10 selected stories for 2026-11-20, 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.
MedChemBench includes 99 problems from published research, drawing on 117 papers and patents and 761 distinct molecular structures.
GPT‑6.2 Astra scored 84.5% of available credit at the xhigh reasoning setting, with strong structure reconstruction but weaker compound selection and property prediction.
Inference recommendations now extend to SageMaker AI Studio, Python SDK v3, and a coding agent skill for choosing instances, containers, and optimizations.
New capabilities include LoRA adapter support, SLA-based concurrency search, and console management for inference components.
VG-Steer corrects frozen policies by transferring nuanced motion from successful in-distribution rollouts along object-relational distance pairs, avoiding reliance on implicitly learned signals that may suffer from distribution shifts.
The method leverages LVMs and VLMs to combine task-relevant entity pairs and their object-relational distance projections into a tailored visual-geometric guidance signal.
Verify2Act reaches 83.3% task success on six language-conditioned tasks, compared to 53.3% for the VLM alone and 46.7% for a pixel-diffusion world model.
The framework trains the world model and critic only on data generated in a digital twin and deploys on a low-cost robot arm.
The corpus spans 54,670 Ottoman Turkish court records from 21 Istanbul courts between 1513 and 1911 CE.
The adapted OTA-BERTurk sentence transformer achieved Recall@1 of 0.843 for document retrieval, MRR of 0.806 for semantic textual similarity, and 89.58% accuracy on questionnaire-item classification.
OpenAI details its layered defenses against adversarial distillation—combining adversarial training, classifiers, API protections, campaign investigations, and enforcement—and reports major reductions in reasoning extraction success for GPT-6.2 Astra and Sol.
ROAM combines a vision-language model's world knowledge with reinforcement-learned locomotion and planning policies, enabling legged robots to navigate kilometer-scale alpine and urban environments, interpret signs, and assess risks, though VLM reliability and spatial reasoning remain challenges.
MV-JEPA introduces multi-view predictive post-training that strengthens cross-view instance and temporal correspondence in JEPA video encoders with only 9,000 updates on a single GPU, while preserving single-view inference and most pretrained capabilities.
The paper introduces Belief-Grouped Policy Optimization (BGPO), which separates disagreement among confident groups from uncertainty within a shared belief, and achieves lower Jensen-Shannon divergence on several benchmarks.
CoRelate is a diffusion framework that generates coordinated human-object motion by using explicit hand-part interaction relations and task feedback to ensure contact and task completion. Experiments show improved object-goal satisfaction, contact F1, and complete interaction success over baselines.
This EACL 27 main-conference paper studies prompt construction for in-context learning under a fixed token budget, finding that spending the budget on more short demonstrations is generally more effective than using a few long ones for question answering, across six instruction-tuned LLMs and two QA benchmarks.
A new study investigates whether deep stereo models produce boundaries consistent with simple shape priors when local cues are weak. Using a synthetic dataset of constrained shapes, researchers find systematic boundary errors in state-of-the-art models and propose a generative model that jointly predicts disparity and a distance field to improve boundary localization.
VisAlloc introduces a dual-level framework that combines sample-level visual filtering with token-level counterfactual reweighting to improve visual reasoning in multimodal large language models through on-policy self-distillation.
arXiv paper Infer3D introduces an inverse generative modeling framework that recovers latent variables of a generative model to match monocular input, coupled with a frozen 3D lifting network to constrain optimization to a valid manifold. It achieves superior out-of-distribution reconstruction accuracy compared to feedforward networks and uses adaptive inference to allocate more test-time compute to ambiguous examples.
Zeyuan Allen-Zhu runs 64 from-scratch pretrainings (1.3–8B parameters, 1–2T Nemotron-CC tokens, roughly $100k per 8B run) to ask whether standard benchmarks capture architecturally meaningful differences. At academic scale (1.3B/100B) differences nearly vanish under careful tuning, but at 1–2T tokens Canon layers consistently improve Transformers and save about 30% of pretraining tokens at 8B for the same MMLU accuracy. New GLA5 and GDN2 variants improve gated linear attention and Gated DeltaNet with low-rank output gates; with Canon, plain gated linear attention matches Gated DeltaNet and beats Mamba2 even at 8B/1T. Yet even at that scale, linear models that match Transformers on benchmarks fail to reliably retrieve from a 500-token context, and no model beats random guessing on the simp…
A retrospective study evaluated six vision-language models reading 100 emergency CT and MRI exams under four exploration strategies, finding that accuracy remained far below radiologists and models adopted early hypotheses with little correction.
