built into the declustering method.

Whether one policy difference transfers depends on the target: Qwen3 targets gain positively while Qwen2.5 targets under a Qwen3 reference gain almost nothing, and swapping in a Qwen2.5-3B reference turns those gains positive.

Whether one policy difference transfers depends on the target: Qwen3 targets gain positively while Qwen2.5 targets under a Qwen3 reference gain almost nothing, and swapping in a Qwen2.5-3B reference turns those gains positive.

A mesh-native tokenisation layer that matches classical convolutions lets grid-pretrained physics foundation models deploy on unstructured meshes without retraining, reaching usable accuracy zero-shot and beating mesh baselines after a few fine-tuning epochs.

A mesh-native tokenisation layer that matches classical convolutions lets grid-pretrained physics foundation models deploy on unstructured meshes without retraining, reaching usable accuracy zero-shot and beating mesh baselines after a few fine-tuning epochs.

Automatic differentiation can now run without fixing a coordinate basis: writing the geometric-algebra multivector derivative as a JAX transformation lets the same operator compute curvature, divergence and rotor-group updates.

Automatic differentiation can now run without fixing a coordinate basis: writing the geometric-algebra multivector derivative as a JAX transformation lets the same operator compute curvature, divergence and rotor-group updates.

Amazon Web Services introduces priming, which initializes hybrid state space models from pretrained Transformer weights and cuts the pretraining cost of hybrid architectures by more than 99%, enabling controlled comparisons of three state space layer types.

Amazon Web Services introduces priming, which initializes hybrid state space models from pretrained Transformer weights and cuts the pretraining cost of hybrid architectures by more than 99%, enabling controlled comparisons of three state space layer types.

Across five open-weight models, holding graph facts fixed and changing only the task instruction redistributes how the same attention heads influence later computation; in Gemma-2-9B and Qwen3-14B, editing one early internal activation vector switches which fact the model reads while the written instruction and weights stay fixed.

Across five open-weight models, holding graph facts fixed and changing only the task instruction redistributes how the same attention heads influence later computation; in Gemma-2-9B and Qwen3-14B, editing one early internal activation vector switches which fact the model reads while the written instruction and weights stay fixed.

SmolWAM compresses the 2B-parameter Cosmos Policy video world-action model to 0.88B, succeeding in 95.5% of 400 LIBERO episodes at 57 ms, 2.3 TFLOP and 2.4 GiB per 16-action chunk, 11 times faster than its teacher while still forecasting scene motion.

SmolWAM compresses the 2B-parameter Cosmos Policy video world-action model to 0.88B, succeeding in 95.5% of 400 LIBERO episodes at 57 ms, 2.3 TFLOP and 2.4 GiB per 16-action chunk, 11 times faster than its teacher while still forecasting scene motion.

The MCES-10 structure-disjoint split used to evaluate mass-spectrum structure prediction percolates once the public library grows from 29,000 to 227,690 molecules, so a single component holds 95.5% of the library, the usable holdout falls to 4.5%, and the surviving compounds are systematically heavier.

The MCES-10 structure-disjoint split used to evaluate mass-spectrum structure prediction percolates once the public library grows from 29,000 to 227,690 molecules, so a single component holds 95.5% of the library, the usable holdout falls to 4.5%, and the surviving compounds are systematically heavier.

A modeling scheme that attaches local games to places of a colored Petri net, so that a game's outcome becomes typed tokens that enable, synchronize, or block later games; in an intrusion campaign, a newly discovered phishing scenario is composed with the existing campaign through a single entry value, matching a full re-solve.

A modeling scheme that attaches local games to places of a colored Petri net, so that a game's outcome becomes typed tokens that enable, synchronize, or block later games; in an intrusion campaign, a newly discovered phishing scenario is composed with the existing campaign through a single entry value, matching a full re-solve.

Sorting creativity benchmarks by how rare their items are under a reference corpus makes the tail subset's model ranking agree more closely with LM Arena creative-writing human preferences than the peak subset, so creativity evaluation can be read as task performance in the long tail.

Sorting creativity benchmarks by how rare their items are under a reference corpus makes the tail subset's model ranking agree more closely with LM Arena creative-writing human preferences than the peak subset, so creativity evaluation can be read as task performance in the long tail.

