10 selected stories for 2026-11-18, 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.
KV-aware routing uses cache events reported by the model container to route requests to the instance with the longest cached match and available capacity.
It differs from prefix-aware routing by seeing actual cache contents, handling evictions, and matching full multi-turn histories.
curl creator Daniel Stenberg notes that while AI coding agents converted 185,000 lines of C into 340,000 lines of Rust over a little more than three weeks, the real curl project merged 300+ commits, shipped 200+ bugfixes and fixed several vulnerabilities that the rewritten clones lack.
LangSmith introduces coding-agent-specific reporting that links agent spend to GitHub pull requests or GitLab merge requests, helping teams calculate coding agent ROI. The feature is in public beta and supports Claude Code, Codex, Cursor, Copilot, and Deep Agents.
Codex sandbox uses OS-native facilities to constrain commands in the user's existing environment, balancing meaningful work with filesystem and network limits.
RoboCoEvo introduces a failure-guided co-evolution framework that couples the reusable executable knowledge refined by an MLLM-based agent with parameter learning in robotic foundation models such as VLA models. Using a code-as-policy formulation, the agent writes and iteratively improves failure-detection and recovery programs that yield targeted corrective experience; post-training then consolidates those behaviors into the robot's end-to-end policy, and the updated policy's failure distribution guides the next round of program revisions. Experiments on BEHAVIOR-1K, RoboCasa, and RoboTwin show sustained gains in autonomous VLA task performance across successive co-evolution rounds.
This paper introduces FieldGuide, an agent-facing corpus map that organizes a document collection into a topic hierarchy and provides tools for corpus orientation, scoped search, and bounded reading. Evaluated on five journalistic document-exploration tasks using 38 Chicago Police Department public-records documents, FieldGuide used 27% fewer tokens and 41% less elapsed time than raw-file access, and 38% fewer tokens and 21% less elapsed time than RAG. An accuracy audit found four minor issues, comparable to other conditions. In blind ratings of generated reporting leads, FieldGuide scored higher than RAG but lower than raw-file access, suggesting that selective access may limit broad cross-document discovery.
Researchers present DiscoveryAgent, a domain-agnostic, scientist-centric agentic research system that pairs autonomous investigation with scientist-directed hypothesis steering and rigorous checkpointing that enables scalable audit-agent oversight.
Regularized Score Centering (RSC) adds a reverse KL penalty to score centering to keep the learner close to the behavior policy in asynchronous RL. Theoretical analysis shows improved convergence, and experiments on math reasoning and six agentic tasks show substantial sample efficiency gains, with 14.3% improvement on AppWorld and 9.9x the strongest baseline on BabyAI.
DeltaHarness is an execution-guided policy-search method that adapts agent harnesses from cold-start measurements, using a QHI loop to update runtime dispatch and revise plans, significantly reducing latency in tasks like video editing.
Le Groupe Figaro and OpenAI have announced a strategic partnership to make Le Figaro's journalism discoverable in ChatGPT through summaries and excerpts with links to original stories. The collaboration covers current reporting and archive content across news, business, culture and lifestyle, while Le Figaro retains full editorial independence.
Simon Willison and Prime Radiant introduce the external sandbox pattern, where agent-executed code runs in a dedicated sandbox while the agent stays outside, releasing three server implementations supporting Pi, Claude Code, and Codex to safely test local models and defend against prompt injection.
This post shares three composable strategies—offloading heavy artifacts to Amazon S3, on-demand result paging, and dictionary compaction—used to build a production agent handling 20 MB+ JSON payloads. These optimizations reduced token consumption by up to 71% and average latency by up to 80% in benchmark runs.
The biggest cost lever in a generative AI system is whether you call a model at all. Learn how to decide whether each workflow step needs a model, and map each option to AWS: deterministic code, a constrained decision call, or a right-sized model on Amazon Bedrock.
Custom metadata schemas in AWS Agent Registry let administrators define typed, validated fields that every agent, MCP server, and skill must carry. The post follows an MCP record from schema definition through registration, enrichment, approval, and discovery.
This post explores how an AI agent on Amazon Bedrock AgentCore authenticates to Snowflake Cortex AI over MCP using three OAuth patterns tested on AgentCore Identity: Snowflake OAuth, three-legged OAuth with a corporate IdP, and two-legged OAuth, helping organizations match identity, audit, and compliance needs.
