Skip to content
AI News HubLIVE

This edition’s highlights

Models

Full Fine-Tuning (FFT) on SageMaker AI Serverless Model Customization

  • FFT updates all model parameters, overcoming LoRA's low-rank representational limits and excelling on complex tasks with large datasets
  • Experiments show FFT achieves 90.0% on LLM-as-Judge semantic correctness for code generation at 100K+ training examples, leading LoRA across all quality dimensions
AWS Machine Learning BlogIn-site articleFull Fine-Tuning (FFT) on SageMaker AI Serverless Model Customization

Introducing the Amazon SageMaker JumpStart Model Leaderboard

  • The leaderboard covers 70+ models from providers including Meta, Google, Qwen, IBM, and Arcee, showing quality, speed, and cost directly in SageMaker Studio
  • AWS runs BIG-Bench Hard, GPQA Diamond, and MMLU-Pro with a unified evaluation harness so scores are comparable across models
AWS Machine Learning BlogIn-site articleIntroducing the Amazon SageMaker JumpStart Model Leaderboard

MedHarness: A Unified Budgeted Runtime for Reliable Medical Imaging Agents

  • Reframes post-answer control as shared decision-object attachment for a unified cross-modality controller
  • Cuts wrong accepts by ~47% and raises accepted accuracy to 83.2% on a 512-case live pool
arXiv Computer VisionIn-site articleMedHarness: A Unified Budgeted Runtime for Reliable Medical Imaging Agents

Reference-Guided Data Curation for 3D Medical Vision-Language Pretraining

  • Proposes an offline, training-free reference-guided curation recipe that extracts per-modality quality signals from a small reference set and rank-aggregates them for sample selection
  • On 3D CT, matches Merlin's zero-shot disease classification with 3.3× less training data and a 5k-sample reference set
arXiv Computer VisionIn-site articleReference-Guided Data Curation for 3D Medical Vision-Language Pretraining

llm-gemini 0.38

  • llm-gemini 0.38 adds support for Gemini 4 Neon
  • Google's new lineup graphic omits the Argon model entirely
Simon Willison's WeblogIn-site articlellm-gemini 0.38
Policy

Learning the Dynamics of Deformable Linear Objects Through Coupled Oscillators for Bimanual Robotic Manipulation

  • Φ-CON combines a floating-base coupled oscillator network with a learned affine diffeomorphism, yielding shape dynamics that are stable by construction.
  • The model's Lagrangian structure enables a PD controller that compensates learned potential forces with closed-loop stability guarantees.
arXiv RoboticsIn-site articleLearning the Dynamics of Deformable Linear Objects Through Coupled Oscillators for Bimanual Robotic Manipulation

Biologically-Inspired Spike Encoding and Optimization for Spiking Neural Networks

  • Difference-of-Gaussians and DVS-style delta modulation encoders produce much sparser spike trains than rate coding.
  • Rouser's learnable per-neuron thresholds lift Fashion-MNIST accuracy from 87.37% to 88.13% via homeostatic firing-rate regulation.
arXiv Machine LearningIn-site articleBiologically-Inspired Spike Encoding and Optimization for Spiking Neural Networks
Robotics

RESCUE Jumper: A Centimeter-Scale Rolling and Jumping Robot for Search and Rescue

  • Dual-actuator design reuses wheel motors: forward motion rolls, reverse motion winds a spring steel tail for jumping
  • Jump direction varies with tail winding speed, transitioning from backward to forward jumps between 0.75 and 8.73 rad/s
arXiv RoboticsIn-site articleRESCUE Jumper: A Centimeter-Scale Rolling and Jumping Robot for Search and Rescue
Research

datasette-arrows 0.1a0

  • Datasette 1.0a43 lets plugins implement streaming export formats via register_output_renderer().
  • The plugin was suggested and almost single-shot built by Claude Fable 5.2.
Simon Willison's WeblogIn-site articledatasette-arrows 0.1a0

You have reached the end of this edition.

Your reading list
Other updates (53)
Agents

LangSmith Tool Gateway: Secure, governed tool use for agents

LangSmith Tool Gateway, now in Public Beta, handles credentials, identity, and access for every agent tool call, with policies, approvals, and traces.

