10 selected stories for 2026-11-11, grouped by topic.
Edition details
Dates use UTC+8. The latest edition uses a day with at least 5 completed reports when available; until then an earlier edition may appear. There is no fixed delivery time.
Genie plugins work natively in ChatGPT, Claude, Microsoft Copilot Cowork, and Grok Bot; other agents can access Genie via the MCP server.
Genie Ontology provides a governed context layer with certified definitions, relationships, and source context to help agents interpret business terms correctly.
Introduces a unified benchmark with aligned data splits and explicit task contracts across 56 datasets, 46 methods, and 20 capability axes.
Rank comparisons link imputation and prediction via numerical recovery and dependence preservation, and connect generation to both through predictive utility, while distributional fidelity follows different rankings.
The Universal Agent Evaluation Framework (UAEF) is a framework-agnostic system for evaluating AI agents. UAEF converts each framework's traces into one canonical format, scores them with 36 built-in metrics across 7 dimensions, and stores results so teams can compare agent versions and catch regressions.
AWS Marketplace now supports usage-based metering for Amazon SageMaker AI bidirectional streaming inference, allowing model providers to charge for actual consumption during live streams. Deepgram is the launch partner, bringing usage-based speech-to-text and text-to-speech billing to AWS Marketplace.
Databricks, in collaboration with Turing, built a shareable enterprise data-agent benchmark comparing Genie with two frontier agents in a live Databricks workspace. Genie passed 91% of runs versus 45% and 58%, while using less than half the tokens and finishing 3–4× faster, thanks to business context extracted from notebooks, dashboards, and saved SQL by Genie Ontology.
Databricks adds unstructured data analysis, MCP connectors, and cross-organization sharing to Genie Agents, enabling governed agents to reason across tables, documents, and tools and take action.
Databricks announces the general availability of Genie Ontology, an agentic ontology built by agents and governed by people to give every AI agent a shared, governed understanding of the business. The release adds Genie Search, external BI semantic model ingestion, organizational directories, and Ontology Manager. In a benchmark built with Turing, Genie with Genie Ontology passed 91% of task runs, 1.6× the strongest frontier agent, while using less than half the tokens.
DOBB introduces an online geometric representation that maps state estimates, uncertainty, and agent-specific characteristics into a dynamic, direction-dependent safety margin, achieving a mean total failure rate of 2.6% across four environments and keeping collision-based failure at or below 2.3% under out-of-distribution shifts.
ASPIRE is a framework in which a large language model agent selects and executes nonlinear system identification methods, producing closed-form models with parameters, estimation records, interpretations, and error evaluated on data the agent cannot access. All 30 runs across three measured systems yielded valid models at 0.96 to 3.38 USD per run; an author assessment found 132 of 150 items fully satisfied, with model-structure justification the most frequent weakness.
Robo-MLT is a dual-system robotic framework that couples explicit workflow monitoring with learned action generation to improve reliability and recoverability in long-horizon medical laboratory tasks. It achieves 82% and 70% success rates on blood gas analysis and blood sample tube sorting, significantly outperforming baselines, and can replan and continue after human intervention.
A study demonstrates autonomous synchrotron X-ray experiments on nanoparticle superlattices using coding-agent infrastructure, with full transcripts enabling audit of decision provenance and cognitive labor division.
Kimi K3.1 Pro Highspeed from Moonshot AI is now available on Amazon Bedrock with a 1-million-token context window, 2–3x faster output, OpenAI-compatible APIs, Flex tier pricing, explicit prompt caching, and OpenCode integration for coding assistance.
JioStar Chief Architect Vijay Seshadri shares lessons on building an AI discovery assistant for 500M users, covering evals, guardrails, and taking AI to production.
This post shows how to build a daily dependency watcher as a LangGraph graph on the Amazon Bedrock AgentCore Web Search API. It researches three package ecosystems in parallel, keeps rules like date window, blocklist, and search budget in code, and checks each finding against its cited sources.
This post explains how to evaluate production agents on Amazon Bedrock AgentCore Evaluations by starting from business outcomes, breaking them into observable behaviors, and mapping each behavior to built-in, DeepEval, AutoEval, or custom evaluators across on-demand, online, and batch modes.
