AI Research Paper Feed

TLDR: The paper introduces a bias-reduced evaluation protocol for LLMs in multiple-choice questions that improves robustness to answer permutations with minimal performance loss.

TLDR: The 'Mine and Refine' framework enhances e-commerce search retrieval by optimizing semantic text embeddings through a two-stage process involving contrastive training and policy-aligned fine-tuning.

TLDR: InstructKG is a framework that automatically constructs knowledge graphs from course materials to capture learning dependencies and aid personalized learning.

TLDR: The paper presents a unified framework for addressing locality in scalable multi-agent reinforcement learning (MARL) by introducing a policy-dependent approach to the exponential decay property (EDP) of value functions.

TLDR: The paper introduces a novel grey-box method that integrates incomplete physics models into generative models for learning dynamics from observational data without relying on ground-truth physics parameters.

TLDR: The paper introduces Phase-Aware Mixture of Experts (PA-MoE) to enhance reinforcement learning by allowing expert specialization for complex tasks without being dominated by simpler tasks.

TLDR: This paper proposes a Privacy-by-Design framework for developing AI applications for children that integrates privacy regulations to ensure data protection and legal compliance.

TLDR: KLong is a new LLM agent designed to tackle long-horizon tasks using a novel training method combining trajectory-splitting SFT and progressive RL, outperforming existing models on various benchmarks.

TLDR: ASTERIS, a self-supervised denoising algorithm, enhances astronomical imaging detection limits by leveraging spatiotemporal data, improving detection by 1 magnitude and identifying previously undetectable features in deep space images.

TLDR: This paper presents a novel approach to crystal structure prediction using large language models and constrained optimization to improve symmetry inference and enforce physical validity, achieving state-of-the-art results without relying on existing databases.

TLDR: This study compares Deep Reinforcement Learning (DRL) and Mean-Variance Optimization (MVO) for portfolio allocation, showing DRL's strong performance across various financial metrics.

TLDR: This paper presents an AI-enhanced pipeline using In-Context Learning and tensor methods to improve data analysis in behavioral neuroscience, particularly for studying fear generalization in mice, which can help understand PTSD.

TLDR: The paper identifies that weight sign persistence is a bottleneck in sub-bit model compression and proposes methods to reduce sign flips while maintaining performance.

TLDR: The paper introduces Successive Sub-value Q-learning (S2Q), a method that improves adaptability in multi-agent reinforcement learning by retaining multiple high-value actions, outperforming existing algorithms.

TLDR: This paper introduces reusability and verifiability as new metrics to evaluate the quality of Chain-of-Thought reasoning in multi-agent IR pipelines, revealing that these metrics are not correlated with traditional accuracy measures.

TLDR: Certain transformer attention heads in language models act as membership testers, identifying repeated tokens with high precision, similar to Bloom filters.

TLDR: CRAFT is a parameter-efficient fine-tuning method using Tucker decomposition on pre-trained attention weights, achieving competitive performance with minimal adaptation parameters.

TLDR: This study introduces a taxonomy for fine-grained uncertainty quantification in long-form language model outputs, revealing that claim-level scoring and uncertainty-aware decoding improve factuality in generated content.

TLDR: This study proposes a collaborative AI workflow for humanities and social sciences research, using Taiwan's Claude.ai data to validate its feasibility and effectiveness.

TLDR: This paper introduces One-Shot Incremental Federated Learning (OSI-FL), a framework that addresses communication overhead and catastrophic forgetting in federated learning by using category-specific embeddings and selective sample retention.