The Archive

Variance-optimal off-policy evaluation for contextual bandits interpolates between doubly robust and cluster-level weighting under local reward model correctness.

Evaluation of Jev evidence verifier for financial LLM outputs reveals citation misattribution risks when numerically-identical cells enable paraphrased wrong-role claims.

Pre-registered study of 248 scenarios shows post-training determines whether LLM agents act against stated moral judgment under social/authority pressure.

MemFLoRA optimizes CNN fine-tuning at the edge by reducing activation memory overhead rather than parameter count during backward passes.

Semantic behavioral watermarking embeds provenance in LLM agent actions, resisting observation paraphrasing and defending against trajectory forgery attacks.

ParanoiaEval benchmark systematically evaluates risk-treatment capabilities in coding agents using software engineering risk management frameworks.

Neuro-semantic verification framework ensures biomedical AI explanations in glioblastoma radiogenomics are grounded in patient-specific evidence.

Selective RL update transfer improves model merging efficiency for reasoning capability transfer to vision-language models.

DireSMC uses sequential Monte Carlo to efficiently sample rare events from diffusion model surrogates in weather and materials simulation.

Case study documents production healthcare platform governance using coding agents and human-led meta-agent supervision without formal engineering background.

SquidAgent addresses parallelization overhead in multi-agent LLM systems by reducing context re-exploration and coordination costs.

Mirror Particle will launch at TechCrunch Disrupt's Startup Battlefield 200 with a world model built from scratch to predict human behavior, arguing that LLM role-play falls short for market research and brand strategy.

HygieneRoboBench evaluates household robot planners on contamination-aware task execution across 624 instances and 134 task families.

With AI hardware, tech companies are pushing the definition of what does and doesn't constitute a recording. For most of gadget history, it'd be reasonable to assume that a device with a microphone or camera is either recording you or it isn't; it's either on or off, without much gray area. Microphones capture sound. Cameras capture images. That data is saved onto a device. Sometimes, it's also sent to the cloud - hopefully with end-to-end encryption - for processing. That data is a preserved instant in time and is therefore a recording. Perhaps not anymore. Last week, Bloomberg's Mark Gurman...

In-parameter memory augmentation stores post-pretraining knowledge in model parameters and adapters as alternative to context-window learning.

Parallel Predictive World Models enable long-horizon planning by predicting trajectories in parallel while preserving causal constraints.

Feature information dynamics framework uses I-MMSE identity to quantify when features emerge during diffusion model generation.

Method for selecting Adam optimizer memory parameter β from pilot training via local gradient modeling.

E4 system uses constrained agentic workflows for explainable root-cause analysis in internet-scale services.

Recursive Game Creator orchestrates Designer, Builder, Player, Reviewer agents to iteratively develop entertaining games.

GPU-accelerated Kubernetes clusters depend on compatible versions across dozens of components, each on its own release cycle: host kernels, GPU drivers,... GPU-accelerated Kubernetes clusters depend on compatible versions across dozens of components, each on its own release cycle: host kernels, GPU drivers, container runtimes, networking, storage, operators, and workload frameworks. A configuration that works for one service, GPU generation, and Kubernetes release may silently fail for another, and tracing version conflicts after deployment is… Source

InterCorrect merges demographic-specific ASR models with intersection-aware corrections for fair speech recognition.

CropSentry multi-robot system detects crop stress using multimodal leaf sensing and swarm coordination.

Experiment evaluates generative AI translation of emergency earthquake instructions into Chinese/Spanish for safety-critical messaging.

CMDS coordinates frozen pretrained diffusion models via stochastic optimal control for coherent structured output generation.

ResNet-BiLSTM deep learning model detects multiple perceptual bugs in video game footage.

"We created this program because we believe the benefits of AI will reach most people through the companies that build on top of models, rather than through the models alone."

Atlassian and OpenAI expand partnership to integrate frontier models with enterprise knowledge for team workflow and project delivery.

CNet C++/CUDA framework for training complex-valued neural networks with Wirtinger autodifferentiation.

MINDSET energy-based schema for long conversational agent memory that tracks stale/active knowledge and evolves context over time.