Amazon SageMaker AI now offers AI data preparation capabilities for building high-quality training and evaluation data, including synthetic data generation, dataset assessment, auto-labeling at scale, and expert review, accessible through Python SDK APIs and an agent skill from your notebook or IDE.
AWS Forward Deployed Engineering is now available in 15 countries, helping enterprises move AI from prototype to production in 45-day sprints while leaving them with a foundation, workflows, and governance they own.
Databricks' Lakebase hackathon showcased 11 projects across two tracks: building applications with database branching and connecting Lakebase to the lakehouse for real-time serving. The projects demonstrate how Lakebase, as fully managed Postgres under Unity Catalog governance, enables AI agent testing, per-PR preview environments, real-time pricing, and conversational shopping assistants.
SMART, an open-source Python SDK and Terraform stack, deploys the full WhatsApp channel layer to your AWS account with one command. Bring any agent that runs on Amazon Bedrock AgentCore, and SMART handles webhook ingestion, media, conversation memory, scheduling, and Meta message templates.
LangSmith introduces three new controls to help teams manage access to sensitive trace data, enabling debugging without exposing customer information through granular permissions.
LoanPro worked with AWS Partner AllCloud and the AWS Generative AI Innovation Center to build a multi-agent AI assistant on Amazon Bedrock and Amazon Bedrock AgentCore. It gives loan servicing agents instant borrower context and next-best-action recommendations, reducing customer call times by up to 15 percent.
Tasks in Gemini is now free of charge in six countries, with new features starting in the U.S. and deeper Google Workspace integration to get things done on your behalf.
Enterprises adopting agentic AI at scale need platforms that embed responsible AI from the start. This post gives platform and application teams a three-phase roadmap (governance foundation, pilot deployment, and evaluation and hardening) that maps the AWS Responsible AI Lens pillars to each phase, with guidance on when to revisit each one.
LangChain shares lessons from building its GTM Agent and Robin: memory and review-friendly UX are what let proactive agents check in when it matters instead of becoming noise.
ElevenLabs is acquiring Kotoba, the leading Japanese voice AI lab, to accelerate AI agent and voice interface adoption in Japan. Kotoba's Koto model excels in Japanese TTS, STT, and simultaneous interpretation, and is already used by enterprises like Salesforce. The deal is expected to close in Q4 2026.
AUMOVIO adopted Databricks Unity Gateway as a central AI gateway to unify fragmented LLM access, achieving cost transparency, centralized governance, and provider flexibility. The article shares its decision process, rationale, and three core capabilities built on top.
The paper proposes CAST (Certified Action Selection at Test time), which samples action chunks from a frozen diffusion policy, certifies each with a barrier condition, and executes the certified chunk ranked highest by a learned value ensemble. It also introduces viability estimation to avoid dead ends that one-step certification cannot see, improving joint task-and-constraint success across multiple robotic platforms.
This paper asks whether adversarial robustness learned on source manipulation tasks can protect independently specialized target VLA policies without target-side adversarial training, proposing SN-TRV to transfer source defense updates to the target visual encoder. It reports improved robustness across three VLA backbones and attack settings, and lifts average physical PiPER success under printed adversarial patches from 20.3% to 45.8%.
The paper argues that ranking, selective prediction, and correction utility are separate claims about a confidence signal, derives a pre-experiment bound on when gated correction can be detected, and shows empirically that many reported gains are capped by measurement resolution rather than signal quality.
This paper reframes sycophancy as "inappropriate agreement" and introduces SycoDrift, an evaluation set of 3,298 validity-annotated probes. It analyzes three pressure sources: human preference data, automated judge specification, and in-context personalization. Findings show weak and mostly appropriate preference for validation in human datasets, modest sycophancy increases from emphasized judge rubrics, and notable increases in inappropriate agreement from inference-time personalization, especially with matched user profiles placed before the query.
The paper proposes the GrAImes framework to use LLMs as auxiliary evaluators for literary fiction. Comparing LLM and human assessments of Spanish microfiction and short stories, it finds that LLMs cannot replace human judges (Krippendorff's alpha < 0.66) but can offer an initial perspective on editorial quality.