LangChain BlogIn-site articleLangSmith Tool Gateway: Secure, governed tool use for agents

Introducing CustomerLake Conversions: Improve advertising performance with first-party customer context

Databricks introduces CustomerLake Conversions, a new capability in its Agentic CDP that connects first-party customer data and ML models directly to advertising activation and measurement. Marketers can define value-based conversion signals, such as a first purchase from a customer with high predicted lifetime value, and send them to ad platforms via CAPI integrations, with campaign insights returning to Databricks through OpenSharing and Lakeflow Connect.

Databricks BlogIn-site articleIntroducing CustomerLake Conversions: Improve advertising performance with first-party customer context

Tencent Hy4 is live on Together AI

Tencent's flagship Hy4 model is now available on Together AI with text and image input, a 1M-token context window, and pricing at $0.83 per 1M input tokens and $2.50 per 1M output tokens. Together AI tested its chart extraction and single-prompt browser game generation.

Together AI BlogIn-site articleTencent Hy4 is live on Together AI

How Abundant designed efficient rollout architecture with Modal Sandboxes

Abundant builds RL environments for frontier labs like OpenAI, Anthropic, Google, and Meta. It has created over 75,000 rollouts, each running in a Modal Sandbox, with Modal Functions powering its scheduler API.

Modal BlogIn-site articleHow Abundant designed efficient rollout architecture with Modal Sandboxes

Improving Developer Experience in the Age of Agents

Weaviate found that coding agents often receive incomplete documentation, leading to errors. They propose optimizing both developer experience (DX) and agent experience (AX).

Weaviate BlogIn-site articleImproving Developer Experience in the Age of Agents

Introducing workload recovery and expanded infrastructure checks on Amazon SageMaker HyperPod

Amazon SageMaker HyperPod adds workload recovery and expanded infrastructure checks—MultiNodeTrainingValidation, Kubernetes Resource Monitoring Agent, and hung job detection—to catch silent data corruption, hung jobs, and cluster-level failures, improving training goodput.

AWS Machine Learning BlogIn-site articleIntroducing workload recovery and expanded infrastructure checks on Amazon SageMaker HyperPod

How Airbnb uses cost per task to operate coding agents on Amazon Bedrock

Airbnb runs evals on real engineering tasks and uses Amazon Bedrock observability to choose models for its coding agents and control production costs. This post shows how cost per task guides model selection, why production spending can differ from eval results, and how to monitor latency, manage quotas, and attribute spend.

AWS Machine Learning BlogIn-site articleHow Airbnb uses cost per task to operate coding agents on Amazon Bedrock

Introducing Genie Migrate for Serverless

Databricks launches Genie Migrate for Serverless (Beta), an AI agent that automates the migration of classic workloads to serverless compute by scanning, translating, and applying changes reversibly, with reviewable diffs and a 30-day rollback.

Databricks BlogIn-site articleIntroducing Genie Migrate for Serverless

Private certificate authority support in Amazon Bedrock AgentCore Identity

Amazon Bedrock AgentCore Identity now supports private certificate authorities (CAs), so your agents can connect to self-hosted identity providers that use a private CA. This post shows how to issue an IdP certificate with AWS Private CA, configure a credential provider to trust it, rotate the CA, and protect the certificate.

AWS Machine Learning BlogIn-site articlePrivate certificate authority support in Amazon Bedrock AgentCore Identity

Connect Amazon Bedrock AgentCore Gateway to private APIs in a hub-and-spoke network with centralized inspection

This post presents a tested pattern for connecting Amazon Bedrock AgentCore Gateway to private APIs and MCP servers in enterprises where a central network team owns all VPCs and a central firewall inspects all inter-VPC traffic bidirectionally. The pattern places the VPC Lattice resource gateway in the network account, uses self-managed mode, and shares resource configurations via AWS RAM.

AWS Machine Learning BlogIn-site articleConnect Amazon Bedrock AgentCore Gateway to private APIs in a hub-and-spoke network with centralized inspection

How Stretto built an AI voice agent with Amazon Nova 2 Sonic

Stretto replaced its web-form lead capture with Tara, an AI voice agent built on Amazon Nova 2 Sonic that conducts bankruptcy intake interviews by phone and qualifies leads. See how Stretto, the AWS Generative AI Innovation Center, and CI&T designed the voice architecture, validated it for sensitive conversations, and moved it to production.