Gemini Intelligence is rolling out to more devices including motorola signature 27 and Pixel 10, with new features like Create My Widget and Device Assistance, plus expanded availability of Rambler and task automation.
Databricks announces that Unity Catalog metric views can now be published directly to Power BI semantic models, enabling organizations to define business metrics once in Databricks and get consistent, governed results across Power BI, AI/BI dashboards, and Genie.
Trovata built autonomous AI agents for corporate treasury on Amazon Bedrock and AgentCore, using an RBAC-secured architecture and bank-branded deployment to cut TMS costs, save team time, and reduce reporting from hours to under a minute.
Tata Consultancy Services (TCS) built an agentic test orchestrator on Amazon Bedrock AgentCore that turns requirements into executable, governed tests for its Banking, Financial Services, and Insurance (BFSI) clients. Specialized agents, per-tenant isolation, and workload-aware execution cut testing effort by 88% in an internal benchmark.
Workday used LangGraph and LangSmith to build an internal AI agent platform, cutting prototype time from months to less than a day while meeting strict compliance needs.
Cursor upgrades Projects to support multiplayer collaboration, allowing teams to work alongside their agents with shared context and agent-to-agent communication. The beta rollout is available to all Cursor users starting today.
Unstructured adds an image description enrichment to its Transform parse endpoint, letting a single flag return written descriptions of every detected figure, including values read off charts, so chart-only findings become searchable.
Anthropic's IPO roadshow video reveals previously unreleased details about its multi-chip training strategy and compute multipliers, with co-founder Tom Brown explaining how Claude runs on AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs.
Google Arts & Culture launched the first online retrospective of South African artist Dr. Esther Mahlangu, featuring over 45 curated stories, virtual tours, and AI-recreated early memories to preserve Ndebele culture.
This paper presents GRF-MPC, a framework that integrates live ground reaction force and center of pressure measurements into a human centroidal dynamics model to optimize assistive forces via real-time quadratic programming. Deployed on the RobUST with an underactuated 4-cable configuration, it outperformed a fully actuated 8-cable baseline, reducing peak CoM error by 19.6%, tracking error by 24.1%, net mechanical work by 28.7%, and increasing voluntary excursion area by 45.9%.
PhysMoE introduces a novel architecture with modality-decoupled Mixture-of-Experts, single-token coordinate encoding, and differentiable SDF supervision, addressing bottlenecks in coordinate prediction for embodied VLMs and achieving 15.7% average gain and 8.31x throughput speedup on 9 grounding benchmarks.
The paper proposes a selective localization pipeline that combines heatmap and coordinate supervision with a logistic-regression scorer for surgical action-center localization. On 5,000 frames from 50 surgical sources, at an 80% acceptance quota and a tolerance of 0.10 times the image diagonal, the method retains 46.08 correct outputs per 100 inputs versus 44.20 for the baseline, a gain of 1.88. The work argues that evaluating retained correct outputs, not just the supply of correct predictions, is necessary to tell whether a selection rule preserves or erodes localization improvements.
The paper proposes PR-OVD, a partition-and-refine framework that explicitly reserves capacity for novel categories and strengthens category separation, achieving an average 6.5% proportional improvement on novel categories over state-of-the-art methods and integrating seamlessly into different detector architectures.
Researchers propose URL-Sim, a unified and regulated dynamic LiDAR simulation framework that uses a dynamic Gaussian scene graph and a regulated Gaussian densification strategy to overcome the limitations of existing 3DGS methods, achieving higher quantitative accuracy and visual quality on Waymo and KITTI-360 datasets.
The paper identifies three blind spots in OKS-mAP, the dominant metric for 2D human pose estimation: it does not distinguish visible from occluded keypoints, does not measure joint-angle error, and ignores temporal stability of tracking-free per-frame inference. The authors propose a three-module supplementary protocol and evaluate it on four datasets and four models, finding that aggregate mAP overestimates occluded-scene performance in every combination and that angle and temporal metrics provide independent information.
