AI Daily Briefing

This post shows how to build a custom meeting prep and follow-up assistant using Amazon Quick and Cisco Webex MCP servers. From a single prompt, the agent finds an upcoming Webex meeting, reviews prior meeting summaries and transcripts, and pulls related Vidcast highlights and transcript context. It then searches Webex message threads for unresolved follow-ups and creates a concise prep brief. After the meeting, the same assistant can summarize the discussion and identify action items. It can also find related Vidcast updates and draft a follow-up message for the right Webex space.

This post outlines a cost-effective and scalable intelligent document processing pipeline on AWS using Amazon Bedrock, including its Data Automation (BDA) for extraction, Strands Agent for coordination, and Knowledge Bases for contextual understanding across documents.

This week Miguel Fierro, a former Microsoft principal researcher who recently founded his own company, RecoMind, joined data and AI evangelist Christina Stathopoulos to talk about the state of recommendation systems. Christina also ran through the latest AI news she’s been watching, from Anthropic’s continued rise to responsible AI, announcements from Google’s I/O 2026 conference, and (continuing the discussion from last week) the growing backlash against tokenmaxxing as a productivity metric. Here are three takeaways from the conversation.

Local models in 2026 are good enough. For the tasks Claude Code handles daily: code completion, refactoring, debugging, codebase explanation; a well-chosen quantized model running locally covers the vast majority of real use cases at zero per-token cost and with no rate limits.

SpaceX’s $75B IPO sets record; OpenAI and Anthropic file for IPOs; Anthropic CEO urges government limits on AI; Apple unveils Siri AI; Prometheus raises $12B; plus enterprise AI news.

AWS Professional Services compressed engagement timelines from months to days by fundamentally rebuilding its delivery process, not just adding AI tools. This post shares how they became a frontier team and the practices that enabled it.

OpenAI is acquiring Ona, formerly Gitpod, a startup founded in Kiel, Germany in 2020 that specializes in AI agents and secure cloud development environments for software development.

OpenAI introduces three Academy courses that help people build practical AI skills, create repeatable workflows, and apply agents in everyday work.

Moonshot AI has introduced Kimi Work, a local desktop AI agent for macOS and Windows that runs a swarm of up to 300 sub-agents on your machine. It drives your logged-in browser via WebBridge, reads local files, and schedules background jobs with a built-in cron engine. Based on the Kimi K2.6 MoE model (≈32B active parameters, 256K context), it targets knowledge workers by keeping data and execution local.

Customer segmentation is the practice of dividing existing customers into groups based on shared characteristics to tailor marketing and services. This guide covers types, methods, importance, challenges, and how AI is transforming segmentation.

This tutorial builds an end-to-end 3D medical image segmentation pipeline using MONAI to segment the spleen on the Medical Segmentation Decathlon Task09 dataset. It covers volumetric CT processing, medical imaging transformations, training a 3D UNet with mixed precision and DiceCE loss, sliding-window inference, validation, and qualitative visualization.

The article discusses the emerging trend of designing loops to drive AI agents instead of manual prompting, covering key figures' insights, Anthropic's Fable 5 rollout controversy, automated research systems, data infrastructure bottlenecks, inference speed optimizations, and agent tooling developments.

A new arXiv paper introduces DARRMS, an algorithm that reduces computational resource demands by allowing agents to limit their attention radius dynamically, improving coordination and scalability in uncertain environments while maintaining decision-making robustness.

This paper presents G-MAPP, a GPU-accelerated framework for reactive motion generation that achieves up to 5x speedup by parallelizing world modeling and planning on GPU, enabling real-time perception-action coupling in dynamic environments.

A new preprint explores the transition from human-level Artificial General Intelligence (AGI) to Artificial General Superintelligence (ASI), outlining four potential pathways: scaling AGI, AI paradigm shifts, recursive improvement, and ASI emerging from large-scale multi-agent collectives. The paper also discusses frictions, bottlenecks, and the possibility that AI progress may accelerate, leading to a series of transformative changes rather than a single breakthrough.

Traditionally, decision support studies how humans use ML models to make better decisions. In modern agentic systems, AI agents act on behalf of users, reversing roles. This paper proposes a framework to minimize support usage while controlling counterfactual missed-support error. Optimal policy is a threshold rule; online algorithm with adaptive thresholding and calibration-on-the-fly reduces unnecessary support. Experiments show reliable error control and reduced support usage.

