Results for "trial-and-error"

Inferring sensitive features of training data.

Extracting system prompts or hidden instructions.

Detecting unauthorized model outputs or data leaks.

Learns the score (∇ log p(x)) for generative sampling.

GNN using attention to weight neighbor contributions dynamically.

Controls amount of noise added at each diffusion step.

Diffusion performed in latent space for efficiency.

Maps audio signals to linguistic units.

Identifying speakers in audio.

Repeating temporal patterns.

Directed acyclic graph encoding causal relationships.

Simple agent responding directly to inputs.

Flat high-dimensional regions slowing training.

Restricting updates to safe regions.

Converts constrained problem to unconstrained form.

Model exploits poorly specified objectives.

Maximizing reward without fulfilling real goal.

Learned subsystem that optimizes its own objective.

Model behaves well during training but not deployment.

Prompt augmented with retrieved documents.

Control without feedback after execution begins.

System returns to equilibrium after disturbance.

Stability proven via monotonic decrease of Lyapunov function.

Finding control policies minimizing cumulative cost.

Optimal control for linear systems with quadratic cost.

Modifying reward to accelerate learning.

Sampling-based motion planner.

Ability to correctly detect disease.

Failure to detect present disease.

Agents have opposing objectives.