AWS Machine Learning BlogIn-site articleHow Stretto built an AI voice agent with Amazon Nova 2 Sonic

From reporting to reasoning: Getting enterprise data ready for AI agents

A recap of the inaugural AWS Quick User Group DC meetup, where data leaders explored getting enterprise data ready for AI agents. It covers what's new in AWS Quick, Financial Independence Group's AI analytics journey, and hands-on app and no-code flow building.

AWS Machine Learning BlogIn-site articleFrom reporting to reasoning: Getting enterprise data ready for AI agents

What's new for Genie in Microsoft: Copilot Cowork and Chat, Microsoft Teams, and Excel

Databricks announces deep integration of Genie across Microsoft tools, including general availability in Teams, plugins for Copilot Cowork and Chat, an Excel add-in, and a Copilot Studio connector, all powered by Genie Ontology and Unity Catalog for governed insights.

Databricks BlogIn-site articleWhat's new for Genie in Microsoft: Copilot Cowork and Chat, Microsoft Teams, and Excel

Your Data Was Never Just Documents

Unstructured introduces a Multimedia Partitioner that turns audio and video into source-timed JSON elements, enabling AI agents to use speech and visuals from recordings. The article details its element ontology, diarization, scene descriptions, and model architecture.

Unstructured BlogIn-site articleYour Data Was Never Just Documents
Chips

Season’s Streamings and Savings: Black Friday Deals Come to GeForce NOW

GeForce NOW launches Black Friday promotions with 30% off the first six months of Performance and Ultimate memberships through Nov. 30. The service also expands to Amazon Fire TV Stick HD, 4K Select, and 4K (2026 model) with a 90% off first-month Amazon offer, and adds nine games including five Bethesda titles via GOG.

NVIDIA BlogIn-site articleSeason’s Streamings and Savings: Black Friday Deals Come to GeForce NOW

Introducing Together Storage Accelerator: fast, free data caching for your GPU cluster

Together AI launches Together Storage Accelerator (TSA), a distributed caching layer for Together GPU Clusters that delivers about 20x faster warm reads, one origin fetch per cluster, no application changes, and is included free.

Together AI BlogIn-site articleIntroducing Together Storage Accelerator: fast, free data caching for your GPU cluster
Robotics

How OpenAI’s Performance Factory keeps ChatGPT fast

OpenAI details its Codex-powered Performance Factory, an autonomous system that automates the discovery and resolution of performance regressions to keep ChatGPT fast.

OpenAI NewsIn-site articleHow OpenAI’s Performance Factory keeps ChatGPT fast

LuxTac: A Flexible, Scalable, and Open-Source Optical Taxel Array with Distributed Three-Axis Force Sensing for Tactile Robot Learning

LuxTac is a 3.5 mm thick optical taxel array built from off-the-shelf parts that, to the authors' knowledge, is the first to resolve shear at individual taxels. Normal sensing is linear (R² = 0.999) from 104 mN to at least 5.8 N, and three-axis force estimates achieve summed R² of 0.92–0.96. Adding LuxTac pressure readings to an Action Chunking Transformer policy substantially improves peg-in-hole success over a vision-only baseline.

arXiv RoboticsIn-site articleLuxTac: A Flexible, Scalable, and Open-Source Optical Taxel Array with Distributed Three-Axis Force Sensing for Tactile Robot Learning

Panini: A Feasible Grammar for Robot Designs

The paper introduces a typed production grammar that uses case-role relations, parameter inheritance, and specificity-based rule arbitration to generate both manipulation skills and robot morphologies, plus PANINI, a cross-entropy search over grammar-based morphology representations. Its demonstrated advantage is structural — legal-by-construction, geometrically coordinated designs — rather than superior optimization performance.

arXiv RoboticsIn-site articlePanini: A Feasible Grammar for Robot Designs
Research

How Uncertain Is My Action? Self-Correcting Visual Planning with Frozen World Models

This paper shows that a frozen latent world model can be run in reverse to detect and compensate for actuation shifts. By predicting what should have been observed under a hypothesized shift and comparing it with actual observations, the method quantifies shifts, estimates them with Cramer-Rao error bars, and corrects the planner. Experiments on simulated tasks and a real robotic arm demonstrate significant improvements in success rates and precision.

