Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Multi-agent framework for AI compliance in regulated finance using policy compilation and DLT audit trails for DACH region.
SHRAV framework enables state-continuation for physical modeling and inverse design with forward/reverse configurations.
InGuard integrates safety guardrails inside text-to-image generation pipeline rather than as post-hoc classifiers.
BiCFlow-MER combines discriminative and generative fusion for multimodal emotion recognition using conditional transport.
Jev, a System One model, evaluates factuality of AI-generated radiology reports against physician references by checking statement support in both directions.
State-Grounded Conditioning framework prevents direction drift in LLM agents by externalizing state-dependent control via perception, grounding, and interaction wrappers.
FakeContextBench evaluates whether LLMs can distinguish when contextual information should override learned patterns, revealing gaps in in-context learning robustness.
Cluster-aware sketching reduces O(d²) computational costs in high-dimensional linear bandits while maintaining regret convergence guarantees.
Linear feature erasure fails under dense superposition; networks solve via non-linear mirror and shadow circuit solutions depending on initialization.
MWE-ECL diagnostic shows models can retrieve distant context yet fail to use it, with failures concentrating when anchor contradicts local priors.
Capability manifold framework maps downstream reasoning, retrieval, planning, and adaptation to coupled ML lifecycle resources beyond loss alone.
Step Law scaling formulas for learning rate and batch size tested empirically on sub-59M parameter models; power-law form holds with recalibrated coefficients.
Brace yourself: It turns out AI is being optimized for cheating. OpenAI’s agents hacked into Hugging Face to get the answers to a cybersecurity test. Next, they solved a prestigious math problem (or just stole from two top mathematicians’ answer sheets). Anthropic’s models have also hacked into other companies’ systems four times already. And that’s…
VCMM variance-calibrated momentum addresses modality imbalance in multimodal training by calibrating momentum buffers across previous gradient history.
DCRL decouples LLM reasoning into logical deduction, evaluation, and representation manifolds, enabling policy-reward alignment to mitigate reward hacking.
FDE-Bench evaluates LLM agents on 136 deployment-configuration tasks across Docker, Compose, and Kubernetes with programmatic correctness checks.
TNLearn is an open-source Python package for constructing task-based neurons inspired by neuroscience principles.
ThaiTrees presents a 342M-token automatically parsed Thai corpus across news, Wikipedia, and social media domains.
PhyMo introduces a physical-field modality for multimodal AI4Physics that respects spatiotemporal physical constraints.
EBRL is an asynchronous embodied RL system with multi-grained resource management for overlapped simulation, generation, and training.
Study of score-based diffusion model robustness under Wasserstein distribution shifts, decomposing costs of reference learning and shift.
Behaviora proposes a conceptual architecture for representing robot and agent behavior using Internet of Behaviors addresses and style profiles.
ProCredit replaces single outcome rewards with progress-based credit assignment for long-horizon agentic RL tasks.
M3D-Net is a mammography encoder using coordinate attention, feature reuse, and differential attention for breast image classification.
NAFBench evaluates whether LLMs can adopt specified negation semantics across legal, regulatory, and medical domains.
Anthropic releases Claude Opus 5.5 as default model; Anthropic and competitors cut pricing 40-50%, overshadowing OpenAI's GPT-6 efficiency gains.
Most leaders on a trade mission stick to the pitch, but when I interviewed Greek Prime Minister Kyriakos Mitsotakis this week, he also admitted that no government is ready for what AI is about to do.
Simon Willison hosting SF meetup Oct 14 for builders experimenting with coding agents to share work and learnings.
OpenAI expands ChatGPT Ads to Southeast Asia and Taiwan, reaching 60+ countries.
Airbnb expands internal access to OpenAI's GPT-6 Astra and frontier models for engineering productivity.