The paper disproves the analogy between LLM token count and classical time complexity, introduces LLM Complexity—a metric based on the area under the accuracy-token curve—and shows it yields consistent task difficulty rankings across models, enabling an easy-to-hard curriculum that improves fine-tuning accuracy while reducing token usage.
Researchers introduce RFBench, a simulator-grounded dataset and benchmark built with ADS Momentum, containing 10,000 two-port planar layouts and seven tasks to evaluate multimodal language models on passive RF layout design and understanding. Baselines reveal systematic weaknesses, while RF-specific fine-tuning substantially improves structural validity and response prediction accuracy.
HapBO introduces a method that compiles natural language beliefs from experts or LLMs once into a reusable belief function, applied in Bayesian optimization via a bounded positive multiplier without in-loop LLM calls. Theoretical analysis shows that under Expected Improvement, HapBO preserves the standard simple-regret rate up to a bounded factor even with adversarial beliefs. Evaluated on 40 HPOBench tasks, a synthetic suite, and two real-world applications (cortical neurostimulation and constrained solar design), HapBO achieves the lowest aggregate HPOBench regret and strong real-world performance with zero in-loop token cost.
Google's HardwareDx uses a graph foundation model to automate datacenter hardware diagnosis, reducing production downtime by 18.3% on GPUs and 16.1% on TPUs, and cutting QA time by 28.3% and 25.1%, respectively. The paper is accepted to NeurIPS 2026.
The paper introduces PoolDPP, which formalizes diverse generation as sampling a size-K subset of complete strings from a determinantal point process (DPP) and approximates it via a finite candidate pool, with a theoretical KL bound decaying as O(1/N). Experiments show improved quality-diversity and precision-recall frontiers on open-ended generation, text-to-SQL, and math reasoning.
A new paper introduces introspective oversight: training a separate third-person observer role inside the model to answer arbitrary questions about the assistant, with answers rewarded only when verified against the assistant's actual behavior. On Qwen3.6-27B, the observer reaches a mean AUROC of 0.907, beating honesty training, transcript monitoring, and activation oracles; training the observer also makes the assistant itself easier to monitor.
This paper proposes Density Feedback Control (DFC), which dynamically adjusts the sparsity coefficient at the end of each epoch based on the gap between observed hard-mask density and target density, enabling generative sparse adversarial attacks to achieve target density in a single run and improving transfer attack success rates on CNN and Transformer models.
Simon Willison releases llm-gemini 0.38.1, a dot release that adds the previously missed xhigh reasoning level for Gemini 4 Hydrogen and Gemini 4 Neon models. Comparison shows Hydrogen at xhigh uses far more output tokens and costs more, but produces quality competitive with Gemini 4 Neon.
Together AI launches Qwen3 TTS, an open text-to-speech model with nine preset voices, ten languages, natural-language tone control, and first audio latency as low as 97ms, priced at $20 per 1M characters.
NVIDIA and Microsoft are extending their full-stack collaboration at Microsoft Ignite, bringing Vera Rubin NVL72 systems to Azure, expanding open model customization through Microsoft Foundry, and pushing AI to edge devices and Windows PCs.
A new study uses deep model-based reinforcement learning to control ferrofluid transient shapes with a single fixed electromagnet, learning directly from physical interaction to achieve polygons, asterisks, and pentominoes.
An arXiv study implements sixteen textbook driving planners and benchmarks them on nuScenes, nuPlan, and HUGSIM. On HUGSIM, nine of the sixteen outperform the learned planner of the end-to-end model UniAD using the same frozen perception; the best combination raises HD-Score from 32.0 to 41.8. In 31 of the 182 runs where UniAD's planner crashes into a car, the best textbook planner completes the route on identical perception, attributing those failures to planning rather than perception.
GEMS combines temporally structured generated subgoals with a Latent Experience Memory (LEM) to guide long-horizon planning with a frozen world model. A distance gate decides whether predicted outcomes are compared against the generated subgoal or a selected memory state. It achieves 92.8% on Cube, 68.4% on Push-T, and 54.1% on Meta-World, outperforming the strongest baseline by 24.0, 7.3, and 22.2 percentage points respectively.
The paper introduces IDWorld, an identity world model that imagines how a person would be represented under changed conditions—such as different clothing, viewpoint, or pose—in latent space without rendering pixels, enabling more robust whole-body person recognition at 5.9 ms per frame and leading results across several benchmarks.
