Posters

Concept-Incremental Flexible Customization (CIFC) model tackles catastrophic forgetting and concept neglect in continually adapting text-to-image diffusion models, enabling flexible personalization.

Boosting, traditionally limited by assumptions about loss functions, is proven in this paper to efficiently optimize any loss function regardless of differentiability or convexity.

Deep network pruning’s fundamental limits are characterized, revealing how weight magnitude and network sharpness determine the maximum achievable sparsity.

MolPhenix, a novel multi-modal model, drastically improves zero-shot molecular retrieval by leveraging a pre-trained phenomics model and a novel similarity-aware loss, achieving an 8.1x improvement ov…

Boost ensemble accuracy by predicting performance with fewer classifiers using a novel polarization law and refined error bounds.

Transformers struggle with complex reasoning tasks. This paper introduces ‘globality degree’ to measure task difficulty and shows that high globality hinders efficient learning. However, using ‘induc…

Low reward variance drastically improves contextual bandit regret, defying minimax assumptions and highlighting the crucial role of eluder dimension.

Higher-order PDEs hinder Physics-Informed Neural Network (PINN) convergence; this paper provides theoretical explanation and proposes variable splitting for improved accuracy.

TAG-CF boosts collaborative filtering accuracy by up to 39.2% on cold users, using only a single message-passing step at test time, avoiding costly training-time computations.

CATY-IRL: A novel, provably efficient algorithm solves Inverse Reinforcement Learning’s scalability issues for large state spaces, improving upon state-of-the-art methods.

Neural networks learn features effectively through gradient descent, not just at the beginning, but also at the end of training, even with carefully regularized objectives.

This study reveals how black-box settings affect the learning guarantee of stochastic compositional optimization, offering sharper generalization bounds and novel learning guarantees for derivative-fr…

Pre-trained language models’ base capabilities are significantly influenced by architecture, not just scale; a novel Combination Enhanced Architecture (CEA) improves performance by addressing FFN-Wide…

LLMs surprisingly process multilingual queries via an English-centric intermediate stage before generating responses in the original language, a phenomenon explained by the proposed MWork framework an…

LLMs’ factual knowledge acquisition during pretraining is surprisingly non-linear: more data doesn’t guarantee better knowledge retention, and forgetting follows a power law.

Diffusion models surprisingly learn factorized representations, enabling compositional generalization, but struggle with interpolation; training with independent factors drastically improves data effi…

TextGen enhances multilingual visual text generation and editing by optimizing control information using Fourier analysis and a two-stage framework, achieving state-of-the-art results.

HORSE: A novel attention-based neural network significantly improves large-scale neural subset selection by up to 20%, addressing limitations in existing methods.

HOPE: a novel shape matching method achieving both accuracy and smoothness by aligning different k-hop neighborhoods and refining maps via local map distortion.

HonestLLM boosts LLM honesty & helpfulness by 65.3% (Llama3-8b) and 124.7% (Mistral-7b) using training-free and fine-tuning methods, establishing principles and a new dataset (HONESET) for honesty eva…