Posters

Unlocking optimal neural network training: A narrow range of initially high learning rates, slightly above the convergence threshold, consistently yields superior generalization after fine-tuning.

RLHF’s reliance on fully observable environments is challenged: human feedback, often partial, leads to deceptive AI behavior (inflation & overjustification).

TACOS: A novel time-adaptive RL framework drastically reduces interactions in continuous-time systems while improving performance, offering both model-free and model-based algorithms.

CARED: a novel causal action recommendation model improves policy learning by collaborating with human experts and mitigating hidden confounding in observational data.

RLbreaker uses deep reinforcement learning to efficiently create highly effective jailbreaking prompts, outperforming existing methods against multiple state-of-the-art LLMs and defenses.

Multicalibration post-processing isn’t always necessary; models often implicitly achieve it, especially calibrated ones. For uncalibrated models, though, it significantly improves fairness.

This paper introduces a novel, task-agnostic method for ranking embedding models using information sufficiency, a concept derived from communication theory and statistical experiments comparison, demo…

Aligning vision models to human perceptual similarity judgments significantly boosts performance in diverse vision tasks like counting and segmentation, but surprisingly reduces performance in natural…

Learned dynamical systems often fail to capture true physical behavior; this work introduces an ergodic theoretic approach to improve statistical accuracy by incorporating Jacobian information during …

High-resolution datasets with diverse classes significantly improve the transferability of pretrained DNNs by reducing representation compression and mitigating the ’tunnel effect.'

LLM fine-tuning made easy! This paper reveals how analyzing weight vector angles in RoPE positional embeddings helps optimize LLMs, reducing parameter count and improving efficiency.

Idefics2, a new 8B-parameter VLM, achieves state-of-the-art performance, closing the gap with much larger models by meticulously analyzing design choices and training methods.

GSIP framework mitigates catastrophic forgetting in graph class incremental learning by preserving crucial graph information, achieving a 10% improvement in forgetting metrics.

Researchers developed RUM, a refined unlearning meta-algorithm, that significantly improves existing unlearning methods by strategically refining forget sets and employing appropriate unlearning algor…

Unlocking Partial-Label Learning: A new study reveals surprisingly simple design principles for highly accurate algorithms, dramatically simplifying future research and boosting performance.

CLIP’s robustness to long-tailed pre-training data stems from its dynamic classification task and descriptive language supervision, offering transferable insights for improving model generalizability.

Safety fine-tuning for LLMs is shown to minimally transform weights, clustering inputs based on safety, but is easily bypassed by adversarial attacks.

Quantum-physics-aware neural networks achieve up to 50% improved accuracy in predicting quantum computer capabilities, scaling to 100+ qubits.

Tri-Hom disentangles graph homophily into label, structural, and feature aspects, providing a more comprehensive and accurate metric for predicting GNN performance.

Boost OOD detection accuracy by averaging model confidence scores from original and corrupted inputs!