🏢 Princeton University

This paper provides a tight characterization of inductive inference, proving it’s possible if and only if the hypothesis class is a countable union of online learnable classes, resolving a long-standi…

Vision-language models struggle with multi-object reasoning due to the binding problem; this paper reveals human-like capacity limits in VLMs and proposes solutions.

Researchers reveal how transformers learn word co-occurrence using a novel gradient flow analysis, uncovering a two-phase training process that leads to near-minimum loss and improved model performanc…

This paper presents an algorithm achieving Õ(√T) optimal regret for bandit non-stochastic control with strongly-convex and smooth cost functions, overcoming prior limitations of suboptimal bounds.

Revolutionizing machine learning, Schedule-Free optimization achieves state-of-the-art results without needing learning rate schedules, simplifying training and improving efficiency.

SWE-agent achieves state-of-the-art performance on software engineering benchmarks by creating a custom agent-computer interface that enhances LM agents’ ability to use computers.

SureMap, a new method, significantly boosts accuracy in single and multi-task disaggregated evaluations of AI models using limited data by transforming the problem into Gaussian mean estimation and cl…

SimPO: a simpler, reference-free reward algorithm significantly outperforming existing offline preference optimization methods, achieving higher accuracy and efficiency in aligning LLMs with human pre…

Probabilistic Federated Prompt Tuning (PFPT) significantly improves federated learning accuracy on heterogeneous and imbalanced data by using a probabilistic model for prompt aggregation, outperformin…

This paper introduces a novel method for creating highly accurate and narrow prediction intervals even when data distribution shifts unexpectedly, significantly improving machine learning model reliab…

One-layer transformers provably learn the one-nearest neighbor prediction rule, offering theoretical insights into their in-context learning capabilities.

SGD can train neural networks to learn low-dimensional polynomials near the information-theoretic limit, surpassing previous correlational statistical query lower bounds.

FRLC: a novel algorithm for low-rank optimal transport using latent coupling, enabling faster computation and better interpretability for diverse applications.

Aligning AI’s world representation with humans enables faster, safer learning of human values, improving both exploration and generalization.

Linear transformers efficiently learn sparse contextual bigrams by leveraging both in-context and global information, achieving polynomial sample complexity.

Kraken: A new Transformer architecture boosts multi-device inference speed by 35.6% by cleverly overlapping communication with computation.

Robots using LLMs for task planning often make unsafe or wrong decisions due to LLM hallucination and ambiguity in instructions. This paper introduces ‘introspective planning,’ a novel method that us…

Contrastive learning enables efficient probabilistic inference in high-dimensional time series by creating Gaussian representations that form a Gauss-Markov chain, allowing for closed-form solutions t…

GREATS: a novel online batch selection method significantly speeds up LLM training by greedily selecting high-quality data batches in every iteration, improving both convergence and generalization per…

This paper provides a novel optimization framework for guided diffusion models, proving Õ(1/K) convergence for concave objective functions and demonstrating structure-preserving guidance.