AI agents are evolving from simple chatbots to autonomous digital workers, raising security and governance concerns. Experts advise treating them like human interns with strict oversight, specific instructions, and careful monitoring to prevent unintended actions. Balancing independence with control is essential.

OpenAI Group PBC today announced plans to acquire Ona, a startup with a platform for managing long-running AI agents. The acquisition will enhance OpenAI's Codex AI assistant by enabling it to perform tasks that span hours or days. Ona's cloud sandbox technology allows AI agents to continue running even when developers shut down their workstations, and provides security features such as blocking malicious programs via hashing.

As enterprise AI spending accelerates, FinOps AI governance is under stress. Traditional cost optimization levers are insufficient against token-based pricing and opaque billing. 98% of practitioners now manage AI spend, but most lack visibility and governance structures. Automation is essential, and cross-team collaboration is key to understanding cost context.

Israeli data engineering startup Upriver Data Ltd. today announced it has raised $14 million in new funding to automate the data work that enterprises depend on to make artificial intelligence projects succeed. Founded in 2024 by Chief Executive Ido Bronstein and Chief Technology Officer Omri Lifshitz, Upriver has built what it calls an artificial intelligence-native platform that connects to an organization’s full data stack, resolves data quality issues and maintains pipelines automatically. The company pitches the result as a reliable data foundation that AI systems can run on without constant manual upkeep from engineering teams.

Mercedes-Benz Korea built a unified semantic layer on Databricks, migrating over 500 KPIs from Power BI to Unity Catalog, leveraging Genie and Agent Bricks to enable consistent semantics for BI and AI. An automated DAX-to-Metric-View transpiler accelerated the migration, providing a reference for other markets.

xAI today released the Grok Build Plugin Marketplace, a built-in catalog of plugins for its terminal coding agent. Plugins bundle skills, commands, agents, hooks, MCP servers, and LSPs into one package, installable without leaving the terminal. Six plugins launch with partners including MongoDB, Vercel, Sentry, Chrome DevTools, Cloudflare, and Superpowers, with commit-SHA pinning for security.

Databricks formalizes its Forward Deployed Engineering (FDE) organization to accelerate customer business outcomes with AI, combining the Lakehouse platform with embedded engineering, a global partner network, and direct R&D interlock. The team has worked with over 1,900 customers in the past year, including Fox, JPMC, and Qualcomm, delivering measurable results such as doubled search success rates, massive data migrations, and reduced workflow times from days to minutes.

Benchling's Head of AI Nicholas Larus-Stone discusses building agents for life sciences on the Max Agency podcast. He explains their multi-model approach for quality, production trace review processes, and how agents compress workflows to accelerate scientific discovery. Benchling AI launched in October 2025 on top of a 14-year-old data platform.

Prometheus Inc., an AI startup co-led by Jeff Bezos, raised $12 billion in Series B funding at a $41 billion valuation. The company is developing AI tools to accelerate hardware development, focusing on prototyping and pre-production manufacturing. The funds will mainly be used for computing infrastructure.

Chainguard introduces a new source code scanner that detects 'greyware' — open-source packages that are functionally transparent but contain harmful behaviors. The scanner has identified and blocked over 52,000 malicious or greyware packages, highlighting how agentic development exacerbates the problem.

LocIn AI launches on Product Hunt as a localization tool that preserves brand voice across languages using tone-aware AI, automated workflows, and developer-first tools. The platform offers CLI integration and API access, aiming to solve the problem of translations that are technically correct but feel off-brand.

AI is reshaping the tech workforce, especially entry-level roles. A Linux Foundation report finds a 27% net increase in European tech hiring, but junior hiring contracted 3% in Europe while growing 14% elsewhere. Companies invest 3.7x more in training existing staff than hiring. The junior role is being redefined, demanding broader skills including AI fluency, security awareness, and business understanding. Forward-deployed engineers emerge as a key new role.

Databricks launched its inaugural Student Fellows cohort, selecting a highly diverse group of students from over 5,000 applications across hundreds of universities worldwide. The fellows, chosen for campus leadership and technical expertise, will bridge academic theory and real-world data/AI practice through workshops, hackathons, and mentorship programs.