On a controlled binary benchmark with supplied report lineage, LQET makes models update beliefs by evidence provenance rather than report count, cutting new-evidence update RMSE by 41-44% and log-odds drift after eight relay steps by 89-91% relative to task-only training.

On a controlled binary benchmark with supplied report lineage, LQET makes models update beliefs by evidence provenance rather than report count, cutting new-evidence update RMSE by 41-44% and log-odds drift after eight relay steps by 89-91% relative to task-only training.

Four LLMs rewrote the Kessler warm-rain microphysics Fortran code into GPU-executable JAX, reproducing the Fortran reference to near floating-point precision while differing widely in GPU performance, with the fastest reaching about 357 times the speed of serial Fortran in cached GPU execution.

Four LLMs rewrote the Kessler warm-rain microphysics Fortran code into GPU-executable JAX, reproducing the Fortran reference to near floating-point precision while differing widely in GPU performance, with the fastest reaching about 357 times the speed of serial Fortran in cached GPU execution.

Evaluating perovskite solar-cell machine learning with provenance-grouped splits shows a lookup table that learns nothing matches the published state of the art on a random split, while neural models overtake tree ensembles on unseen institutions and later years.

Evaluating perovskite solar-cell machine learning with provenance-grouped splits shows a lookup table that learns nothing matches the published state of the art on a random split, while neural models overtake tree ensembles on unseen institutions and later years.

A 19-day protocol that matures human iPSC-derived macrophages with M-CSF alone yields cells that take up amyloid-beta and zymosan more than standard iPSC-derived microglia and mount stronger TREM2-driven SYK and PLCγ2 signaling, giving Alzheimer's researchers an easier-to-perturb human macrophage model.

A 19-day protocol that matures human iPSC-derived macrophages with M-CSF alone yields cells that take up amyloid-beta and zymosan more than standard iPSC-derived microglia and mount stronger TREM2-driven SYK and PLCγ2 signaling, giving Alzheimer's researchers an easier-to-perturb human macrophage model.

TopoLIA extends attention from how much information is exchanged to how that information is transformed per target, beating its standard-attention counterpart on protein–ligand scoring and docking benchmarks and reaching an 88% docking success rate on CASF-2013.

TopoLIA extends attention from how much information is exchanged to how that information is transformed per target, beating its standard-attention counterpart on protein–ligand scoring and docking benchmarks and reaching an 88% docking success rate on CASF-2013.

A separately trained diffusion generator maps non-precipitation atmospheric states from a frozen weather model to observation-aligned 24-hour precipitation, improving daily accumulation and event detection on Pangu-Weather, GraphCast and HRES without backbone-specific retraining.

A separately trained diffusion generator maps non-precipitation atmospheric states from a frozen weather model to observation-aligned 24-hour precipitation, improving daily accumulation and event detection on Pangu-Weather, GraphCast and HRES without backbone-specific retraining.

Counterfactual prediction can skip causal-graph estimation: CFGNP maps observational samples directly to a counterfactual distribution, achieves stronger predictive performance than existing approaches across synthetic and image-based benchmarks, and a model trained only on 10-node graphs transfers to graphs up to 30 nodes without retraining.

Counterfactual prediction can skip causal-graph estimation: CFGNP maps observational samples directly to a counterfactual distribution, achieves stronger predictive performance than existing approaches across synthetic and image-based benchmarks, and a model trained only on 10-node graphs transfers to graphs up to 30 nodes without retraining.

High-energy-physics detector data can be compressed into discrete tokens, with track and topocluster tokenizers holding jet resolution below 1% at roughly 4–11x compression, while calorimeter cells stay slightly above that target.

High-energy-physics detector data can be compressed into discrete tokens, with track and topocluster tokenizers holding jet resolution below 1% at roughly 4–11x compression, while calorimeter cells stay slightly above that target.

Across five AI data centers with equal completed work and network resources, folding optical circuit switching into joint compute-electricity-network scheduling cut real-time grid purchase cost by 12.50% and raised local renewable utilization by 2.724 percentage points versus fixed optical operation.