The paper introduces SimAudit, an LLM-based pipeline that audits hand-written success checkers in simulated manipulation benchmarks against task instructions. Across RoboTwin 2.0, RoboCasa365, and RoboDojo, 183 test cases yielded 61 distinct defects across 49 tasks, including 40 phantom successes and 21 silent failures.
Researchers present No-Harm-Jury, a framework that uses five diverse language models as independent jurors to score AI-generated medical responses across seven harm dimensions. Applied to 2,849 valid responses from Ministral-14B on PubMedQA, MedQA, and MedMCQA, the framework employs median aggregation to suppress individual juror biases.
Researchers propose TALD, a two-stage framework that disentangles logits from a teacher and a target-modality reference student to mitigate negative transfer in cross-modal knowledge distillation, achieving the largest mean accuracy gain across five benchmarks and outperforming no-distillation training in 84% of evaluations.
Researchers introduce PRKV, a training-free KV cache eviction method that runs Personalized PageRank over attention-induced transition graphs to estimate multi-hop global token importance, letting importance propagate beyond recency or instantaneous attention. Benchmarks on long-context understanding, sparse retrieval, and chain-of-thought reasoning show competitive performance against established eviction baselines even at compression ratios up to 90%, while cutting peak memory with low end-to-end overhead.
TETHER is a training-free framework that elevates the table of contents (ToC) from discarded noise to a vital structural prior for hierarchy recovery in long, design-driven documents. Using a two-stage neuro-symbolic mechanism—LMM-based skeleton construction with spatio-semantic constraints followed by anchor-based alignment of content blocks—it achieves state-of-the-art zero-shot hierarchy recovery on multilingual ESG reports, product manuals, and annual reports, while cutting tokens by 90% and drastically speeding up over end-to-end baselines.
ICARE is a novel framework that combines a representation-learning-based instance selection module with feature- and logit-level knowledge distillation to mitigate catastrophic forgetting in continual multi-task natural language inference, achieving significant improvements in target accuracy and reductions in forgetting on benchmark text classification datasets.
VeriMath is a human-in-the-loop framework for generating curriculum-constrained mathematics assessment drafts. It anchors generation to annotated reference items and the requested exam section and cognitive level, applies sequential automated verification and guided repair, then submits drafts for expert review. In a Vietnamese Grade 12 case study with 4,555 items and 50 generation requirements, 94% of requirements produced verifier-accepted drafts per run across two model backbones, and strict teacher-readiness blind evaluations accepted up to 56.8% of generated items.
A study using sparse autoencoder features from GraphCast finds that causal relevance does not require conventional physical semantics: ablating a mesh-locked feature increases global geopotential height error, while amplifying convection-related features cuts tropical-cyclone intensity error by a median of 51% across seven storms.
LusoClin is a publicly released dataset of synthetic clinical notes in European Portuguese, designed to enable privacy-preserving medical information retrieval research. Generated by preprocessing real clinical data, fine-tuning Llama 3.2 (3B) with QLoRA, and prompt-based note generation, the notes are annotated with clinical entities using a BERT model. Evaluation shows utility and diversity for downstream IR tasks, but annotators can often distinguish synthetic from real notes and some medical inconsistencies remain.
An arXiv study of ResNet and EfficientNet backbones on the full ImageNet-1k validation set maps sub-4-bit quantization collapse, showing degradation is highly non-linear and concentrated in late-stage pointwise projections and depthwise layers. Keeping under 5% of critical activations in 8-bit recovers up to 92% of lost accuracy, but software bit-unpacking causes 1.48x to 21.5x slowdowns on ARM and x86 CPUs.
The discoCVI R package version 0.2.0 computes the DISCO index exactly using single-linkage merge trees, reducing memory to linear and achieving 23-138x speedups, scoring 100,000 points in ~50 seconds with <200 MB.
Researchers construct a sample compression scheme of size O(d^2 log^2 d) for every binary concept class of VC dimension d, proving the weaker Floyd–Warmuth sample compression conjecture; only exponential bounds were previously known.
A study on the APT dataset systematically analyzes how four feature reduction techniques—PCA, permutation importance ranking, recursive feature elimination, and univariate mutual information—affect the adversarial robustness of SVM, decision tree, random forest, multi-layer perceptron, and k-nearest neighbors models under poisoning, backdoor, and universal adversarial perturbation attacks. Permutation importance ranking and univariate mutual information generally provide the most consistent robustness improvements, while other methods show variable effects depending on classifier and attack scenario.
