Search and filter through extracted claims from AI researchers.
Showing 3681-3700 of 4537 claims
Typographic attacks on CLIP pose significant risks to safety-critical applications like autonomous driving
CLIP models exhibit a critical failure mode where irrelevant text in images confounds visual representations, biasing them toward lexical meaning rather than visual semantics
Symbol-equivariant RRMs exhibit improved extrapolation to larger problem sizes compared to standard recurrent reasoning models
Neural guidance with G-RRM improves symbolic solver search efficiency only when problem instances have expansive combinatorial search space and solver architecture can dynamically overwrite branching choices
Descriptor-free matching naturally enables multi-detector training as heterogeneous keypoints can be optimized in shared geometry-only space without aligning descriptor spaces
Descriptor-free visual localization accuracy still lags far behind descriptor-based pipelines due to insufficient geometric discriminability in geometry-only matching
Existing panoramic search methods rely heavily on fragmented local viewpoints and suffer from myopic, inefficient exploration due to rigid initialization and lack of global panoramic priors
Standard MLLMs struggle to effectively model panoramic properties like severe polar distortion and continuous cylindrical topologies, significantly degrading target detection accuracy in 360-degree environments
Acoustic information alone has a robust association with audiobook appeal, even after accounting for title effects
GACR framework with Observation-Anchored Residual Flow enables fast, stable, and faithful cloud removal reconstruction by anchoring generative trajectory to cloudy observation rather than pure noise
Existing cloud removal approaches for optical remote sensing prioritize visual realism while overlooking impact on downstream analytical tasks, leading to semantic drift and degraded performance
Policy-paced learning can regulate synthetic-data training through sample selection and scheduling to balance world-model quality with policy improvement in robot RL
World models grounded on real rollouts can generate high-fidelity synthetic transitions that greatly lower visual hallucination for robot RL data augmentation
Deploying reinforcement learning on real robots remains constrained by high interaction costs, since each physical rollout is costly and reflects only one realized action-outcome path
Generative drifting models can be adapted to accelerated MRI by training physics-conditioned drift fields that predict residual corrections and produce error maps
Global image metrics fail to capture blurred diagnostically relevant structures or reconstruction failures in accelerated MRI
Quantum state encoding and entanglement can effectively model multimodal sensor interactions for multi-agent activity recognition with drastically reduced parameters
Variational quantum circuit fusion modules can achieve 10x parameter reduction compared to classical MLP-based fusion in federated learning sensor fusion
Explicit, compositional and editable spatio-temporal scene graph representations can enable richer grounded activity understanding from first-person video
Existing approaches for understanding human behavior in embodied AI rely on implicit representations and disregard structured reasoning over scene dynamics
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 8,966 pending.