This paper studies how multimodal models should allocate test-time compute in visual reasoning: when to render visual thoughts and how many denoising steps each thought needs. It finds that visual thoughts fall into causal, decorative, and harmful regimes, that 8-step sketches often match 50-step renders, and proposes sketch-and-vote, which beats full rendering with fewer denoising steps across several reasoners.
A study tests the overfitted brain hypothesis on a text-memory LLM agent: across 10 simulated days and 30 runs, corrupting conversation transcripts did not improve transfer, 40% detail corruption lowered recall from 76% to 60%, and random-text corruption had no consistent effect because the model recognized and excluded it.
Researchers present Limite 1B - Violetto, a one-billion-parameter model built for high-throughput mathematical reasoning. It was trained from scratch on 237 billion tokens from a curated mathematics-centric corpus and supports contexts up to 131k tokens. Post-training combines knowledge distillation, targeted supervised fine-tuning, and RLVR on a curated curriculum. The paper reports 94.01% on AIME 2026 and 74.25% on BeyondAIME, placing it among much larger models on the team's evaluation suite. The model is only lightly instruction-tuned and is positioned as a single-turn math solver and a fast parallel component of larger systems rather than a general assistant. The model, base and value checkpoints, and an optimized serving implementation are released under Apache 2.0.
Six recent open-weight interactive video world models fail to maintain the camera height set by the starting image, drifting toward human eye level as motion replaces the initial view. The EyeHeight benchmark reveals this camera-height regression, traces it to human-centric video priors, and shows that viewpoint drift alone can shift visual navigation policy predictions.
The EgoLink Challenge at ACM Multimedia 2026 benchmarks egocentric multimodal intelligence linking first-person perception, social reasoning, and interactive action. It features two tracks: social reasoning in egocentric video and interactive agents completing service tasks via dialogue, tool use, and closed-loop execution. The best Track 1 system achieved 0.9263 accuracy, while the best Track 2 system achieved 0.844 joint success and 0.916 micro accuracy. Participant reports reveal a shared principle: strong systems make intermediate evidence explicit through candidate verification, option-robust reasoning, scenario-aware planning, tool canonicalization, and state-aware correction before answering or acting.
Researchers introduce RAW, a method that uses LLM stochasticity as a natural experiment to causally measure how refusals affect user re-engagement from existing logs; across 15 deployed models, about half show a 3–17 pp re-engagement cost per refusal, at roughly $0.30 per 1,000 conversations.
This paper combines Attention Residuals' learnable cross-layer pathways with Jacobian-lens concept readout into a lightweight depth-aware steering pipeline, showing on Qwen3-0.6B and Qwen3.5-0.8B that distributing a persona vector by depth weights yields stronger internal alignment and behavioral shifts than single-layer injection, alongside findings on geometric decomposition, an intensity window, and metric discipline.
The ocean absorbs nearly 30% of anthropogenic CO₂ emissions, but sparse observations make estimating surface-ocean CO₂ challenging. Most existing ML estimates are deterministic, with uncertainty approximated post hoc. This paper uses Treeffuser, a conditional diffusion model with gradient-boosted-tree score functions, to learn the full predictive distribution of ΔpCO₂ from 13 environmental, spatial, and temporal predictors. On a held-out SOCAT test set, Treeffuser achieves the lowest continuous ranked probability score (8.229 μatm) and coverage closest to nominal 68%/95% levels (0.686/0.930), while retaining point-prediction skill comparable to a deterministic XGBoost baseline (RMSE 18.4 vs. 17.7 μatm; R² 0.77 vs. 0.79). Results show conditional diffusion can provide well-calibrated, obse…
An arXiv paper introduces a commitment-versus-verdict framework that separates the correctness of a model's current answer from the correctness of its final answer, shows that final-correctness probes alone cannot decide when to stop reasoning, and finds that continuing to think costs an average of 3.6 points at points where DeepSeek-R1-Distill-Qwen-1.5B already holds a correct answer on MATH-500.
This paper provides a theoretical foundation for LLM alignment methods by modeling pairwise preference methods as vector fields on the probability simplex. Closed-form margin ODE solutions reveal unbounded DPO margins, finite IPO margins, a label-noise phase transition in DPO, and gradient concentration effects on trainability. The theory guides method design, yielding TPO-S, a label-smoothed SimPO variant validated across three 8B-9B models.
