OVO-S-Bench hierarchically benchmarks multimodal LLMs on streaming video spatial reasoning, testing capabilities like depth, layout, and object-relation understanding over temporal sequences.

RAMP provides a runtime evaluation framework for assessing agentic AI models in live production environments, capturing failure modes invisible to static offline benchmarks.

Uber imposed usage caps on AI coding tools including Claude Code as a direct cost-control measure following higher-than-expected enterprise spending.

MCP tool integration was added to the Reachy Mini robot, enabling it to invoke external AI-powered tools via the Model Context Protocol.

Direct Preference Optimization techniques are applied outside conversational chatbot settings to align generative models in other domains such as code, images, or structured outputs.

OpenAI published its formal public policy agenda outlining its positions on AI regulation, safety standards, and government engagement priorities.

A governance blueprint proposes democratic oversight mechanisms—such as public participation and accountability structures—for decisions made about frontier AI development and deployment.

Wasmer engineers used OpenAI Codex to accelerate building a Node.js-compatible JavaScript runtime optimized for edge computing environments.

GPT-Rosalind receives new capabilities, likely expanding its functionality for biology or genomics-related AI tasks.

A year-long empirical mapping of AI-enabled cyber threats yields policy-relevant findings about attack patterns and defensive implications.

Anthropic launches a Services Track and Partner Hub to formalize and expand the Claude Partner Network ecosystem.

An analysis identifies the conditions and memory allocation patterns that trigger fragmentation in CUDA's caching memory allocator.

A method computes function vectors—representations capturing model behavior for specific tasks—more quickly while preserving faithfulness to the original approach.

AutoLab benchmarks frontier models on long-horizon automated research and engineering tasks, evaluating autonomous scientific problem-solving capability.

Language models are applied to automatically generate concise, relevant titles for research papers given their content.

A minimal-pair dataset probes whether language models can distinguish light verb constructions from full verb uses in matched phraseological contexts.

FoeGlass uses simple in-context learning to generate adversarial audio examples that fool deepfake detection systems during red teaming.

A method rapidly identifies genuine Roman-era gems from the RAPID collection using automated recognition or classification techniques.

A structured knowledge index is built for the Noah's Ark corpus, organizing and making its contents systematically queryable or retrievable.

Techniques or curricula for teaching arithmetic reasoning are developed or analyzed to improve language models' mathematical computation accuracy.

Leverages existing metadata (e.g., geolocation, timestamps, tags) as supervision signals to fine-tune vision foundation models without requiring manual annotations.

Extends disentangled representation learning frameworks to handle more than two modalities simultaneously, scaling the approach across richer multi-modal data combinations.

Assesses how accurately LLMs diagnose and make treatment decisions when presented with structured standardized patient case scenarios mimicking real clinical encounters.

Trains a model on egocentric video and speech data recorded from a child's perspective to study continual multimodal learning mimicking child development.

Applies set transformer attention mechanisms to graph-level tasks, enabling permutation-invariant aggregation over sets of graph elements for improved graph representation.

Introduces a model for processing and generating audio through interactive turn-taking or contextual audio exchange, enabling conversational or responsive audio understanding.

Extracts latent self-evaluation capabilities already encoded in base LLMs using minimal labeled examples, enabling calibrated judgment without dedicated RLHF-style training.

Separates appearance attributes from geometric structure within 3D Gaussian Splatting representations, allowing independent editing and optimization of geometry and visual appearance.

Applies fully homomorphic encryption to causal structure learning algorithms, enabling discovery of causal graphs over sensitive data without ever decrypting individual records.

Uses an LLM-driven agent to generate interpretable, evidence-grounded mobility predictions while reducing computational overhead compared to standard deep learning approaches.