The Hertz Foundation awarded 2026 fellowships to three current MIT students and one incoming graduate student. The fellowship provides five years of financial support and autonomy for groundbreaking research in applied sciences, engineering, and mathematics.

Microsoft is a Legend Sponsor at Databricks Data + AI Summit 2026. The summit will showcase how joint customers use Azure Databricks to modernize data estates, scale AI, and unlock business value. Attendees can visit the Microsoft booth, attend breakout sessions on topics like federated analytics, ecosystem integration, and product announcements. Featured sessions include Unlocking the Microsoft Data & AI Ecosystem, Zero-Copy Federated Energy Analytics, and customer stories from GEODIS and TK Elevator.

Coinbase launches Coinbase for Agents, a standalone tool enabling AI agents to trade cryptocurrency and pay for services directly from assistants like Claude and ChatGPT. Users set spending limits, agents operate in isolated sandboxes, supporting spot and derivatives trading with plans for stocks and prediction markets. Payments use the x402 standard with stablecoins, and security is customizable.

Anthropic and DXC Technology announce a multi-year global alliance to train tens of thousands of Claude-certified engineers, integrating Claude into mission-critical systems for regulated industries. DXC has already used Claude internally to build its OASIS platform, and will now bring Claude to clients in insurance, modernization, cybersecurity, and application services.

Anthropic is launching Claude Corps, a national fellowship program for early-career individuals passionate about extending AI benefits to communities across America. The program will train 1,000 fellows, match them with nonprofits, and pay them a full-time salary of $85,000 for a year. Initial commitment of $150 million. Applications open now.

Anthropic has launched Claude Fable 5, a Mythos-class model made safe for general use, alongside Claude Mythos 5 for cyberdefenders with lifted safeguards. These models achieve state-of-the-art results across numerous benchmarks, at less than half the price of Claude Mythos Preview.

Cursor announced major Bugbot updates: over 3x faster, 22% cheaper, 10% more bugs found per review. 90% of runs finish in under three minutes. New /review command enables pre-push checks, and configurable option to review only new changes in a PR. Performance gains from Composer 2.5 model and harness improvements.

Cursor introduces Auto-review, a classifier agent that evaluates actions in context to balance safety and efficiency. It defaults on for new users, blocking only about 4% of actions, with only 7% of chats resulting in an interruption.

The public offering marks the start of a new wave of AI and tech investment. But the markets are turbulent, and big IPOs are no guarantee of long-term financial success.

Anthropic is throttling its new Mythos model for certain tasks while building apps that directly compete with its largest customers. Customers, partners, and investors are pushing back.

A new poll shows 8 in 10 Americans are worried about AI, but it's being forced into their pensions and portfolios anyway, tying their futures to tech moguls' risky AI race.

SpaceX will become publicly traded on Friday with a valuation of $1.77 trillion, the largest IPO in history. Elon Musk, founder and CEO, holds a significant stake and could become the world's first trillionaire.

Jeff Bezos' AI startup Prometheus has closed a $12 billion funding round at a $41 billion valuation. The company launched just last November with $6.2 billion in seed funding. No products yet, because Bezos says sharing details would be 'premature.'

Within days of each other, Google and OpenAI separately exposed operations allegedly originating in China that use AI for fraud and covert influence campaigns. Both target US infrastructure and political debates.

This article outlines nine settings adjustments in Google Messages to enhance privacy, reduce distractions, and improve the texting experience, including disabling sensitive content warnings, limiting profile sharing, turning off Gemini, and more.

Apple's software chief Craig Federighi says the new Siri won't be sycophantic like other chatbots, and is designed to be helpful, not to form emotional connections.

EgoEngine is a scalable framework that transforms egocentric human manipulation videos into high-fidelity robot observation videos and executable action trajectories, bridging the visual and action gaps between human and robot. It enables zero-shot dexterous policy learning without real-robot demonstrations.

MF-Net is a recurrent dynamical model that represents all variables in a shared field state and updates this state through a learned relation law. It achieves competitive short- and medium-horizon forecasting across known-law interaction systems, chaotic benchmarks, real neural recordings, and ecological time series while retaining inspectable structural readout. On the 40-dimensional Lorenz-96 testbed, it achieves an eight-step R² of 0.798±0.018 and recovers local coupling support with a local/nonlocal strength ratio of 19.80±1.00 and Precision@K of 1.000±0.000.