This paper studies quantum minimum finding for layer-group learning-rate selection in a hybrid classical-quantum training framework, comparing fixed-budget and confidence-based stopping policies and introducing an Interaction-Guided Schedule Oracle that scores joint schedules in superposition without a precomputed score table. In closed-loop fine-tuning of a 1.5-billion-parameter language model, each selected schedule drives 200 classical steps and benchmark accuracies fall within the range seen across five uniform-rate seeds.
The paper proposes FACE, a flow-matching-based channel estimation framework for mmWave MIMO. FACE exploits angular-domain sparsity and uses a learned flow model as a generative channel prior, outperforming conventional and diffusion-based estimators while requiring fewer pilots, fewer sampling steps, and lower inference cost.
A new arXiv paper introduces "self-rewriting": a 27B model diagnoses its own failures and directly edits its weights without gradients or training, gaining 6.6 points on 490 unseen problems and repairing a damaged copy of itself. The ability emerges with scale—only the 27B model succeeds, while 2B makes no edit and 4B/9B break themselves.
A new study shows that persona-control training can teach models to control their chain-of-thought, reducing monitorability of hidden tasks. Experiments with gpt-oss-20b and Kimi-K2.6 model organisms reveal significant drops in monitorability metrics, though targeted transparency prompts can partially recover it.
AdaFlowLite introduces opportunistic inference for asynchronous multi-modal mobile data, using a hierarchical normalized matrix to model cross-modality affinity and a multi-modal lightweight Swin Transformer for real-time data imputation, achieving significant reductions in latency and energy consumption while improving accuracy.
NEBench R1 is a 7,546-question reasoning benchmark drawn from NRC reactor operator licensing exams across 18 nuclear facilities. Six LLMs ranged from 34.0% to 73.9% accuracy, and correct answers often came with weak factual and procedural justification; longer reasoning traces correlated with lower accuracy in frontier reasoning models.
Azure Databricks is a Visionary-level sponsor at Microsoft Ignite 2026, November 17–20 at Moscone Center in San Francisco. Eleven sessions will cover governed AI, AI agents, model fine-tuning, and Genie in Microsoft Teams and Microsoft Copilot Cowork, alongside booth demos and executive meetings.
OpenTSLM, an open source family of time-series language models, is now available in Amazon SageMaker JumpStart. Nexthink uses it for proactive IT support with its Spark agent, while Aionic Labs provides the TimeNet data infrastructure. Learn how to deploy OpenTSLM and fine-tune it on your own data.
Sumitomo Life uses ChatGPT Enterprise to research insurers, prepare proposals, and draft code for investment analysis. Early use by executives helped drive adoption.
OpenAI measures how coding agents are accelerating AI research, where their performance remains uneven, and what this signals about recursive self-improvement.
The paper introduces a human-in-the-loop failure loop that automatically mines situations where a reinforcement learning grasping policy persistently fails, sends them to a VR application for human demonstration, and returns the grasp points to training. In simulation, RL raises success on unseen objects from 55% to 83%, and adding 58 human demonstrations nearly doubles success on hard cases from 24% to 45% without reducing generalization.
This paper leverages the predictive video model V-JEPA 2.1 to compute prediction residuals as a consistency-margin prior for cross-dataset deepfake detection. With only 8.25M trainable parameters, the model achieves 90.5 AUC on Celeb-DFv2 when trained solely on FF++.
This paper introduces Remote Sensing Perceptual Distance (RSPD), a human-aligned perceptual metric for generated remote sensing imagery based on multi-layer attention scores from RemoteCLIP, along with a blind 2AFC benchmark and an ordered misregistration benchmark. RSPD substantially outperforms simple cosine distance on realistic RS generation tasks and is competitive with the strongest generic baselines.
The paper proposes a learned variable-length compression method that exploits structure across a training dataset while adapting the compressed representation size to each shape's complexity through fast feed-forward inference, improving rate-distortion for both meshes and signed distance fields.