arXiv RoboticsIn-site articleHow Uncertain Is My Action? Self-Correcting Visual Planning with Frozen World Models

Learning and Control in Physically Coupled Dynamical Systems

This paper introduces TAG (Testbed for Autonomous Games), a low-cost robotic platform for studying learning and control in physically coupled dynamical systems. TAG combines an actively tilted and steered ball-in-maze board with a top-mounted rotary inverted pendulum, where the two subsystems disturb each other. The authors present complementary reinforcement learning paradigms: model-based for the maze and model-free for the pendulum. A key finding is that introducing the pendulum causes an immediate sharp drop in the maze policy's success rate, but the policy adapts online and recovers to pre-coupling performance within a small number of episodes. Experiments show both independently trained policies successfully control their subsystems despite physical coupling.

arXiv RoboticsIn-site articleLearning and Control in Physically Coupled Dynamical Systems

Confidence-Based Materials Data Quality Assessment via Domain Knowledge and Graph-Structured Learning

This paper proposes a confidence-based framework for materials data quality assessment that integrates domain knowledge with graph-structured learning, formulating quality as continuous confidence estimation and demonstrating significant improvements in model performance through rule calibration and graph refinement on multiple benchmarks.

arXiv Machine LearningIn-site articleConfidence-Based Materials Data Quality Assessment via Domain Knowledge and Graph-Structured Learning

BiCAT: Bimanual Coordination Action Tokenization

BiCAT is a new action tokenization method for bimanual manipulation that shares one motion representation across arms, conditions each arm's tokens on its partner via symmetric cross-arm attention and a joint decoder, and predicts each arm's tokens in parallel. Across ten RoboTwin 2.0 tasks and six real-world tasks on three embodiments, it achieves the highest success, improving mean simulation success by 17.6 percentage points over the strongest of eight baselines while requiring only 40.5 ms per action chunk.

arXiv RoboticsIn-site articleBiCAT: Bimanual Coordination Action Tokenization

Exploring Parkinson's Diagnosis Using Machine Learning Techniques

A study on a 58-person dataset evaluated several machine learning algorithms for Parkinson's disease diagnosis, finding decision tree and logistic regression most accurate (72-76%), all models achieving high sensitivity (up to 97.56%) for positive cases, but poor performance in distinguishing healthy controls.

arXiv Machine LearningIn-site articleExploring Parkinson's Diagnosis Using Machine Learning Techniques

Budgeted Proper Loss Design for Uncertain Decision Thresholds

This paper studies the choice of a proper loss when the threshold distribution belongs to a finite specified family. Under integrable loss weights with a positive pointwise floor and fixed total mass, the authors derive the weight that minimizes the sharp, distribution-free multiplier from surrogate regret to threshold-averaged decision regret. The solution clips a rescaled upper envelope of the threshold densities at the prescribed floor, with normalization found by a one-dimensional root search. Proofs of sharpness, optimality, and stability under density misspecification are given, along with a limitation for atomic threshold laws. An exact histogram implementation and synthetic experiments verify the construction.

arXiv Machine LearningIn-site articleBudgeted Proper Loss Design for Uncertain Decision Thresholds

mmTherm: Absolute 3D Human Pose Estimation from Low-Resolution Thermal Images and mmWave Radar

mmTherm introduces a lightweight framework that explicitly combines body structure and joint direction cues from low-resolution thermal imaging with range measurements from mmWave radar for absolute 3D human pose estimation. On OctoNet, it reduces cross-user absolute joint error by 33% and root localization error by 38% while requiring substantially fewer parameters and less inference computation.

arXiv Computer VisionIn-site articlemmTherm: Absolute 3D Human Pose Estimation from Low-Resolution Thermal Images and mmWave Radar

Chebyshev Polynomial Zonotope Abstractions for Certified Reachability of Learning-Enabled Systems

This paper proposes Chebyshev polynomial zonotope abstractions to replace Taylor expansions for certified reachability of learning-enabled nonlinear systems. The authors prove that Taylor expansions have a set-width cap due to complex-plane singularities, while Chebyshev abstractions have no such cap and gain width geometrically with degree. Experiments on closed-loop control and MNIST classifiers show certificates survive where Taylor ones diverge, flipping specifications from unknown to proven.