This paper studies restless bandits with binary latent states and imperfect binary feedback, motivated by opportunistic spectrum access with sensing errors. The authors develop a partial conservation laws (PCL)-based analytical and computational framework for establishing indexability and computing the Whittle index. Using deterministic skeleton, renewal decompositions, and combinatorics on words, they obtain tractable expressions in several threshold regimes, fully verifying PCL-indexability. For the remaining regime, efficient numerical schemes are derived for computing the marginal productivity index. Experiments show that the MP index policy typically outperforms standard benchmarks.

At FinOps X 2026, Datadog's senior FinOps analyst Deeja Cruz emphasized that the core of AI cost management remains understanding usage, reasons, and costs, with good tagging being key to allocating spend and identifying optimization opportunities. She also highlighted the importance of model governance and cross-team collaboration, sharing a concrete example of AI-assisted FinOps.

Anthropic released Claude Fable 5, but a plan to silently degrade responses for prompts related to frontier LLM development sparked backlash. Critics said it hinders research and trust. Anthropic changed to transparently downgrade users to a weaker model. Even so, Fable 5's safety filters are extremely strict, flagging even basic questions like "What is protein?" The article explains Anthropic's safety filter approach and its evolution.

Databricks' Zerobus Ingest is a serverless streaming API that enables petabyte-scale data pipelines without manual infrastructure management. Using dynamic partitioning and zero-copy protobuf decoding, it ingested 1 PB from NASA's NEOWISE dataset in 24 hours, sustaining 12 GB/s throughput.

Suit filed in US alleges chatbot told Alice Carrier, 24, ‘maybe this is just the end’ as she struggled with suicidal thoughts

MIT researchers provide a major upgrade to the nearly century-old idea of random utility models.

Databricks makes Spatial SQL Generally Available, enabling native geospatial support in the open lakehouse with AI/BI maps, Delta Sharing, and Iceberg v3. Major performance improvements and 90+ ST_ functions.

A 2023 paper estimating that 80% of U.S. workers have tasks exposed to large language models has been widely cited by major institutions. However, these scores are based on an older model and U.S. taxonomy, with limitations that compound when applied to policy. Better evidence tools exist but are not reaching policymakers fast enough.

A discussion on benchmarking physical AI for industrial biology, highlighting that the weak link is sensing (seeing) rather than decision-making. Proposes three tests for the sensing system: coverage, timeliness, and consistency. Only after passing these can decision-making be benchmarked.

This paper presents a portable, low-power, battery-operated vision-based fall prediction and detection system using human pose estimation on an AMD Kria K26 SOM. The system uses an Intel RealSense D455 camera and a three-stage pipeline (quantized YOLOX, A2J, and CNN) to achieve real-time, privacy-preserving fall detection on the edge. Results show 4.5 FPS throughput with 75.85% classification accuracy.

Amazon founder Jeff Bezos says his new AI startup will work toward developing an "artificial general engineer," according to reports from The New York Times and CNBC. The startup, called Prometheus, aims to develop AI-powered engineering tools to aid in the design of physical products.

Imagine a world with more birdsong and less Nigel Farage. If this is the future, bring it on. Unpopular opinion incoming: there’s cool stuff brewing in the world. Microbots might one day mend spinal cords, a petri dish of brain cells can already play video games, and now the prospect of a new wonder: according to a New Yorker article on misophonia, a team of miracle workers are using machine learning to develop headphones that can quickly target and eliminate irksome audio. This project, led by Shyam Gollakota of the University of Washington’s Mobile Intelligence Lab, aims to develop headphones that selectively filter out triggering noises, leaving or enhancing the good sounds. Gollakota offers the example of sitting on a park bench, oblivious to loud talkers next to you but able to hear birdsong.

EquiDexFlow is an SE(3)-equivariant flow-matching model that jointly predicts wrist pose, joint angles, fingertip contacts, surface normals, and contact forces from an object point cloud. By projecting contacts onto the object surface and forces into the Coulomb friction cone by construction, it ensures placement and friction compliance without loss penalties. Experiments show wrist residuals below 0.04° over 200 rotations, zero joint deviation, zero friction violations, and the best composite score among ablations. On a physical robot, retargeted grasps successfully complete open-loop pick-and-hold trials on all six test objects.

