🏢 Harvard University

PocketFlow: a novel generative model designs high-affinity protein pockets using prior-informed flow matching, outperforming existing methods.

FINCON: an LLM-based multi-agent system uses conceptual verbal reinforcement for superior financial decision-making, generalizing well across various tasks.

TRAC: a parameter-free optimizer conquering lifelong RL’s plasticity loss!

Eye-gaze data boosts medical image-text alignment!

New metrics reveal that generative models often possess surprisingly incoherent world models, despite seemingly accurate next-token predictions. This incoherence leads to fragility in solving related …

Generative models learn hidden capabilities suddenly during training, which can be explained and predicted using a novel ‘concept space’ framework that analyzes learning dynamics and concept signal.

Researchers dissected attention paths in Transformers using statistical mechanics, revealing a task-relevant kernel combination mechanism boosting generalization performance.

A new Covariate Shift Corrected Pearson Chi-squared Conditional Randomization (csPCR) test accurately assesses conditional independence even when data distributions vary between source and target popu…

Truthful comparison data is hard to obtain without ground truth. This paper presents novel peer prediction mechanisms using bonus-penalty payments that incentivize truthful comparisons, even in networ…

This paper presents efficient algorithms to detect whether an agent updates beliefs optimally (Bayesian) or exhibits bias towards their prior beliefs, using information design and signaling schemes.

This paper revolutionizes AI alignment by applying social choice theory axioms to RLHF, exposing flaws in existing methods and proposing novel, axiomatically guaranteed reward learning rules.

Latent Mutual Information (LMI) approximation accurately estimates mutual information in high-dimensional data using low-dimensional learned representations, solving a critical problem in various scie…

A lightweight knowledge alignment module enables two pre-trained LLMs to mutually learn and improve clinical language representation, exceeding individual model performance on various downstream tasks…

Soft labels, not sophisticated data synthesis, are the key to successful dataset distillation, significantly improving data-efficient learning and challenging existing methods.

Debunking a common myth, this paper proves that AUPRC is not superior to AUROC for imbalanced datasets, and in fact, can worsen algorithmic bias.