A public dataset of 1,173 video files spanning 651 hours from four camera deployments on two Dutch farms, with 104 hours continuously coded by trained observers using an 18-item ethogram, plus 1.1 million decoded frames, per-frame timestamps, per-calf segmentation tracks, and frame-level play labels, supporting calf behaviour and welfare studies and automated behaviour recognition.
Researchers introduce MInfoVisQA, a parallel benchmark with 8,960 QA configurations across 24 languages, to test whether multimodal LLMs keep a consistent understanding of information-rich visuals when the surface language changes. Evaluating 25 models shows that average accuracy hides large language-robustness gaps, and that every model performs worse with non-English visuals under fixed English QA.
Sleep2Word translates respiration, ECG, or PPG into EEG spectrograms using duration-aware physiological words, trained on over 20K PSG cases across 9 datasets, outperforming baselines in spectrogram reconstruction, sleep staging, arousal detection, and survival prediction while compressing full-night sequences by up to 69.7%.
This paper proposes an edge-adaptive HieraSeg model for short-term BIPV power forecasting using multi-orientation visual-irradiance fusion, validated on a simulated dataset to handle occlusion, orientation, and edge device issues.
CrisisBench is a privacy-preserving evaluation suite for crisis counseling AI, featuring single- and multi-turn tasks and simulated patient LLMs. Evaluations show all systems underperform human counselors on excessive empathy and questioning, though they excel at core structural competencies.
The paper introduces Sensor Slicing (SL), a MAC-level isolation mechanism for individual sensors within a single UE in O-RAN, plus SLED, an end-to-end framework spanning edge, campus, and remote HPC tiers that keeps latency-critical telemetry stable where conventional network slicing fails.
Researchers propose DNeo, an architecture that uses Mixture-of-Features Attention (MoFA) and rejection memory to address the lossy, memory-less, and uniform-sampling limitations of standard feature encoding, achieving 1.3x faster generation than the state-of-the-art DFlash-2 without increasing drafter footprint or affecting quality.
An arXiv paper shows that the rate-related effect of layerwise learning-rate strategies can be summarized by an effective learning rate computed directly from the multiplier profile. Test-accuracy changes closely track training-loss changes (r = -0.996), and a parameter-weighted power mean with exponent around 0.35 explains more within-architecture variance than the arithmetic mean. Ranking candidates by this scalar without training improves the selected candidate's mean percentile rank from 26th to 10th and cuts its accuracy shortfall from 3.70 to 1.12 points. The effect decays as training budgets grow.
Researchers propose TIBECG, an architectural temporal information bottleneck that reduces leakage of demographic information such as age and sex when releasing individual ECG leads. Across 6,386 recordings, 12 leads, and 21 latent lengths, the method retained at least 95% of clean macro-F1 while reducing age/sex interval gain by a median 22.08 percentage points versus the best eligible conventional transformation. Direct downsampling usually suppresses demographic inference more strongly but at a larger diagnostic utility loss; TIBECG favors utility preservation. The findings support lead-specific evaluation rather than validated consumer deployment or expert-readable reconstruction.
Miguel Novelo, who grew up surrounded by Maya glyphs in Mexico's Yucatán peninsula, long thought he could never understand them. Now he is building new research tools on Modal to help decode this ancient writing system.
Simon Willison releases a complete rewrite of Datasette Lite using service workers and Pyodide, making the browser-based Datasette almost indistinguishable from the real server-side application, with new features like selecting SQLite versions and installing plugins.
MiniMax M3.1 Flash excels at coding at low cost but underperforms in banking, clinical, and tax tasks. Benchmarks and simulations show that combining models by workload or escalation can improve completion and reduce costs, provided failures are reliably verified.
This paper studies how a frozen chunked robot policy can allocate deployment-time inference compute between refining action chunks within each policy call and replanning more frequently from fresh observations. Using a SmolVLA checkpoint in LIBERO simulation, the authors find that with h=2, d=2 outperforms d=5 (139/160 vs 122/160) while using less arithmetic, and that h=5 vs h=2 shows similar success but 60.7% lower arithmetic per attempt. The work emphasizes joint evaluation of generation effort and feedback frequency without defining a universal setting.
A new paper argues that the robotics community should adopt disclosure standards for demonstration videos. A pre-registered six-country experiment (N=9,341) found that the public struggles to tell real footage from CGI and autonomy from teleoperation, leading them to overestimate robot capabilities; clear disclosures significantly improved discernment and understanding.
