Results for "theory discovery"

Learning from data by constructing “pseudo-labels” (e.g., next-token prediction, masked modeling) without manual annotation.

Minimizing average loss on training data; can overfit when data is limited or biased.

When information from evaluation data improperly influences training, inflating reported performance.

Breaking documents into pieces for retrieval; chunk size/overlap strongly affect RAG quality.

Techniques to understand model decisions (global or local), important in high-stakes and regulated settings.

Selecting the most informative samples to label (e.g., uncertainty sampling) to reduce labeling cost.

Structured dataset documentation covering collection, composition, recommended uses, biases, and maintenance.

Tracking where data came from and how it was transformed; key for debugging and compliance.

Observing model inputs/outputs, latency, cost, and quality over time to catch regressions and drift.

A broader capability to infer internal system state from telemetry, crucial for AI services and agents.

Time from request to response; critical for real-time inference and UX.

Stochastic generation strategies that trade determinism for diversity; key knobs include temperature and nucleus sampling.

Converts logits to probabilities by exponentiation and normalization; common in classification and LMs.

A model is PAC-learnable if it can, with high probability, learn an approximately correct hypothesis from finite samples.

A measure of a model class’s expressive capacity based on its ability to shatter datasets.

A measure of randomness or uncertainty in a probability distribution.

Measures how much information an observable random variable carries about unknown parameters.

Updating beliefs about parameters using observed evidence and prior distributions.

Estimating parameters by maximizing likelihood of observed data.

Optimization problems where any local minimum is global.

Optimization with multiple local minima/saddle points; typical in neural networks.

Neural networks can approximate any continuous function under certain conditions.

Fundamental recursive relationship defining optimal value functions.

Coordination arising without explicit programming.

Required human review for high-risk decisions.

Recovering training data from gradients.

Compromising AI systems via libraries, models, or datasets.

Embedding signals to prove model ownership.

Neural networks that operate on graph-structured data by propagating information along edges.

Predicting future values from past observations.