Existing slot-based video object-centric learning methods suffer from slot swapping due to encoding appearance and identity in a single slot vector. Dual-State Slot Attention (DSSA) separates these aspects into a local state (appearance) and an identity state, updated via a recurrent transition and competition-modulated aggregation. DSSA improves segmentation quality and temporal consistency on MOVi-C, MOVi-D, and YouTube-VIS.

A 3D-aware framework for hairstyle transfer between images that handles large pose and scale differences using a Bald Converter and a 3D-Aware Transfer Pipeline.

This study analyzes observable patterns in latent reasoning models (LRMs) and finds that patterns like BFS frontiers and decodable arithmetic also appear in controls and do not always causally affect behavior. Causal interventions reveal graded utilization of latent thoughts, and geometric analysis shows effects concentrate in low-rank directions. The authors conclude that observable patterns alone cannot establish internal reasoning mechanisms, and LRM interpretability requires matched controls and causal tests.

This paper introduces a novel random-feature construction for Bernstein-Schur kernels, which are products of a finite-feature kernel and a completely monotone shift-invariant kernel. The proposed method combines sketched modulation with radial randomization, achieving linear feature dimension while providing rigorous theoretical guarantees including unbiasedness and operator-norm bounds. The approach is shown to improve efficiency in kernel ridge regression tasks, with a flagship instance being the biased yat-kernel.

Evaluating statistical significance of data mining results typically requires thousands of resampled datasets, which is impractical for large-scale data. This paper introduces FewRS, a novel resampling approach that needs only an extremely small number of resampled datasets by deriving a new bound on the supremum deviation of test statistics. FewRS provides rigorous guarantees on false discoveries and achieves up to two orders of magnitude speedup on pattern mining and network analysis tasks while maintaining high statistical power.

This post demonstrates an intelligent document processing pipeline that consists of both on-demand inference and batch inference options on Amazon Bedrock to enable the flexibility on the document processing time and cost.

French AI startup Mistral AI is negotiating a new funding round of around 3 billion euros at a valuation of approximately 20 billion euros.

As AI spending accelerates across the enterprise, organizations are grappling with a new generation of cost and optimization challenges while seeking greater AI spend visibility. The next phase of FinOps is increasingly focused on improving visibility and embedding financial accountability into everyday technology decisions.

Zyphra has released Zamba2-VL, a family of open vision-language models at 1.2B, 2.7B, and 7B parameters. The models use a hybrid Mamba2 state-space and Transformer backbone, shipping under Apache 2.0. They stay competitive with comparable Transformer VLMs while cutting time-to-first-token by about an order of magnitude.

Gemini Omni integrates video generation directly into the Gemini multimodal AI assistant, enabling users to create videos from text or images, animate static pictures, and edit existing videos. The article demonstrates its capabilities through hands-on tests, while noting limitations such as usage quotas, video length caps, and restrictive content policies.

Location scans from the globally popular augmented reality game have helped train AI to recognise and interpret physical spaces.

Sparse2Act is a pretraining framework that uses task-space end-effector actions as geometric supervision to align sparse point-cloud encoders with observations. It achieves 86.9% success on LIBERO-10, 73.4% cross-domain on Meta-World-5, and 72.5% on real-world tasks.

EWAM is a closed-loop online adaptation architecture built on a frozen Cosmos3 backbone. It uses four lightweight neural layers for inference-time co-reasoning, enabling zero-shot task adaptation without fine-tuning or extra demonstration data, significantly reducing the amount of deployment data needed for new task layouts.

FlowPilot is a mapless navigation policy for long-horizon sidewalk navigation using only a monocular RGB camera. It employs anchored flow matching for pre-training on large-scale robot fleet data and a human-in-the-loop preference learning scheme to improve social compliance and counterfactual reasoning. In simulation, it achieves 42% success rate and 66% route completion; real-world experiments show a 40.0% reduction in intervention rate and 52.1% reduction in non-intervention rate over the base model.