Across five AI data centers with equal completed work and network resources, folding optical circuit switching into joint compute-electricity-network scheduling cut real-time grid purchase cost by 12.50% and raised local renewable utilization by 2.724 percentage points versus fixed optical operation.

JWST/NIRCam F444W slitless spectroscopy over 0.33 square degrees in COSMOS has yielded a homogeneous catalog of 12,275 emission-line galaxies spanning redshift 0 to 9, now publicly released.

JWST/NIRCam F444W slitless spectroscopy over 0.33 square degrees in COSMOS has yielded a homogeneous catalog of 12,275 emission-line galaxies spanning redshift 0 to 9, now publicly released.

Continuous bitstream diffusion writes images, speech and proteins as binary codes and, with one training and sampling protocol, reaches 3.5% word error rate in joint speech-text generation, while the fold-level agreement of generated protein sequence-structure pairs is reported as 75% in the results section and 69% in the conclusion.

Continuous bitstream diffusion writes images, speech and proteins as binary codes and, with one training and sampling protocol, reaches 3.5% word error rate in joint speech-text generation, while the fold-level agreement of generated protein sequence-structure pairs is reported as 75% in the results section and 69% in the conclusion.

An echocardiography diagnosis model now learns a dedicated token for each quantitative measurement and selects the diagnosis-relevant tokens at inference; average AUC rises from 0.659 to 0.748 on MIMIC-IV-Echo and from 0.730 to 0.778 on an internal cohort.

An echocardiography diagnosis model now learns a dedicated token for each quantitative measurement and selects the diagnosis-relevant tokens at inference; average AUC rises from 0.659 to 0.748 on MIMIC-IV-Echo and from 0.730 to 0.778 on an internal cohort.

Embedding probability measures into a Hilbert space before learning a sparse dictionary turns training from repeated Wasserstein distance solves into one autoencoder optimization, cutting reconstruction loss by roughly two to three orders of magnitude against Wasserstein dictionary learning under the same time budget.

Embedding probability measures into a Hilbert space before learning a sparse dictionary turns training from repeated Wasserstein distance solves into one autoencoder optimization, cutting reconstruction loss by roughly two to three orders of magnitude against Wasserstein dictionary learning under the same time budget.

PanTact lets a Unitree G1 humanoid learn whole-body dexterity from robot-free human demonstrations, using any suitable body surface such as forearms and knees for embracing, pushing, sweeping and load-sharing.

PanTact lets a Unitree G1 humanoid learn whole-body dexterity from robot-free human demonstrations, using any suitable body surface such as forearms and knees for embracing, pushing, sweeping and load-sharing.

The same Fock-space inequality argument, with its auxiliary operators chosen by a semidefinite-program search, raises the certified lower bound on the gap above the Laughlin state to roughly five times the previous constant, with the new certificate checked in exact rational arithmetic.

The same Fock-space inequality argument, with its auxiliary operators chosen by a semidefinite-program search, raises the certified lower bound on the gap above the Laughlin state to roughly five times the previous constant, with the new certificate checked in exact rational arithmetic.

PhySeqNet trains U-Net detectors on semi-synthetic images built from an instrument-matched optical model plus real sample-free noise, so BARseq fluorescent rolonies can be localized without manual annotation, and with barcode-aware non-maxima suppression it recovers a median of 51 counts per cell versus 47 for BarDensr.

PhySeqNet trains U-Net detectors on semi-synthetic images built from an instrument-matched optical model plus real sample-free noise, so BARseq fluorescent rolonies can be localized without manual annotation, and with barcode-aware non-maxima suppression it recovers a median of 51 counts per cell versus 47 for BarDensr.

Feed-forward neural networks trained on state-to-state rate coefficients for 24 linear molecules colliding with para-H2 extend existing datasets to 500 K and 696 cm-1 and give a first estimate within a factor of 3 for CS and CF+, which were excluded from training entirely.

Feed-forward neural networks trained on state-to-state rate coefficients for 24 linear molecules colliding with para-H2 extend existing datasets to 500 K and 696 cm-1 and give a first estimate within a factor of 3 for CS and CF+, which were excluded from training entirely.