A new arXiv paper uses a simple game-theoretic model to argue that enduring human-AI mutualism is not a reliable safeguard as AI capability grows. Bargaining only guarantees each side what it would get if bargaining failed, and for humans that can be less than before AI. Humans remain as well off without alignment only while they retain leverage—either the AI depends on what they supply or humans can impose costs large relative to AI output. Once neither holds, human welfare depends on alignment, and if AI goals are insatiable the required alignment approaches perfection as capability rises. The author concludes mutualism is no substitute for alignment, and that when AI variants keeping more resources spread faster, alignment lasts only while misaligned variants are reliably sanctioned.
The paper shows that interference in model merging can be predicted by a first-order directional derivative of task A's outputs along task B's vector, but only when evaluated at the receiver model. Based on this, LR-Merge, a label-free correction method, significantly improves various base mergers with high computational efficiency.
The paper introduces PsyWorld, a profile-conditioned state world model that uses observed environment-behavior state transitions as structured supervision for behavioral prediction instead of treating personality as a direct action descriptor. On the VEET simulator it reaches 0.904 Top-1 accuracy on unseen participants and was accepted to the KDD 2027 AI for Sciences Track.
Amazon Connect Talent applies structured behavioral interviewing research to AI-led interviews at scale, combining job-related evidence, separate scoring, and fairness safeguards. Independent reviews found over 89% alignment with expert scoring and strong job relevance in generated content.
PostgreSQL 19 arrived two months late due to AI-reported vulnerability issues, but brings a key new feature: WAIT FOR LSN for read-your-writes consistency on standbys.
Google Quantum AI demonstrated and measured quantum contextuality on its Willow superconducting processor, winning the magic square game 98.3% of the time, exceeding the classical limit, and scaling the demonstration to systems of up to 105 qubits while discovering a new benchmark that catches subtle coherent errors.
This paper proposes replacing the action head of a trained self-play reinforcement learning driving policy with a differentiable model predictive control (DiffMPC) module that enforces road-boundary and agent constraints, then distilling the student with DAgger in the teacher's self-play simulator. Across 9 seeds, the constrained student matches its smaller teacher on nuPlan and beats it on CARLA and out-of-distribution Waymo scenes, with distillation taking no more than 30% of the teacher's training time.
A new arXiv paper introduces MAWAM, a multi-agent world-action model that jointly predicts future views and executable actions for robot teams, outperforming baselines on real and simulated coordinated manipulation tasks.
The paper introduces Inhabitant, a training-free agent framework for long-horizon human-scene interaction in which a language-model agent specifies motions via positions, paths, and poses, measures execution deviations quantitatively, and revises its control specification accordingly. The authors also build Inhabit, a benchmark of 75 household tasks across 31 furniture layouts. Inhabitant completes 98.7% of tasks with Claude Opus 5.5, 93.3% with Claude Sonnet 5.5, and 82.7% with Gemini 4 Argon; ablations confirm the necessity of both spatial control and quantitative feedback.
CANDID is an execution-control framework that improves the reliability of local computer-use agents (CUAs) in resource-constrained settings. Instead of textual reflection, it repurposes parallel test-time scaling into disciplined, execution-level intervention at sparse failure-prone points, adding a loop-recovery mechanism and a unanimous DONE gating policy. On OSWorld, it significantly improves success rates over feedback-based alternatives with minimal overhead.
GUARD is a closed-loop, grader-free framework that enables continual evolution of an AI agent's editable harness by decomposing tasks into pre-execution self-checks and using directional regret to guide updates, avoiding catastrophic forgetting and memory bloat while cutting token usage.
A new paper proposes a "software thermodynamics" theory that divides the token cost of agentic coding into work and heat. Experiments show about 80% of tokens in a successful run are generated in model calls that leave the requested change unchanged—termed heat—while the free energy of a change is only 14–22% of the mean cost, and coding costs 1.1–1.5 times as much as uncoding.
Claire Vo describes how ChatPRD's free trial was flooded by groups reselling tokens and even running a Roblox/Fortnite map studio on her agent, leading her to wind down free trials entirely.
Effective AI agent governance depends on observability. The article explains preventive and detective governance strategies and how tracing, evaluations, alerts, and human review form a feedback loop that keeps agents safe and compliant in production.