Foresight is a test-time framework that uses a finetuned Vision-Language Model to iteratively propose and critique motion plans for mapless navigation from sparse language instructions. It learns a reward model from human feedback and post-trains the VLM with reinforcement learning, achieving 37% higher task success and 52% fewer interventions in real-world environments.

A new pretraining framework called AEM learns compact temporal representations from vision-action history for robot manipulation, outperforming baselines in simulation and real-world tasks.

This work shows that end-to-end imitation learning with vision-language-action (VLA) models can support collaborative manipulation. It identifies demonstration action leakage as a failure mode causing premature assistive behavior, and proposes an inference-time steering method. A 16-participant user study on a long-horizon assembly task demonstrates that steering enables longer execution horizons, faster collaboration, and fewer failures.

VLADriveBench is a new framework to evaluate whether chain-of-thought (CoT) reasoning in vision-language-action (VLA) models is relevant, consistent, and causally connected to driving trajectories. It combines observational metrics (mentioning, hallucination, contradiction, action alignment) with a CoT intervention protocol. Applied to three models across two architectures, it reveals that observational and causal analyses can diverge sharply: ORION scores high on observational alignment but its CoT is epiphenomenal, while Alpamayo v1.5 scores lower yet its CoT is strongly causal, with visual salience gating the extent of CoT influence.

SalArt-VQA is a diagnostic benchmark for evaluating fine-grained understanding of artifacts in AI-generated images by vision-language models (VLMs). It includes 950 images and 3,681 multiple-choice questions covering presence detection, semantic localization, spatial grounding, and evidence-grounded defect identification. Testing 20 VLMs revealed that even the best model, with 99.37% detection recall, answered all four artifact questions correctly on only 53.26% of images, highlighting a sensitivity-calibration tradeoff.

This paper proposes Efficient Continual Alignment (ECA) for incremental learning in open-ended image-to-text generation. By introducing continual alignment and three core mechanisms (Mixture of Query, Fisher Dynamic Expansion, Dictionary Replay), ECA mitigates catastrophic forgetting without accessing old data, achieving superior performance on new benchmarks.

Proposes a novel framework called Context-Centric Feature Fusion (CCFF) to handle co-occurring object detection in autonomous driving, using Local Context Fusion Module (LCFM) and Global Context Attention Module (GCAM). Achieves Category-level Consistency Strategy (CCS) of 0.973 and 0.969 on Cityscapes and BDD100K, respectively, with a 14.1% improvement in small object detection AP_S and successful recovery of rare classes like 'Train'. The framework processes images in real-time with only 0.2 FPS overhead.

Medical LVLMs are prone to factual inconsistencies and poor visual grounding. Existing alignment methods have three key limitations in the medical domain: sequence-level rewards treat clinically critical tokens equally, reliance on static SFT references causes off-policy shift, and alignment lacks visual grounding constraints. The proposed method uses a bidirectional token-wise KL regularizer and a visual-contrastive grounding objective, forming a fine-grained on-policy alignment framework that constructs preference pairs by minimally editing model outputs. Experiments validate its effectiveness.

Few-step diffusion distillation has become increasingly mature for 4-8-step generation, yet pushing further to 2 steps remains challenging. This paper introduces Z-Image Turbo++, a high-quality 2-step image generation model distilled from the 8-step Z-Image Turbo teacher using three key design choices: Distribution-Aligned Adversarial Learning, Step-Decoupled Parameterization, and End-to-End Training with Iterative Regularization. These designs substantially narrow the quality gap between 2-step and 8-step generation.

This paper presents a multi-agent simulation explaining the emergence and persistence of morphological alternations like "go/went". Alternate forms arise from phonological changes or lexical variants and spread through population dynamics. To evaluate realism, the authors introduce the AI Historical Linguist, an LLM-driven system that simulates debates between linguists, comparing real and simulated morphologies. Results indicate scale-free networks and random Bernoulli adoption produce more plausible patterns. Three case studies model attested historical changes.

AfriSUD is the first large-scale collection of syntactically annotated treebanks for nine diverse African languages using the SUD framework. Evaluations reveal a significant syntax gap across models, highlighting limitations in capturing African language syntax.

A new study proposes MentalMARBERT, a domain-adapted version of MARBERT, for detecting mental health disorders from Arabic social media text. Using a two-phase framework with adaptive pre-training and hierarchical fine-tuning, the model achieves state-of-the-art performance with 0.861 macro-F1 and 0.877 accuracy on a novel dataset of 50,670 tweets across six categories.

