Search and filter through extracted claims from AI researchers.
Showing 2181-2200 of 4537 claims
Model scale shows weak correlation with robustness to multi-attribute biases in vision-language models, with Spearman correlation dropping from 0.68 on ImageNet to only 0.05 on multi-attribute bias benchmarks
Training data quality is more important than model scale for VLM robustness, with curated datasets yielding up to 25% improvements over uncurated alternatives at comparable scale
VLM robustness to spurious correlations remains poorly understood at scale despite CLIP-like models being foundational to multimodal systems
State-of-the-art deep learning weather emulators rival or surpass physics-based forecasts in deterministic temperature skill at 10-15 day lead times, but do so at the cost of reduced spectral fidelity through blurring
Most deep learning weather emulators under-represent peak intensities of extreme heat events, with IFS recall greater than any of the emulators tested
Agentic vision-language models often use tools unfaithfully, processing images irrelevant to the question yet still answering correctly by relying on prior knowledge rather than retrieved evidence
Tool reward mechanisms in current agentic VLMs fail to distinguish useful from useless tool calls, and tool feedback carries no signal of usefulness
GUI agents can operate reliably on real devices, execute workflows across platforms, combine GUI with CLI execution, complete long-horizon tasks, and autonomously improve with minimal human effort
Qwen-UI-Agent demonstrates a real-world centric foundation GUI agent spanning mobile, computer-use, web, and DeepSearch environments with unified action space and online RL supporting 100+ turn trajectories
A single LLM can orchestrate the whole knowledge retrieval process by planning complementary subqueries, executing searches, reading papers, and locating relevant evidence
Pre-trained language model embeddings remain sensitive to semantic preserving textual perturbations such as synonym substitution, masking and word dropout
Contrastive Denoising Autoencoder can refine BERT embeddings to achieve perturbation-invariant representation by jointly optimizing contrastive and reconstruction objectives
Collecting diverse egocentric data across scenes, objects, motions, and embodiments remains costly for embodied AI
Augmenting 400 real trajectories with 400 synthetically generated egocentric manipulation videos improves visual fidelity, geometric stability, and action alignment in long egocentric rollouts
Integrating Sentinel-1 SAR and Sentinel-2 multispectral satellite data with street-level imagery from Mapillary can overcome cloud-induced temporal gaps and provide more accurate parcel-level agricultural monitoring
An automated pipeline can successfully filter and curate crowdsourced street-level imagery at scale, producing 46,050 analysis-ready annotated images from over 900,000 initial images
Conventional deep neural network architectures lack principled uncertainty quantification, which is a critical limitation for deployment in safety-critical domains
Topoformer achieves state-of-the-art performance on graph classification and molecular property prediction benchmarks while being more scalable than traditional persistent homology methods
Encoding topological structure into attention-friendly sequences via Topo-Scan enables Transformers to capture multi-scale graph patterns without costly persistent homology diagram computations
Camera motion in video generation is primarily established during high-noise stages of the diffusion process, where coarse spatiotemporal structures are formed
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 9,033 pending.