10 selected stories for 2026-11-14, 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.
The attack began with an access key Cursor issued to Fireworks that Fireworks exposed, leading to exposure of API credentials for less than 0.01% of customers.
The attacker used stolen credentials to place malicious Claude Code extension builds in the marketplace cache; at most 5,000 users were potentially affected over roughly 11 hours.
GenPFN keeps TabPFN weights fixed and generates candidates purely through in-context learning from labeled evaluated designs.
Across fourteen problems it achieves higher class satisfaction and lower realism distance than adversarial, VAE, and diffusion generators, though VAE offers better coverage and diversity.
SynthCorp simulates business events on a virtual calendar to generate complex data estates comprising relational databases, document stores, and data catalogs.
The platform automatically generates analytical tasks and derives verified gold programs from a simulation log, eliminating manual annotation.
A NeurIPS 2026 MATH-AI workshop paper evaluates certify-or-decline proof-step verification on open-weight models. Certification boosts answer-key precision by 31–46 points on GPQA and AIME but only answers 35% and 21% of attempts; on near-ceiling MATH-500, gains are at most 5 points at a 13–27 point coverage loss. The authors also identify invalid proof construction as a major execution failure and caution that current audits lack human or formal-kernel ground truth.
This paper introduces a method to construct a θ-confidence Minimal Belief Set (MBSθ) from the probability outputs of pre-trained perception models like VLMs, enabling robust task planning under uncertainty without determinizing predictions or maintaining full distributions.
A new study uses a cooperative card game, The Mind, which forbids explicit communication, to compare how 150 humans and two classes of AI models (computer-use agents and standard text-prompted LLMs) coordinate using only temporal signals. Human behavior is best captured by a Bayesian strategy that infers a partner's belief from timing, and frontier models show similar patterns. However, unlike humans, none of the tested models efficiently adapt their behavior to partners over the course of the game. The findings suggest humans perform implicit probabilistic reasoning about continuous-time signals in dynamic social settings, and such mechanisms may emerge without real-time experience.
A new arXiv paper systematically studies three decisions for integrating LLM-provided semantic information into observational causal discovery: where to allocate semantic queries, which orientation proposals to admit, and which causal graphs to retain. Benchmarks show semantic information improves recovery over random orientations, but targeting unstable edges offers no established advantage over matched-random allocation; the tested admission certificate cannot support nontrivial admission with available calibration tasks; and learned references contain the true DAG in only 3 of 80 planned tasks.
Together AI introduces Together Compiler (TKCC), the first truly continual agent loop for AI inference optimization. It addresses the throughput-fidelity tradeoff with a Tiered Persistent Engine and a cost model, outperforming hand-written Megakernels by up to 1.7x on Llama 1B and achieving up to 2.6x higher TPS than the best open source engines on frontier models, with state-of-the-art batch-size-1 performance of 1111 TPS on M3.1 Flash and 1205 TPS on DeepSeek v4.1 Flash.
Amazon SageMaker HyperPod Data Capture records live inference inputs and outputs to Amazon S3. This post walks through one end-to-end example: capturing a support router's production traffic, surfacing a new-product blind spot with Amazon Athena, fine-tuning on the captured data with the HyperPod training operator, and measuring the before-and-after lift.
arXiv paper reveals that arena-style leaderboards using Bradley-Terry models are highly vulnerable to strategic voting, prompting, and nomination, which can boost a model's rank by up to 45 places.
A new arXiv paper shows through controlled experiments on AgentCE-Bench and AppWorld that editing an agent's history only helps when the real-world state changes along with it; making a history look clean while the error persists can leave a receiver agent worse off than no edit at all.
Zugwise reframes handwritten chess scoresheet digitization as game reconstruction, using chess engine machinery to detect silent legal-but-wrong errors and merging both players' sheets to reduce mistakes dramatically.
GS Caltex connects Amazon SageMaker Canvas to Amazon SageMaker Pipelines so process engineers' no-code models run in production. This post shows how ML engineers reuse validated preprocessing flows and models, cutting deployment from about a week to a few hours and running batch inference every 30 minutes.
OpenFold3-preview is now deployable from Amazon SageMaker JumpStart, letting users predict 3D structures of proteins, DNA, RNA, and ligand complexes inside their own AWS account. The post covers asynchronous inference, multi-GPU parallelism, bringing your own MSAs for accuracy and privacy, and scaling to zero to reduce cost.
AssemblyAI runs cost-optimized speech-to-text inference at scale on AWS using Amazon ECS, EC2 Spot GPU instances, and an SQS-based retry architecture, achieving 70% cost savings while maintaining SLA commitments across European Regions.