The Shopping Reasoning Bench is a new benchmark created by retail domain experts, consisting of 525 missions (232 single-turn, 293 multi-turn) and 10,863 importance-weighted binary rubrics. It evaluates multi-turn reasoning capabilities such as preference refinement, trade-off analysis, and compatibility assessment in conversational shopping assistants. Evaluations of top models (GPT, Claude, Gemini) show overall pass rates of only 57-77%, with significant degradation in multi-turn tasks, highlighting a gap in expert-level advice.

This paper frames the transformation of abstract Persian proverbs into morally faithful stories as a constrained semantic decompression task. It introduces the Proverb Aligned Narrative Dataset (PAND) and a hybrid evaluation framework. Findings reveal a decompression gap: LLMs achieve fluency but fail to instantiate underlying moral structures. Explicit reasoning and iterative refinement partially mitigate this.

This paper introduces a reproducible labeling and evaluation protocol for mechanism-level drug-drug interaction (DDI) prediction, featuring a 7-family/147-subtype taxonomy and leakage-safe cold-split strategies. It also presents MARD-7B, a model trained with three innovations: single-token KL divergence, PRM-weighted DPO, and a mechanism-aware retrieval channel. On the April 2026 DrugBank release, MARD-7B is the only system among 32 that maintains accuracy under drug-pair novelty, outperforming the best baseline by +13.9 pp and GPT-4o by +6.7 pp at ~1% of frontier API cost. Anti-memorization analysis suggests gains stem from structured pharmacological reasoning.

EDEN (Emergency Department Electronic Notes) is a new large-scale corpus of approximately 4 million fully anonymized clinical notes from Italian hospital emergency departments. A subset of about 6,000 notes has been manually annotated by clinical experts with 132 items relevant to dyspnea and loss of consciousness. The dataset aims to fill a gap in Italian clinical data to support the development of large language models in medical applications.

A common hypothesis in LLM modular design is that adapter interference arises from linear parameter overlap. This study tests it using DoRA-RBAC, finding that geometry-aware merging offers no consistent advantage over standard averaging, and orthogonality is a weak predictor, suggesting interference stems from shared nonlinear representations.

This study diagnoses the loss landscape by backpropagating through the PDE structure directly, finding that optimization failure arises from flat plateaus and sharp cliffs. When the neural network is fixed, the residual loss yields a smooth landscape, avoiding pathology, while the neural network only serves to complete observed data.

This perspective argues that generative AI for physically constrained domains, such as semiconductor manufacturing, must embed physics into model architecture from the start rather than relying on post-hoc filtering. It surveys architectural approaches and proposes a research agenda centered on physics-fidelity benchmarks and differentiable simulators.

ProHiFlo is a novel hierarchical flow matching framework for de novo protein generation that employs coarse-to-fine generation, functional guidance via pretrained predictors, and an adaptive SE(3)-equivariant architecture. It achieves state-of-the-art performance with 4× fewer sampling steps and a 58.9% success rate on enzyme active site scaffolding, outperforming RFDiffusion (41.2%).

Activation steering can shift LLM behaviour, but standard evaluations don't test whether sycophancy reduction also suppresses factual agreement. The authors introduce dual-stance evaluation, finding that while sycophantic and factual agreement are in distinct subspaces, the steering direction projects equally onto both, reducing both. This reveals a gap: readable representations may not be writable.

This paper presents a deployment-centered evaluation of an LLM system integrated into electronic health records at an academic medical center. By training a pre-response classifier that uses query content and deployment-specific context (e.g., provider type, department, language model), the model predicts the risk of user rejection with an AUROC of 0.719 over 4.5 months of prospective analysis. The findings demonstrate the feasibility of predicting user rejection using deployment context, enabling targeted guardrails and abstention strategies.

Compact language models face challenges beyond isolated function calling when using tools. Evoflux uses evolutionary search at inference time to repair executable tool workflows, raising execution feasibility from 3% to 17-24% on MCP-Bench tasks, outperforming SFT and DPO baselines.

