Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
FORTE is a production framework for feature lifecycle management in recommender systems addressing deprecation under feature interactions and data drift.
Gaussian process solvers for linear PDEs with inequality constraints (nonnegativity, monotonicity) via Borel measure conditioning.
Two-phase adaptive variance reduction algorithm for non-convex optimization without prior knowledge of smoothness constants or variance bounds.
Role-aware transformer for detecting sentiment, emotion, politeness, formality in dialogue via semantic role labeling.
Study of LLM sensitivity to discourse particles (just/only) revealing gaps in understanding subtle conversational context.
OccludeBench evaluates VLMs on inferring occluded objects; five models average 42.7% (direct) vs 62.7% (inferential) accuracy.
Randomized momentum transport optimization method reducing oracle calls via cheap predictors with importance weighting.
Optimal algorithms for nonsmooth nonconvex stochastic optimization achieving Goldstein stationary points with improved complexity.
Discrete graph diffusion models exhibit memorization on training graphs despite equivariance, measured via WL kernel similarity.
UkrMeme dataset (200 Ukrainian wartime memes) for persuasion detection; VLMs evaluated zero-shot on visual-textual propaganda.
ReplaySSM caches SSM inputs instead of state, reducing I/O bottlenecks in hybrid models and enabling cheaper speculative decoding.
Multi-objective optimization using norm-constrained oracles with entropic weight updates achieves O(T^-1/2) ergodic stationarity.
Information-theoretic bounds on adversarial reconstruction from DP-SGD models, computable pre-training independent of (ε,δ).
DualQuant enables calibration-free 4-bit weight-activation quantization for LLMs by jointly optimizing weight MSE and per-channel activation rescalings.
Geospatial foundation models (TerraMind, Prithvi-EO-2.0, DOFA, DINOv3) show significant accuracy drops under held-out disaster evaluation, revealing shortcut learning in damage assessment.
Joint Graph Diffusion with Latent Sociability Priors addresses fundamental limitations of Erdős-Rényi priors in discrete graph generation by incorporating latent node heterogeneity.
Prior-Data Fitted Networks (PFNs) decompose amortization error into prior, context, and deployment components; show residual errors persist in rare settings.
Moore sees a huge opportunity in consumer AI, particularly if the industry can tap into revenue streams beyond just subscriptions and API charges.
Matched-Context Group Policy Optimization (MCGPO) improves LLM agent credit assignment from sparse terminal feedback using context-matched action comparisons and rubric progress.
IntWorld and PermWorld are mathematical task ecosystems designed to study how neural networks learn under task diversity similar to modern language model training.
Critique-conditioned on-policy self-distillation for reasoning shows scoring signals fail to localize errors or provide corrective credit without structural modifications.
SmolWAM compresses Cosmos Policy to 2B parameters for edge robotic manipulation using latent world-action models with cached text features and reduced transformer blocks.
SafeDiSTL uses Signal Temporal Logic robustness for credit assignment in constrained multi-agent RL with joint safety violations, enabling decomposed constraint handling.
Orthros attention dynamically switches computation patterns across sequence processing phases to maintain in-context recall while reducing quadratic complexity bottleneck.
Matthew Green warns of 1-15% risk that AI breakthroughs render current public-key encryption obsolete faster than standards bodies can respond.
President Donald Trump has a knack for turning words against his enemies. His first successful presidential run was built on monikers like "Little Marco" and "Crooked Hillary"; he changed "fake news" from a phrase describing scammy media outlets to a derogatory term for the press at large. Over the past few weeks, he's clearly decided he can work the same magic to promote artificial intelligence - branding a technology he wants to accelerate "super", while turning "artificial" into his latest go-to pejorative and declaring resisters "THE ENEMY." Some of the biggest names in AI are going along...
Instinct’s agent is always just a text away. Before there were cute little guys, there was Instinct. In August, the startup got its AI agent to market with an unusual playbook: invite-only, no marketing, and barely so much as a website. And yet, Instinct quickly became the buzziest thing in AI, garnering praise for its straightforward, text message-based interface and its ability to handle chores like booking DMV appointments and sending follow-up emails. Then Muse arrived, followed not long after by Dots. The same products, more or less, from two far more powerful companies. With Big Tech pl...
Simon Willison built a blog feature using ChatGPT voice mode in local development, demonstrating hands-free coding via conversational AI.
RESCUE framework evaluates robot runtime monitors via counterfactual simulation to assess whether alarms enable practical recovery interventions.