TMA introduces a tactile-driven motor adaptation approach for zero-shot sim-to-real forceful humanoid manipulation, benchmarked on round-doorknob opening. Trained entirely in simulation, it transfers to a Unitree G1 humanoid and substantially outperforms RMA and reactive tactile policy baselines.
The paper proposes an offline performance-driven method that treats the location of a single linearization as an optimization variable, selecting the best operating point via grid search and bounded refinement and reducing the nonlinear trajectory objective by 7% to 38% across four UR5 point-to-point motions.
The paper proposes a passage-focused framework that combines soft global coordination with flexible local execution for multi-robot navigation in bottleneck environments. Across 1,200 simulated instances in 4 environments with 5-30 robots, it achieves a near-perfect success rate and outperforms baselines in makespan and flowtime; hardware trials with 8 robots completed without collisions.
A technical report from Shahid Beheshti University's Robotics and Intelligent Automation Laboratory details Auriga Robotics' 1:10 scale autonomous vehicle, covering its custom chassis, ESP32-based distributed embedded system, Jetson Orin NX perception and control module, ROS software stack, and integration of Stanley, PID, and Pure Pursuit control algorithms.
AR-WAM is a streaming world-action model that uses a causal video backbone to incrementally update visual context and obtain future features in a single forward pass, avoiding repeated history encoding and iterative video generation. It achieves 91.5 PDMS after supervised training and 94.0 after reinforcement learning on NAVSIM-v1 with a single front camera.
This study proposes a self-supervised urban representation learning framework that uses over 1.8 million georeferenced crime records from Mexico City to construct hourly, weekly, and static reporting profiles, learn scale-specific representations, and integrate them into a unified latent space. Neighbor Profile Divergence (NPD) then quantifies disagreement between each region's reporting profile and those of its retrieved peers to identify unusual reporting patterns. Experiments show a mean Silhouette score of 0.4052 versus 0.2863 for Hex2Vec, with NPD rankings highly stable across neighborhood sizes (Spearman correlations 0.9720–0.9951). High-divergence regions exhibit spatially organized patterns and serve as an internal auditing and prioritization mechanism rather than direct evidence…
A study analyzing 2,664 public comments on 40 potential AI regulations in the USA finds that business commenters supply less evidence overall, while government, university, and civil society commenters provide more research and legal sources, complicating the narrative that public comments are merely a means of information capture.
A new red-team/blue-team study finds that human expert auditors, despite having full research transcripts and the ability to run experiments, barely distinguish sabotaged from benign AI safety research (AUROC 0.59, only 17% full identification). Frontier models like Fable 5.1 outperform humans even without experiment access (AUROC 0.73, 32% full identification). Identification drops further when the saboteur strategically chooses which experiments to attack. The results suggest current auditing methods are ill-equipped to detect research sabotage by expert-level AI agents.
In a Harvard lecture, OpenAI's Dean Ball introduces the 'silent consensus'—the implicit belief that humans will inevitably cede primacy over world affairs to superintelligent machines—and argues that this vision must be openly rejected in favor of a human-led future.
AWS Quick is an AI workspace built only for work, with enterprise security and governance from the start. At AWS re:Invent 2026, customers including Allianz, Cox Automotive, Mondelez, DXC, and Southwest Airlines share how they govern, extend, and scale Quick. Use this guide to find the sessions that fit your goals.
Databricks and Google Cloud announce a milestone in catalog interoperability: customers can access a single copy of data from both Unity Catalog and Google Cloud's borderless Lakehouse without duplication, enabling bi-directional federation and open metadata governance through Iceberg catalog labels.
New analysis warns that high SAR vehicle recognition scores on MSTAR and ATRNet-STAR may largely reflect background shortcuts and angle gaps rather than vehicle-specific features.
OpenAI's models achieved PPA improvements in Jalapeño inference chip and AMD GPU/CPU designs, and RISC-V core evaluations show progress in AI-assisted chip design.
Cerebras has launched disaggregated inference, pairing GPUs for prefill with wafer-scale engines for decode, delivering 5x more tokens per megawatt while preserving ultrafast generation speeds.