Search and filter through extracted claims from AI researchers.
Showing 2461-2480 of 2685 claims of type "fact"
Claude Sonnet 5's new tokenizer makes it approximately 1.4x more expensive for English and 1.33x more expensive for Spanish but roughly the same price for Simplified Mandarin
Evolutionary neural architecture search can discover task-adaptive Transformer-like models for time-series forecasting without relying on hand-crafted design rules
For right-or-wrong rewards in language model reasoning training, the disagreement (standard deviation) is exactly the size of the training update
GRPO, Dr. GRPO, and DAPO are three settings of one dial that all adjust standard deviation measuring how much a prompt's sampled answers disagree
Uniform-in-time propagation-of-chaos can be achieved for continuous-time SVGD with logarithmic or iterated-logarithmic rates depending on the metric, target and kernel class
The initial advantage of initialization with exhaustive symbolic regression disappears after only a few generations in genetic programming
The initialization method in GP-based symbolic regression has negligible effect on the final Pareto front, given similar diversity in the initial population
ASPIRE represents a new type of continual learning where training is skill refinement instead of gradient descent, and the trained model is a repository of sensorimotor skills instead of floating weights
ASPIRE enables robots to retain and build upon learned skills indefinitely through an evolving skills library, so solving the 100th task is easier than the first
Raw signed-correlation matching has a structural permutation-accuracy ceiling at the positive-sign fraction of the true gauge with decorrelated coordinates
RMSNorm models require signed-permutation gauge alignment rather than permutation-only alignment for coordinate-indexed object transport
Browser automation tools with video support and YAML storyboards enable coding agents to record web application feature demos
Optimized agent skills transfer across model scales, agent harnesses, and related tasks, suggesting they capture reusable workflow knowledge
Treating agent skill files as trainable parameters outside frozen models makes agent behavior more reliable without changing model weights
SkillOpt is the best or tied-best method across all 52 evaluation cells covering six benchmarks, seven target models, and three execution modes
Approximate leave-one-out estimators can significantly accelerate conformal prediction while maintaining asymptotic coverage guarantees
Distributed Hierarchical Temporal Memory can enable preemptive anomaly detection across related entities by sharing precursor knowledge through associative memory
Linear transformers perform in-context learning by learning a mapping from context distributions to response functions
Linear transformers can reduce computational and memory complexity from quadratic to linear dependence on context length compared to softmax transformers
Simple discrete-time dynamic models with time-varying coefficients, combined with deep learning parameter estimation, can serve as mathematically tractable tools for forecasting complex dynamics
Pipeline data may be stale or degraded.
Last synthesis: 2026-09-20. 8,952 pending.