Remoscope images 10 microliters of unstained whole blood into 20,000 pictures in about 12 minutes, and a three-class model separated sepsis from non-infectious critical illness and healthy controls in 29 critically ill adults, reaching 100% sensitivity at 85.3% specificity when tuned for rule-out.

Remoscope images 10 microliters of unstained whole blood into 20,000 pictures in about 12 minutes, and a three-class model separated sepsis from non-infectious critical illness and healthy controls in 29 critically ill adults, reaching 100% sensitivity at 85.3% specificity when tuned for rule-out.

A pre-annotation stage that anchors a short axis-guided conversation in a scenario vignette and returns a first-person micro-narrative raised identity resonance and harm sensitivity among 104 annotators labelling 30 ambiguous hate speech items, while label entropy rose and inter-annotator agreement fell.

A pre-annotation stage that anchors a short axis-guided conversation in a scenario vignette and returns a first-person micro-narrative raised identity resonance and harm sensitivity among 104 annotators labelling 30 ambiguous hate speech items, while label entropy rose and inter-annotator agreement fell.

Two pretrained time-series foundation models skip whole Transformer blocks at inference, leaving mean forecasting quality nearly unchanged across 1,088 held-out windows while cutting median A100 batch-1 latency by 11.2% and 13.4% at moderate sparsity.

Two pretrained time-series foundation models skip whole Transformer blocks at inference, leaving mean forecasting quality nearly unchanged across 1,088 held-out windows while cutting median A100 batch-1 latency by 11.2% and 13.4% at moderate sparsity.

An MIT team let a randomly connected network self-organize its learning machinery from simple rules such as strengthening connections that fire together, matching backpropagation-designed networks on image classification.

An MIT team let a randomly connected network self-organize its learning machinery from simple rules such as strengthening connections that fire together, matching backpropagation-designed networks on image classification.

AI in drug discovery is shifting from single-task prediction tools to agents that decompose a goal, call tools and replan on experimental feedback, and a review grades these systems on a six-level autonomy spectrum, placing today's most advanced work at Levels 2 to 4.

AI in drug discovery is shifting from single-task prediction tools to agents that decompose a goal, call tools and replan on experimental feedback, and a review grades these systems on a six-level autonomy spectrum, placing today's most advanced work at Levels 2 to 4.

A sequence-modeling framework written as a nonlinear time-delay system lets the current input's semantic similarity to past inputs set how far back the model looks, reaching 95.35% accuracy on IMDB sentiment classification with 2.6M parameters, above the 89.90% of LSTM and 83.00% of Mamba in the same table.

A sequence-modeling framework written as a nonlinear time-delay system lets the current input's semantic similarity to past inputs set how far back the model looks, reaching 95.35% accuracy on IMDB sentiment classification with 2.6M parameters, above the 89.90% of LSTM and 83.00% of Mamba in the same table.

LLM pruning now builds its calibration activations from the dense model's own autoregressive decode steps while discarding prefill, aligning the pruning objective with generation-time activations; it beats fixed-text calibration on WritingBench and ClassEval and reaches up to 1.48x end-to-end decoding speedup on A100 GPUs.

LLM pruning now builds its calibration activations from the dense model's own autoregressive decode steps while discarding prefill, aligning the pruning objective with generation-time activations; it beats fixed-text calibration on WritingBench and ClassEval and reaches up to 1.48x end-to-end decoding speedup on A100 GPUs.

On road-network trajectory generation in Porto, Beijing and San Francisco, a model that previews a short future route before choosing the current intersection turn reaches a Porto distance-distribution JSD of 1.01, below HOSER's 1.57.

On road-network trajectory generation in Porto, Beijing and San Francisco, a model that previews a short future route before choosing the current intersection turn reaches a Porto distance-distribution JSD of 1.01, below HOSER's 1.57.

A multistatic terahertz link treats ordinary communication traffic scattered off an object as the observation, identifying shape at orientations never seen in training, with real gains from more transmitters and accuracy that holds when three quarters of the surface is hidden.

A multistatic terahertz link treats ordinary communication traffic scattered off an object as the observation, identifying shape at orientations never seen in training, with real gains from more transmitters and accuracy that holds when three quarters of the surface is hidden.