This paper reveals a candidate-support confound in offline web-behavior benchmarks where inserting the demonstrated action into a finite candidate set allows support membership alone to identify the gold action. It introduces support-identifiability rate (SIR) to quantify this effect, constructs support-balanced candidate sets, and evaluates a target-free dual-path proposer on OPeRA and Mind2Web, showing that controlling support composition is essential when using offline results to reason about deployment.
SELD is a training-free and label-free method that detects silent evidence loss in context compression by comparing a reader model's preferences over unspoken answers before and after compression, achieving 83% identification rate on a benchmark.
The paper introduces SREE, a pipeline that extracts first-person, symptom-specific text spans and selects a representative evidence subset via a facility location-inspired objective before evidence-gated inference, improving the precision and interpretability of depression questionnaire inference from social media.
The arXiv paper 'Third Eye' argues that evaluating persistent self-updates requires looking beyond immediate performance to outcomes after further native updates. Across five 3-8B backbones on math and code, immediate-selected utility is near zero or negative, and continuation reduces it further; the authors recommend treating persistent self-updates as adaptation transitions with independent scoring and a no-update baseline.
Researchers propose MADSNet, a multi-scale axial-attentive dense spiking neural network for spike camera image reconstruction. The hierarchical spiking encoder-decoder outperforms SSIR by 4.3 dB PSNR and Spk2ImgNet by 1.94 dB on SREDS, while achieving about 15× lower latency and using only about 1/40 of the estimated energy compared with Spk2ImgNet.
This paper reframes reward heterogeneity in personalized federated reinforcement learning as a structural asset that catalyzes exploration rather than a hurdle to consensus. The authors propose PF-UCBVI, which achieves optimal linear speedup by decoupling shared dynamics from personalized objectives, and then introduce PF-EFVI, a purely greedy algorithm that leverages reward diversity to ensure state-action coverage. They prove that PF-EFVI attains logarithmic regret without explicit exploration bonuses under sufficient reward diversity.
ConceptTime uses a frozen probabilistic forecaster's predictive surprise to simultaneously locate segment boundaries, build a vocabulary of dynamical-regime concepts, and derive reliability, matching black-box accuracy while beating interpretable baselines across 17 benchmarks.
This arXiv paper re-evaluates TabPFN models and TabICL v2 against Optuna-tuned gradient-boosted trees on five public telecom churn datasets, scoring expected maximum profit and calibration as well as ranking. TabPFN-3.5 earns significantly more profit on Cell2Cell and Iranian, but not on Orange; a larger tuning budget narrows the earlier 6.1-point PR-AUC gap to 4.0 points without closing it, and the TabPFN-3.5 weight licence, not scoring cost, is the main barrier to production.
This master's thesis by Chamani Shiranthika examines reinforcement learning for cancer chemotherapy, first simulating treatment via traditional Q-learning and then proposing a model-based offline RL approach that combines conservative Q-learning with model-based offline policy optimization on real human colon cancer clinical data. Experiments indicate the Chemo RL model handles chemotherapy scheduling well and can support junior oncologists in treatment decisions. The work received the IEEE Taipei Section Best Master's Thesis Award in 2021.
More than 20,000 veterans and transitioning service members have claimed a free year of ChatGPT Plus. They use it to prepare for civilian careers, and OpenAI Academy and partners have hosted trainings.
The paper presents a deployment adaptation method that learns a sensor-return adapter from limited calibration data while keeping a pretrained world model fixed, estimating measurement dropout to guide next-best-view selection. In real-robot experiments with a structured-light sensor, the adapted method achieved higher surface coverage with 60% fewer captures and recovered a region the nominal method repeatedly failed to measure.
The paper proposes DCLDs-EIR, an evidence-grounded iterative reasoning framework that couples diagnostic hypothesis refinement with multimodal evidence acquisition and verification for the differential diagnosis of diffuse cystic lung diseases, combining multiview renderings, axial CT inspection, and disease-specific knowledge graph retrieval in a closed feedback loop. Multicenter evaluation shows consistent accuracy gains over baselines and traceable differential diagnosis.
The paper proposes Trajectory-Regularized Distribution Distillation (TRD), which combines distribution-level and trajectory-level supervision via a momentum-mixture policy predicted in a single student forward pass. On SANA-1.6B and Stable Diffusion 3.5, TRD outperforms consistency and distribution matching baselines in generation quality and training efficiency for one-step and few-step generation, with a trade-off in sample diversity.