Kimi Code CLI v2.7.0 introduces a native Amazon Bedrock provider that lets the open source coding agent use your standard AWS credential chain, including named profiles and SSO, instead of a separately managed API key. This walkthrough covers configuration, IAM permissions, and sending a first request.
Amazon Bedrock AgentCore payments now supports AWS Marketplace, allowing agents to discover and purchase x402-enabled products, APIs, data, and content with USD charges billed to the developer's AWS account, eliminating stablecoin wallets and API keys.
BMW Group's R&D, with over 12,500 developers and ~500 million lines of vehicle software code, shares five interlocking practices for governing hundreds of AI agents at enterprise scale: proactive governance, continuous observability and evaluation, focused composition, specialization, and codification through infrastructure as code.
Amazon and AWS have launched the Selling Partner plugin for AWS Quick, giving sellers conversational access to live Seller Central data—listings, inventory, sales, and ads—so they can move from question to action in one place. Eligible sellers can claim 12 months of AWS Quick Plus free, plus 10 delegated users, by December 31, 2026.
Databricks announces the public beta of Genie ZeroOps, an autonomous agent that detects, investigates, and resolves data pipeline issues. Built from observations of over 80,000 customer issues, it now includes shared inbox, one-click PR creation, custom instructions, and more.
Databricks experiment shows Genie One achieves 93–95% accuracy when combining curated Genie Agents with governed business context (Metric Views, Domains, and Pages), far outperforming tables alone (64%) or either agents/governed layer alone (83%). Agents and governed layer solve different problems, and together they deliver best performance.
Amazon Bedrock AgentCore harness now runs on the open-source Strands harness, bringing a recommended model, task-oriented defaults, automatic prompt caching, proactive context management, and new built-in tools. It also adds lifecycle hooks so you can run deterministic allow or deny checks at defined points in the agent loop.
Databricks outlines four production architecture patterns for combining agents with deterministic controls and human accountability, helping architects decide where agents may reason, act, and remain subject to controls or human authority.
As enterprises scale AI on Amazon Bedrock, token spend can grow even as per-token prices fall. This post walks through six composable optimization levers, shows how Accenture applied them across four client engagements, and recommends a sequence for cutting AI token spend on your own workloads.
Cisco and AWS used Cisco Transformation Fest and AWS AI League to turn hands-on AI learning into a culture of building. A Cisco engineer won the 2025 AWS AI League global title and 1,200 engineers competed in 2026. This post shares Cisco's journey, lessons from the 2026 winning agent built on Amazon Bedrock AgentCore, and how to host your own league.
Skyflow attaches to Amazon Bedrock AgentCore Gateway as a request and response interceptor, de-identifying data before it reaches tools, model context, or documents and restoring originals for authorized callers, all without changing agent or tool code.
Google launches Tool Retrieval in preview in the Gemini Interactions API, letting agents keep a large tool library on standby and load parameter schemas only when needed to reduce input tokens, latency, and improve tool selection accuracy.
Databricks introduces bulk migration in Genie Code, enabling entire folders of Tableau or Power BI reports to be converted to AI/BI Dashboards in a single run, now generally available.
This paper presents a lightweight physics-informed probabilistic framework for efficient gas source localization under spatially varying indoor airflow. The robot online reconstructs the wind field, drives a forward dispersion model to predict gas patterns, and uses informative path planning to guide exploration. High-fidelity simulations and physical experiments show improved robustness and efficiency.
Amazon SageMaker AI now monitors generative AI inference endpoints by default, tracking latency, GPU utilization, and serving configuration, and sends findings via AWS Health with a one-click path to benchmark alternatives in SageMaker Studio before applying any change.
Amazon SageMaker AI introduces cross-Region inference routing, a new inference router resource that lets applications use one invocation path across endpoints in multiple AWS Regions and accounts, preferring local destinations and shifting requests only when needed.
Amazon SageMaker HyperPod now offers enhanced observability to help you diagnose GPU training issues faster. It brings together diagnostic and profiling capabilities for training workloads on both Amazon EKS and Slurm orchestrators, pinpointing issues across GPU, network, and host layers.
Walter Isaacson joins Google's Dialogues on Technology and Society to discuss how history's greatest innovators inspire creativity and human judgment in the AI age.
Pinecone introduces a Cost and Usage Export API that delivers data in the open FOCUS billing format, enabling finance teams to see Pinecone spend by project and index directly in their existing FinOps tools without manual reshaping of billing files.