theoretical framework
A structured approach that provides a set of principles and concepts to guide research and development in AI. It includes defining key notions, proposing models, and establishing theorems that help understand the performance and limitations of algorithms.
- A Statistical Theory of Contrastive Learning via Approximate Sufficient Statistics
- A unified framework for establishing the universal approximation of transformer-type architectures
- AdaptDel: Adaptable Deletion Rate Randomized Smoothing for Certified Robustness
- Belief-Calibrated Multi-Agent Consensus Seeking for Complex NLP Tasks
- Bilevel Optimization for Adversarial Learning Problems: Sharpness, Generation, and Beyond
- Can DPO Learn Diverse Human Values? A Theoretical Scaling Law
- Can Diffusion Models Disentangle? A Theoretical Perspective
- Concentration and excess risk bounds for imbalanced classification with synthetic oversampling
- Convergence Theorems for Entropy-Regularized and Distributional Reinforcement Learning
- Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?
- Escaping Collapse: The Strength of Weak Data for Large Language Model Training
- Evaluating the Inductive Abilities of Large Language Models: Why Chain-of-Thought Reasoning Sometimes Hurts More Than Helps
- FEAT: Free energy Estimators with Adaptive Transport
- Formal Models of Active Learning from Contrastive Examples
- Foundations of Top-$k$ Decoding for Language Models
- Generalization Guarantees for Learning Score-Based Branch-and-Cut Policies in Integer Programming
- Inference-Time Personalized Alignment with a Few User Preference Queries
- Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Model
- Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models
- Online Time Series Forecasting with Theoretical Guarantees
- Private Evolution Converges
- Self-Verification Provably Prevents Model Collapse in Recursive Synthetic Training
- Spend Wisely: Maximizing Post-Training Gains in Iterative Synthetic Data Bootstrapping
- Token Embeddings Violate the Manifold Hypothesis
- Train to Defend: First Defense Against Cryptanalytic Neural Network Parameter Extraction Attacks
- When Does Curriculum Learning Help? A Theoretical Perspective