Search and filter through extracted claims from AI researchers.
Showing 1261-1280 of 2685 claims of type "fact"
ESPP evaluation method raises correlation with human judgment from Pearson r=0.716 to r=0.922 for generative UI quality assessment
Evidence-grounded personas with trait-derived social weighting better approximate diverse human judgment than single-pass or prompt-ensemble approaches
ViewMind3D enables 3D spatial reasoning over multi-view observations without requiring 3D-specific training, fine-tuning, or complete 3D reconstruction
In cross-teacher settings, even a stronger OPD teacher can yield little improvement over the SFT reference
Lightning OPD 2.0 with cross-fitted style residualization can separate useful context-specific teacher evidence from recurring differences in wording, formatting, and reasoning cadence
Autonomous systems experience agnostic collapse where mission failure arises from accumulated hardware degradation rather than single component faults
Graph Neural Networks can serve as data-driven preconditioners for solving large sparse linear systems, though the impact of imposing AMG-style hierarchy remains underexplored
Graph Neural Multilevel Preconditioner (GMP) that adopts AMG hierarchy as structural prior can work as a drop-in preconditioner for standard Krylov solvers on general sparse systems
A framework in Lean theorem prover can verify both formulation-level properties and solver-produced certificates to establish satisfiability without trusting external solvers
Chain-of-thought reasoning degrades label-token probabilities that automated triage relies on, requiring separate calibration
Training LLMs with chain-of-thought reasoning combined with automated prompt optimization, self-training, and RL achieves 82.6% accuracy on security alert triage with 43% improvement in benign recall over direct-label classifiers
VisualRouter's query-grounded visual sampling framework that classifies queries and applies corresponding sampling strategies can improve long video understanding
Vision-Language Models can be trained to reason about AI-generated image edits using reinforcement learning without explicit reasoning supervision
Group Relative Policy Optimization (GRPO) can incentivize models to reason by generating thinking traces prior to giving final answers
EndoCLIP foundation model trained on 125,756 lesion-level image-text pairs from routine colonoscopy reports outperforms general-purpose and biomedical vision-language encoders across multiple clinical tasks
EndoCLIP's linear probe approaches expert reader performance on benign-versus-malignant classification in blinded study with 12 endoscopists
Incremental Context Expansion (ICE) automatic relabeling methodology can identify the point where cumulative social media messages provide enough evidence of mental health disorders
Spanish domain-specific foundational models combined with ICE methodology improve state-of-the-art mental health detection while reducing detection latency
Large language models have a critical vulnerability called 'Salience Bias' where they become hijacked by useless explicit distractors and ignore implicit commonsense prerequisites
All 12 state-of-the-art LLMs tested suffer significantly from salience bias, with severity scaling with distractor density
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 8,986 pending.