AptM-GCD introduces Attribute-Evidence Learning (AEL) and Multimodal Prompt Learning (MPL) to treat attributes as localized visual evidence, avoiding misleading global alignment and achieving state-of-the-art results on six datasets.
A new study measures tokenizer disparity across nine public tokenizers for 25 Indian languages, finding that GPT-4o's tokenizer requires 1.57 to 5.01 times more tokens than English, and 13.7 times for Santali in Ol Chiki script, directly inflating API costs.
Thai researchers propose segmenting Thai at the morpheme (khammoon) level rather than the word level to eliminate compounding ambiguity. They create a formal guideline with near-perfect human agreement, scale annotation with an LLM, and train a 1M-parameter convolutional BiLSTM segmenter that achieves 0.992 boundary F1 on out-of-domain text while maintaining or exceeding word-level baselines on downstream tasks.
An arXiv preprint studies how clinician edits to AI-generated clinical documentation can serve as scalable training supervision. The authors find that over 45% of edited outputs in real deployment data contain user-added information unsupported by model inputs, so naive fine-tuning risks teaching hallucination; they build a taxonomy of editing behaviors, generate semi-synthetic edit data on MIMIC-III discharge summaries, and show that progressively filtering toward grounded edits improves clinician-alignment metrics while preserving faithfulness to the input context.
TimeOmni-o1 bidirectionally couples a pretrained time series foundation model (TSFM) with an LLM: the TSFM encodes series in compressed form for the LLM to reason over with text, while the LLM's reasoning conditions the TSFM's forecast. Trained on the new MMTR dataset (12 domains, 500k+ samples), it outperforms existing models on understanding and forecasting across most domains and three open benchmarks, as well as its LLM and TSFM components used individually.
The paper proposes StepRefine, a plug-and-play decoding strategy that detects entropy spikes during iterative denoising and switches between discrete denoising and latent refinement to mitigate hallucinations in multimodal diffusion language models (MDLMs).
A new paper proposes a four-stage Bayesian pipeline to separately test decodability, shape, ownership, and causal use, offering a stricter evaluation of whether geometric structures found by linear probes actually participate in model computation. Applied to two-digit addition and multiplication in GPT-J 6B, Llama 3.1 8B, and Pythia 6.9B, results show nearly all concepts are decodable, but strongly supported shapes are ribbons or tori, 81% of shapes are inherited, and intervening on shapes retains about 61% of the effect. The authors recommend reporting probe results as verdicts on all four claims rather than a single accuracy.
This survey provides an algorithm-centric and evaluation-grounded review of multimodal attributed graph learning, introducing a unified message-passing formulation across five integration levels, reviewing benchmarks and evaluation protocols, and connecting the field to AI agents and graph engineering.
The paper proposes StatPrompt-LLM, which augments Time-LLM prompting with dataset-level statistics and a zero-initialised autoregressive bypass, reporting average relative reductions of 40.2% MSE and 21.4% MAE under a 10% reduced-prefix setting, while full-prefix results remain dataset-dependent and computational overhead increases.
ADIndex is an adaptive, distillation-driven indexing method for token-level sparse attention that combines layer-wise calibration, group-level distillation, and low-latency inference to significantly reduce indexer latency in long-context inference, while maintaining near-lossless task performance on DeepSeek-V3.2 and GLM-5.1.
SetPrint is a training-free black-box LLM fingerprinting method that jointly exploits within-set relations and between-set correspondence of response changes, outperforming existing training-free approaches on LeaFBench.
A controlled study evaluates when and how adding graph visualizations to prompts improves knowledge graph question answering in vision-language models, finding an average 1.85 F1 gain driven by low-complexity questions but highly sensitive to graph layout.
This study uses a cross-persona paradigm across 14 model variants from Llama 3 and Qwen 2.5 to test whether LLMs internally represent psychological constructs. Larger instruct models reproduce and amplify human cross-scale correlations; internal geometry aligns with the human matrix and correlates with behavioral amplification, but a probe-behavior gap termed readout attenuation remains, and the measured directions are not shown to mediate outputs.