arXiv Machine LearningIn-site articleChebyshev Polynomial Zonotope Abstractions for Certified Reachability of Learning-Enabled Systems

NIRJALA: Draining the Sea out of Gaussian Splatting with Three-Path Optics and Reversible Physics

NIRJALA introduces a three-path underwater image formation model with a range-saturating forward-scattering term and a reversible physics anchor, fixing unbounded inversion gain in Gaussian-splatting water removal and adding a 30-scene path-traced benchmark.

arXiv Computer VisionIn-site articleNIRJALA: Draining the Sea out of Gaussian Splatting with Three-Path Optics and Reversible Physics

Response-Relevant Information Is Not Sufficient for Neural Training-State Equivalence

A new arXiv paper defines training-state equivalence via intervention response and finds that fifteen descriptors, despite carrying predictive information beyond weight norm and loss, show a 16.1% advantage that cannot be separated from a conditional placebo, suggesting decodable information need not form the pairwise geometry required for similarity judgments.

arXiv Machine LearningIn-site articleResponse-Relevant Information Is Not Sufficient for Neural Training-State Equivalence

Where Global Context and Weight Sharing Work in Compact Vision Backbones

This paper investigates where global self-attention should be inserted and whether repeated blocks can share weights in compact vision backbones like EfficientNet-B0. It finds that zero-gated attention adapters at stages P3–P5 improve ImageNet-1k accuracy by 0.954 percentage points, while alternating two shared block templates in Stage 6 cuts parameters by 10.9% with only a 1.84-point drop.

arXiv Computer VisionIn-site articleWhere Global Context and Weight Sharing Work in Compact Vision Backbones

Goemans's Unsplittable-Flow Conjecture: The Anatomy of AI-Assisted Counterexample Discovery

A new paper analyzes the seven-vertex counterexample to Goemans's unsplittable-flow conjecture found by Dmitry Rybin via ChatGPT, and develops a forward construction pipeline that characterizes all counterexamples for a given topology. The pipeline, based on adding shortcut arcs to an arborescence and using a conflict polytope, is shown to be transferable: ChatGPT produced counterexamples on 93 other arborescence shapes, explaining why the original example was hard to find by blind enumeration.

arXiv Machine LearningIn-site articleGoemans's Unsplittable-Flow Conjecture: The Anatomy of AI-Assisted Counterexample Discovery

GridForge: An Open Physics-Informed Digital Twin and Benchmark Suite for UK 11 kV Underground Distribution Cables

GridForge is an open-source physics-informed digital twin and benchmark suite released on arXiv for fault diagnosis and prediction in UK 11 kV underground distribution cables, combining IEC-standard thermal models, a calibrated synthetic dataset of 64 cables (320 cable-years), a PINN surrogate with 0.148 °C RMSE on a held-out set, and a five-task benchmark with three reference baselines; only the virtual-sensor task currently discriminates between baselines.

arXiv Machine LearningIn-site articleGridForge: An Open Physics-Informed Digital Twin and Benchmark Suite for UK 11 kV Underground Distribution Cables

Autoregressive View Model

The paper introduces the Autoregressive View Model (AVM), which generates scenes as a stream of views without explicit 3D representation, using a unified in-context interface and trained on approximately 8M scenes and video clips, evaluated across benchmarks for camera control, text control, visual quality, and long-horizon consistency.

arXiv Computer VisionIn-site articleAutoregressive View Model

Neural Network-Based Modeling of Hysteretic Systems: From Prandtl and Preisach Neural Networks to Neuro-Madelung and Neuroplastic Continuous Models

This arXiv review traces the development of Prandtl neural networks, Preisach neural networks, generalized Prandtl neural networks, the neuro-Madelung model, extended Preisach neural networks, and dynamic plastic continuous neural networks, focusing on how their architectures, activation functions, and training algorithms encode hysteresis memory, degradation, non-congruency, asymmetry, and rate dependence.

arXiv Machine LearningIn-site articleNeural Network-Based Modeling of Hysteretic Systems: From Prandtl and Preisach Neural Networks to Neuro-Madelung and Neuroplastic Continuous Models