Total burn surface area is now regressed directly from frontal and dorsal whole-body and burn-region binary masks, cutting error below the direct pixel-area ratio baseline on a synthetic set of 5,000 human models and 320,000 rendered views.

Total burn surface area is now regressed directly from frontal and dorsal whole-body and burn-region binary masks, cutting error below the direct pixel-area ratio baseline on a synthetic set of 5,000 human models and 320,000 rendered views.

Microsoft plans to scale its Azure Maia family to multi-gigawatt deployments starting in 2027 and make Maia the primary inference system for its products and services, with the third-generation Maia 300 expected to deliver 40% more tokens per megawatt than leading GPU-based systems.

Microsoft plans to scale its Azure Maia family to multi-gigawatt deployments starting in 2027 and make Maia the primary inference system for its products and services, with the third-generation Maia 300 expected to deliver 40% more tokens per megawatt than leading GPU-based systems.

For PFAS substitution, a three-stage multi-agent workflow links requirement analysis, candidate generation and manufacturability assessment in a closed loop, raising the overall score from 2.99 to 5.47 with the backbone model held fixed.

For PFAS substitution, a three-stage multi-agent workflow links requirement analysis, candidate generation and manufacturability assessment in a closed loop, raising the overall score from 2.99 to 5.47 with the backbone model held fixed.

A hybrid intrusion-detection framework that chains causal graph learning, entropy-guided compression, attention fusion, adversarial calibration and trust-weighted federated aggregation into one pipeline reaches 97.4% detection accuracy on CICIDS2017 while cutting the input from 78 features to 18.

A hybrid intrusion-detection framework that chains causal graph learning, entropy-guided compression, attention fusion, adversarial calibration and trust-weighted federated aggregation into one pipeline reaches 97.4% detection accuracy on CICIDS2017 while cutting the input from 78 features to 18.

NVIDIA says RTX Spark laptops arrive Oct. 16 with a Nemotron 3.5 Bolt 74B model quantized to 2 bits and packed into 24GB, letting local agents run on a single machine.

NVIDIA says RTX Spark laptops arrive Oct. 16 with a Nemotron 3.5 Bolt 74B model quantized to 2 bits and packed into 24GB, letting local agents run on a single machine.

Cloudflare's Agents SDK now ships browserTool, giving an agent a browser it can return to across turns with tabs, cookies and login sessions intact, kept alive even while the Durable Object hibernates.

Cloudflare's Agents SDK now ships browserTool, giving an agent a browser it can return to across turns with tabs, cookies and login sessions intact, kept alive even while the Durable Object hibernates.

Distributed LLMs can treat bias as a network state that evolves with communication: nodes exchange a three-dimensional bias state and align by consensus, cutting composite bias by about 48.5% and inter-node bias-state variance by about 87.4% versus FedAvg in a lightweight Transformer pilot.

Distributed LLMs can treat bias as a network state that evolves with communication: nodes exchange a three-dimensional bias state and align by consensus, cutting composite bias by about 48.5% and inter-node bias-state variance by about 87.4% versus FedAvg in a lightweight Transformer pilot.

LangSmith has launched AI Search, which retrieves agent traces from a natural-language description of a behavior, such as whether the user got frustrated or whether PII leaked, instead of requiring a preconfigured evaluator or manual review.

LangSmith has launched AI Search, which retrieves agent traces from a natural-language description of a behavior, such as whether the user got frustrated or whether PII leaked, instead of requiring a preconfigured evaluator or manual review.

AMD introduced the Helion agentic AI rack blueprint, letting customers match CPU core density and power to agentic workloads, with a CPU-only configuration supporting up to 32,768 cores for hyperscale deployments.

AMD introduced the Helion agentic AI rack blueprint, letting customers match CPU core density and power to agentic workloads, with a CPU-only configuration supporting up to 32,768 cores for hyperscale deployments.

vercel-labs released skills, a CLI that installs agent skills into Claude Code, Codex, Cursor and 75 more coding agents with one command, and can emit a skill's prompt directly.

vercel-labs released skills, a CLI that installs agent skills into Claude Code, Codex, Cursor and 75 more coding agents with one command, and can emit a skill's prompt directly.