Researchers propose LandslideFM, the first unified foundation model that connects landslide visual perception with contextual reasoning. It uses semi-supervised learning to adapt to multiple visual tasks and leverages predicted regions to guide a vision-language model, alongside the release of LandslideVL, the first benchmark for landslide analysis.
arXiv paper introduces TopicScope, a framework for training causal topic classifiers on SAE features that assign a topic to every token during generation, generalizing from prefill to decode on Gemma-3 1B and 4B with strong robustness.
EarthRecord is a new benchmark built on the raw record of 56 sites spanning four decades of Landsat, MODIS, VIIRS, and Sentinel-2 data, with 238 land-change questions across ten tasks that require agents to handle cloud masking and cross-sensor harmonization themselves. Evaluations of twelve multimodal agents show that leading agents complete most analyses end-to-end but falter when answers require mapping from imagery; instrument-pair controls reveal that half of correct answers follow altered satellite data, indicating agents have not yet grounded answers in real ground change.
Simon Willison releases llm-gemini 0.36, adding support for Google's Gemini 4 Carbon, now available to all. The model is priced at $2/$10 per million tokens until December 1st, when it doubles. Willison tests it with pelican prompts and a Three.js mech animation, plus a community demo of a 3D game built with Antigravity's /swarm.
Pinecone now supports OpenAI's text-embedding-3-small and text-embedding-3-large models in its integrated embedding feature, letting teams run semantic search without managing an OpenAI account or API key. Tokens are billed on the Pinecone invoice at OpenAI's list price.
This paper introduces Latent Barrier Steering (LBS), a hierarchical semantic-to-path framework for safe flow-matching-based generative planning. LBS first steers the generator's behavior latent toward rollouts with larger safety margin, then applies a path-space control barrier function quadratic program only for residual violations, improving goal reaching and reducing correction burden while preserving safety.
A NeurIPS 2026 position paper introduces the 'competence circuit' framework, decomposing an AI agent's real-world impact into intelligence-mediated discovery, procedural know-how, and direct control. Through simulation analysis, the paper argues that intelligence-driven speedup is capped by the ratio of competing intelligence levels, and that a large speedup multiplier need not imply rapid progress. Consequently, AI governance should prioritize restricting control and enforcing knowledge boundaries rather than limiting intelligence alone, and suggests intentionally 'putting physics in the loop' in system architectures.
This post builds a centralized observability data lake that aggregates Amazon Bedrock invocation logs from every linked account into one queryable table, so you can see prompts, token counts, model choice, and per-invocation cost across your organization.
RoboACE jointly evolves a curriculum harness and an execution harness around frozen foundation models, reaching 74.0% success on LIBERO-PRO, a 15.0-point gain over the initial deployment.
This paper introduces CAFE, a framework that enables scene-driven robot autonomy by formulating and executing tasks from observations under predefined categories (cleaning, safety, storage/organization) without instance-specific goals or action sequences. CAFE maintains online memory of objects, functional zones, and execution states, and uses a rule-guided Trigger-LLM with Draft, Complete, and Audit LLMs to propose goals, resolve details, and check dependencies. Outcomes update memory for correction and discovery. Evaluated in 12 AI2-THOR scenes, Tri-LLM achieves higher GCR scores than baselines across three backbones and task-complexity tiers. Controls validate components, and LimoCobot traces illustrate storage, cleanup, and recovery from a failed grasp.
Maze Runner is an open-source C++/ROS 2 autonomy stack that enables end-to-end maze navigation on a simulated Husarion ROSbot XL with a 2D LiDAR. It extracts wall lines with RANSAC and total least squares, supports EKF localization and EKF SLAM, plans with A* or Hybrid A*, and controls with NMPC plus discrete-time control barrier functions. In simulation, the robot completes a 135 m route with 1.0 cm localization RMSE (known map) and 24 cm (SLAM), planning in 6 ms and NMPC solving in 11.7 ms on average within an 80 ms budget.
At COP31, with a focus on zero waste, NVIDIA highlights four startups using vision AI and robotics to sort mixed waste more precisely and keep valuable materials in circulation.