OCLDA-GS: Object-Centric Local Deformation-Aware Gaussian Streaming

The paper introduces OCLDA-GS, an object-centric streaming 3D Gaussian Splatting framework that dynamically reallocates object- and motion-aware anchors, restricts Gaussian neighborhoods within object groups for sharper boundaries, and combines object-wise rigid transformation estimates with per-anchor local deformation correction. On three dynamic multi-view datasets it improves PSNR over Instant Gaussian Stream under matched refinement budgets and reduces quality drops between refinement timesteps.

arXiv Computer VisionIn-site articleOCLDA-GS: Object-Centric Local Deformation-Aware Gaussian Streaming

How Social Department built culture-aware video discovery

Social Department, an AI operating system for entertainment marketing, worked with TrackIt and AWS to build QuickSearch: single-action queries that surface shareable scenes, quotable lines, memeable frames, and trend-aligned moments across large video libraries, using TwelveLabs models on Amazon Bedrock and Amazon OpenSearch Serverless.

AWS Machine Learning BlogIn-site articleHow Social Department built culture-aware video discovery

pg_pinecone now deploys on Railway in one click | Pinecone

Pinecone released a one-click Railway template that ships Postgres 18 with the pg_pinecone extension installed, preloaded, and enabled. Pinecone says pg_pinecone has lower latency and smaller memory use than pgvector at the same recall, with a benchmark index roughly 15 times smaller than a pgvector HNSW index using halfvec on 1 million vectors.

Pinecone BlogIn-site articlepg_pinecone now deploys on Railway in one click | Pinecone
Policy

Recoverability, Not Invariance: Certified Safety for Robots at the Limits of Performance

The paper introduces recoverability, a contraction-theoretic safety concept that certifies a robot's ability to reengage a trusted recovery mode instead of requiring forward invariance of the deployed controller, enabling formal safety guarantees for aggressive, learned, and experimental behaviors.

arXiv RoboticsIn-site articleRecoverability, Not Invariance: Certified Safety for Robots at the Limits of Performance

Databricks Wins A-Team Award for AI in Capital Markets

Databricks was named Best AI Solution for Data Intelligence & Governance at the 2026 A-Team AI in Capital Markets Awards, highlighting its unified platform for AI and governance.

Databricks BlogIn-site articleDatabricks Wins A-Team Award for AI in Capital Markets

Deploying Daybreak Red

OpenAI publishes a guide on setting up an isolated environment for supervised security investigations with Daybreak Red, covering architecture, safety controls, configuration, and monitoring.

OpenAI NewsIn-site articleDeploying Daybreak Red

Design responsible AI controls for Amazon Aurora PostgreSQL

Learn how to make RAG chatbot answers traceable with Amazon Aurora PostgreSQL, pgvector, and Amazon Bedrock Guardrails, logging sources, scores, and guardrail traces in SQL.

AWS Machine Learning BlogIn-site articleDesign responsible AI controls for Amazon Aurora PostgreSQL
Models

SPICE: Stochastic Policy for Interaction in Cluttered Environments

SPICE is a tactile-conditioned diffusion policy that enables robots to reach targets in dense plant canopies using only tactile feedback, handling uncertain contact and visual occlusion. In over 180 real-robot trials, it achieved 95% success in branch manoeuvring and 87.5% zero-shot transfer on unseen tomato and lime plants.

arXiv RoboticsIn-site articleSPICE: Stochastic Policy for Interaction in Cluttered Environments

RageVLA: Augmentation-Trained Perception and Action for State-of-the-Art Robustness on LIBERO

RageVLA argues that VLA models saturate on LIBERO yet collapse on LIBERO-PRO because what to do stays tied to training-time object positions, surrounding scenes, and instruction wording. The authors cut these ties in the data (general rules that move and rotate every object and fixture, port tasks to other environments and surfaces, and paraphrase every instruction) and in the policy (a perception engine on frozen DINOv3 features hands the action model a decoupled view: a list of objects with positions and a compact gripper-focus code alongside TinyCLIP image and instruction embeddings, with a from-scratch flow-matching diffusion transformer acting on it). The 0.21B-parameter model scores 0.72 on LIBERO-PRO versus 0.53 for the best published model, with a larger variant at 0.73. LIBERO-Ob…

arXiv RoboticsIn-site articleRageVLA: Augmentation-Trained Perception and Action for State-of-the-Art Robustness on LIBERO

Optimize the Cob, Then the Kernel: Optimizing Beam Search for Generative Recommendation Inference

