Summary LOTS uses observed tool outputs to learn reusable, task-specific tool selections that improve agent performance while reducing tool context.

Summary For-Value estimates training-data influence from a single forward pass, avoiding gradient computations for scalable LLM and VLM data valuation.

Summary LatentGym uses environments with known latent structure to measure how LLM agents gather information and adapt across related tasks.

Summary 2F2L combines linearized federated training with certified server-side unlearning to remove a client’s influence without additional client communication or historical model storage.

Summary TEP uses local prompt optimization to mitigate exploding and vanishing textual gradients in deep, long-horizon compound AI systems.

Summary A study of federated feature distortion shows how linear probing followed by fine-tuning can improve the balance between client personalization and generalization.

Summary KGGDG generates clinically plausible but incorrect answer choices from medical knowledge graphs to make medical QA benchmarks more challenging.

Summary FedTextGrad studies federated LLM prompt optimization with textual gradients, identifying aggregation challenges and improving prompt summarization.

Summary DARq and AdamR improve delta-parameter pruning under high pruning rates or large parameter changes for efficient fine-tuned model storage and deployment.

Summary Local Superior Soups improves client training and pre-trained model adaptation in federated learning through regularized model interpolation.

Summary DNE mitigates spurious correlations through noise editing for fairness-aware medical image processing with white-box and black-box foundation model APIs.

Summary FedSoup uses selective model interpolation to balance local and global performance, reduce overfitting, and seek flat minima for improved generalization.

Summary VITA generates diverse on-manifold samples through multi-source vicinal transfer to improve robustness to image corruptions.

Summary The first comprehensive study of five ReID corruption benchmarks compares transformer and CNN robustness and examines cross-dataset generalization.