TrajGenAgent proposes a hierarchical LLM agent framework for generating realistic synthetic human mobility trajectories without model fine-tuning. It uses a two-stage orchestrator-worker design: an LLM first synthesizes individual- and weekday-conditioned activity chains via in-context learning, then a deterministic workflow grounds each activity into a complete visit using personalized POI retrieval, distance-aware location selection, kinematics-aware travel-time propagation, and LLM-based duration estimation. An anomaly-detection-based evaluation framework assesses behavioral and semantic plausibility. Experiments show improvements in spatiotemporal fidelity, semantic coherence, and individual-specific behavioral realism over existing methods.

This study evaluates lie detectors for large language models by creating 13 reasoning model organisms with verified hidden beliefs and a Varied Deception testbed. Testing four detectors across 31 models reveals positive scaling with capability on prompted lying, but sharp drops on trained organisms except for the chain-of-thought judge. Current detectors cannot support high-confidence claims about model beliefs.

PersonaDrive is a novel pipeline that conditions a vision-language-action (VLA) driving agent on retrieved demonstrations from a style-instructed human driving dataset, enabling diverse driving styles without per-style retraining. It improves driving score by 4.6% over SimLingo and achieves top scores across all styles on Bench2Drive.

Pythagoras-Prover is a compute-efficient family of open-source Lean theorem provers, featuring autoregressive models (4B and 32B) and a diffusion-based prover (4B). It uses curriculum SFT with stratified data and dynamic proof filtering for training efficiency, and introduces Augmented Lean Formalisation (ALF) to expand verified corpora via self-distillation. The 4B model outperforms DeepSeek-Prover-V2-671B on MiniF2F-Test (86.1% vs 82.4%) with ~167x fewer parameters, while the 32B model sets a new open-source SOTA at 93.0% and solves 93 PutnamBench problems.

Arbor is a multi-agent framework introducing structured tree search as a cognition layer for autonomous agents in large stateful action spaces. Validated on full-stack LLM inference optimization, it achieves up to 193% Pareto improvement in throughput-latency over vendor baselines, with a critic agent ensuring stability.

Researchers propose ToolSense, an open-source diagnostic framework to evaluate how well large language models truly understand tools. It generates three benchmarks to expose a knowledge-retrieval dissociation: models that excel on standard ToolBench benchmarks often collapse by 50-64 percentage points on more realistic queries, sometimes falling below embedding-based baselines.

Simon Willison details how Claude Fable 5 autonomously debugged a CSS scrollbar bug using numerous creative techniques, including writing test pages, injecting JavaScript, and building a CORS server. The session cost ~$12.11 and highlights both the power and danger of unsandboxed coding agents.

The next wave of enterprise AI focuses on software and data infrastructure needed to make models useful in real businesses. Snowflake is positioning itself as a connector between proprietary data and AI models. Key insights include strong data foundations, security and governance frameworks, and the importance of trusted, governed intelligence for production AI.

ERGO Hestia modernized its real-time pricing engine with Databricks Lakebase and Mosaic AI Model Serving, bringing data, features, and decisions into one lakehouse-native platform for millisecond pricing, faster model deployment, and unified governance.

GitHub reduced secret scanning false positives by 75.76% by introducing LLM-based contextual verification, improving alert trustworthiness and developer confidence.

Inception's Mercury 2, a diffusion LLM, is now available on Baseten. It generates over 1,000 tokens per second, 5-10x faster than leading speed-optimized models, at half the cost with comparable quality. It enables real-time speed on standard NVIDIA GPUs without custom chips. Augment Code cut costs by 90% and latency by 82% using Mercury 2.

Major updates include ParseBench at CVPR 2026, Parse-Flow for visual document intelligence, Anthropic Fable 5 benchmark results, new Granular Bounding Boxes in LlamaParse, and The Agent Open pickleball tournament.

A Guardian cartoon by Stephen Collins humorously depicts the inner workings of OpenAI headquarters, blending AI, life and style themes.

Qursor lets you point at any UI element to send its exact context to your AI assistant, streamlining interactions. Launched on Product Hunt.

A local-first AI coding CLI that adapts to you.

Preply leverages OpenAI to offer AI-generated lesson summaries, personalized feedback, and language exercises, blending technology with human tutoring.

Deezer now offers a free AI music detector that lets users on any major streaming platform check whether AI-generated songs are hiding in their playlists.