Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Hessian spectral properties are a standard tool in analysing neural-network training, with eigenvalues linked to sharpness, generalization, and optimization dynamics. Eigenvalues quantify curvature magnitude, while eigenvectors identify which parameters generate that curvature…
LLM evaluation and AI safety face a shared measurement problem: benchmark scores, reward-model signals, and reported safety metrics can improve while the latent properties they are meant to represent remain difficult to verify. This paper combines a hybrid survey - a systematic…
Proposes adaptive routing between draft and target multimodal models for efficient visual reasoning by learning query difficulty signals.
Introduces sparse cross-modality fusion for RGB-T object detection, reducing computational cost via selective thermal-visible integration.
Multi-center breast cytology dataset with 470 whole-slide images annotated for C1-C5 classification in pathology AI.
Philosophical essay framing data centers as the embodied substrate of modern AI systems using biological analogies.
SHOVIR benchmark evaluates whether radiology report generation models rely on image evidence or learned priors via spatially-grounded metrics.
Analyzes SONAR multimodal embeddings to detect encoding anomalies via dimension sensitivity analysis.
Sim-to-real transfer remains a major obstacle for reinforcement learning (RL), especially for vision-based control where image observations exacerbate the state-distribution shift between simulation and the real world. Domain adaptation (DA) is a promising remedy for this…
How does an agent that can tell self from world come to be durably shaped by that distinction? Recent work shows that a predictive system can detect its own agency (Ye, 2026), but detecting agency does not explain durable, self-shaped behavior. We show that agency-gated slow…
Existing domain-incremental learning (DIL) strategies call for massive amounts of data to adapt to new domains and suffer from the overfitting problem in the case of data scarcity. This paper puts forward a relatively uncharted problem, namely, few-shot domain incremental…
Current multimodal fusion approaches, particularly those based on static Mixture-of-Experts (MoE) architectures, often struggle to provide the adaptive and efficient collaborative reasoning required by complex real-world applications. We introduce the Dynamic Agent-based…
Dynamo enables frozen VLMs to evolve reusable skills and visual tools without retraining, improving reasoning on visual tasks.
MirrorCode benchmark evaluates AI coding agents on full program reimplementation tasks, extending beyond short coding exercises.
Cortex organizes web-scale training corpora via ontological graphs instead of flat filtering, improving LLM data pipeline structure.
NMO Benchmark evaluates generative molecular design on quantum materials science tasks, beyond traditional drug discovery benchmarks.
Federated learning framework uses CRF-based inference to weight heterogeneous client contributions for improved convergence.
PCHS applies physically-constrained signal processing to heart/respiratory rate estimation from wrist PPG under motion artifacts.
Framework disentangles grammatical gender from semantic bias in contextual embeddings for gendered languages like Spanish.
FacePlex generates speech and synchronized facial motion jointly in real-time for conversational avatars via full-duplex tokens.
DNA Language Models benchmark compares transformers (DNABERT2) and CNNs (ConvNova) on genomic tasks, questioning pretraining costs.
Warrant mechanism enforces that attention-retrieved values only contribute to predictions if warranted, addressing relevance-permission gap.
Extends strategic classification to handle decision-dependent manipulation costs, modeling how algorithm decisions affect the economics of gaming behaviors.
Proposes query-aware spreading activation for knowledge graph RAG to improve multi-hop retrieval by maintaining query signals throughout traversal.
Introduces hyper-network neural functional maps for 3D shape matching under partiality and topological noise via unsupervised non-linear alignment.
Empirical study across 25 LLMs showing chain-of-thought gains derive from semantic content, not token length, using in-distribution paired comparisons.
Extends gauge/gravity duality via PINNs to reconstruct holographic bulk potentials under large hierarchies and false vacua in quantum field theories.
Identifies three open problems in Constitutional AI preference reconstruction: principle composition, decision rule specification, and interpretable constraint encoding.
Proposes SA-VLA, a state-aware tokenizer for Vision-Language-Action models that conditions discrete action codes on robot proprioceptive state for improved manipulation.
Introduces AgentCanvas, a framework automating embodied agent architecture search via typed-graph runtime, extending prior text-domain AAS to perceptual agents.