An arXiv paper argues that general-purpose LLM serving engines such as vLLM waste effort on generative recommendation inference, where beams redundantly re-read the shared prompt KV cache and each beam-search step is a separate engine invocation. Using Tree Attention plus compiler- and scheduling-level optimizations, the authors cut end-to-end latency by up to 25x over vLLM on Ascend NPUs without custom kernels, matching a state-of-the-art custom-kernel solution.

arXiv Machine LearningIn-site articleOptimize the Cob, Then the Kernel: Optimizing Beam Search for Generative Recommendation Inference

GRID-UAV: Geometry and Ray-depth Integrated Decision-making for Unmanned Aerial Vehicles

GRID-UAV combines explicit depth lifting with implicit geometric alignment to improve vision-language-action models for UAV navigation. By converting monocular depth into ray-corrected 3D positions and aligning features with a multi-view geometry model during training, GRID achieves state-of-the-art success rates on the TravelUAV dataset.

arXiv Computer VisionIn-site articleGRID-UAV: Geometry and Ray-depth Integrated Decision-making for Unmanned Aerial Vehicles

TransFIRA+: Transfer Learning for Biometric Recognizability Assessment

TransFIRA introduces a lightweight framework that grounds biometric recognizability directly in a pretrained encoder's embedding space, eliminating the need for quality annotations; TransFIRA+ extends it to cross-camera multimodal person identification in untrimmed surveillance video, where a post-hoc sigmoid calibration handles saturated body recognizability scores.

arXiv Computer VisionIn-site articleTransFIRA+: Transfer Learning for Biometric Recognizability Assessment

Bits, Not Boxes: Label-Free Layer Selection for Visual Localization in VLMs

WhereBits identifies which layers in vision-language models localize queried objects by measuring the image-independent component of attention in bits on unlabeled questions, eliminating the need for box annotations. Across eight VLMs, it matches box-tuned pointing accuracy and improves V* accuracy by 6.0 points in training-free zoom-in.

arXiv Computer VisionIn-site articleBits, Not Boxes: Label-Free Layer Selection for Visual Localization in VLMs

Exa boosts embedding throughput by 50% to index billions of documents on AWS AI chips

Exa's neural search engine is backed by an in-house vector database, so the unit economics of embedding tens of billions of web documents drive the business. This post shares how Exa optimized embedding inference on AWS Neuron hardware (Inferentia2 and Trainium2) with custom NKI kernels, reaching a 47% gain in performance and cost per million documents.

AWS Machine Learning BlogIn-site articleExa boosts embedding throughput by 50% to index billions of documents on AWS AI chips

Introducing Unity Gateway Insights

Databricks introduces Unity Gateway Insights, which uses Genie to automatically analyze millions of agent traces alongside business data and ontology to find wasted tokens. It surfaces recurring MCP errors, simple tasks running on expensive frontier models, and repeated workflows worth turning into shared skills, all under existing data governance.

Databricks BlogIn-site articleIntroducing Unity Gateway Insights

How Salesforce is advancing global-scale model serving with cross-Region inference routing in Amazon SageMaker AI

Salesforce uses SageMaker AI's cross-Region inference router to securely add capacity across 30+ Regions while keeping data residency commitments, and gains per-request visibility into where and why each call was served.

AWS Machine Learning BlogIn-site articleHow Salesforce is advancing global-scale model serving with cross-Region inference routing in Amazon SageMaker AI

Announcing Data Annotation Queues

LangSmith introduces Data Annotation Queues to help developers review logs, annotate data, and build datasets for improving LLM applications from prototype to production.

LangChain BlogIn-site articleAnnouncing Data Annotation Queues

Catch production failures early with LangSmith Alerts

LangSmith introduces Alerts for real-time notifications on error rates, run latency, and feedback scores, helping teams spot failures in LLM apps and agents before customers do.

LangChain BlogIn-site articleCatch production failures early with LangSmith Alerts
Tools

The Gemini app is now available on lower-memory Android devices.

Google expands the Gemini app to lower-memory Android (Go edition) devices, bringing Gemini Live voice conversations, camera-based learning, and custom image generation to hundreds of millions more users.

Google AI BlogIn-site articleThe Gemini app is now available on lower-memory Android devices.