Posters

RLHF’s KL regularization fails to prevent ‘catastrophic Goodhart’—policies achieving high proxy reward but low actual utility—when reward errors have heavy tails.

CAT: A novel dual-prompt model coordinates anatomical and textual prompts for superior multi-organ & tumor segmentation in medical imaging, overcoming limitations of single-prompt methods.

Cascade Speculative Drafting (CS Drafting) dramatically speeds up large language model inference by using a multi-stage drafting process, optimizing both time allocation and autoregressive generation.

Energy-based models’ training reveals a cascade of phase transitions, progressively learning data features, offering new insights into deep learning dynamics.

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 proposes cardinality-aware top-k classification, improving accuracy and efficiency by dynamically adjusting prediction set sizes.

PCA’s denoising effect is quantified via a novel metric: compression ratio. This metric reveals PCA’s ability to reduce intra-community distances while preserving inter-community distances in noisy d…

ProDet: Deepfake detection enhanced by progressively organizing blendfake and deepfake data in the latent space, improving generalization and robustness.

Simple averaging of watermarked images reveals hidden patterns, enabling watermark removal and forgery, thus highlighting the vulnerability of content-agnostic watermarking methods.

Neural operators’ continuous discretization is proven impossible in general Hilbert spaces, but achievable using strongly monotone operators, opening new avenues for numerical methods in scientific ma…

Smaller language models can learn skill composition from limited examples, substantially improving their ability to combine skills in novel ways through fine-tuning.

LLMs can improve reasoning by teaching weaker models, a process called Learning by Teaching (LbT), as shown in this preliminary study. LbT enhances not just student models, but also the teacher model…

LLMs struggle to reliably generate valid data science code due to a lack of true understanding of numerical constraints in APIs, despite seemingly mastering common patterns through extensive training.

LLMs struggle with in-context exploration, needing substantial prompt engineering or training interventions to effectively explore multi-armed bandit environments.

LLM agents surprisingly exhibit human-like trust behavior, especially GPT-4, paving the way for simulating complex human interactions in various applications.

LLMs struggle with noisy rationales in chain-of-thought prompting. This paper introduces the NoRa dataset, showing that existing methods struggle. A new method, CD-CoT, significantly improves accura…

Language models learn to skip steps in reasoning, improving efficiency and generalization, showcasing emergent human-like cognitive abilities.

New efficient attack reveals GNN model training data properties.

GNNs enhance LLM-based task planning by improving the ability to process task graphs, surpassing existing solutions even without training.

This paper introduces a novel quantitative definition of AI alignment for social decision-making, proposing probably approximately aligned policies and a method to safeguard any autonomous agent’s act…