Natural Language Processing

Artemis: A new MLLM excels at video-based referential understanding, accurately describing targets within complex videos using natural language questions and bounding boxes, surpassing existing models…

ARKVALE boosts LLM inference efficiency by intelligently evicting and recalling key-value pairs from cache, improving latency and throughput without significant accuracy loss.

More LM calls don’t always mean better results for compound AI; this study reveals performance can initially increase then decrease, highlighting the importance of optimal call number prediction.

Language models (LMs) can now forecast future events as accurately as expert human forecasters! This groundbreaking research unveils a retrieval-augmented LM system surpassing human forecasters in spe…

LLMs’ emotional alignment with humans is assessed using emotion appraisal theory, revealing that while LLMs respond appropriately in some cases, they lack alignment with human emotional behaviors and …

AP-Adapter boosts text-to-image diffusion model generalization by using a two-stage prompt optimization method that leverages large language models and inter-model differences.

This study reveals that modifying optimizers to normalize updates based on angular changes and gradient signal-to-noise ratio significantly reduces the need for learning rate warmup in GPT training.

Steering vectors, while promising for controlling LLMs, show unreliable in- and out-of-distribution performance, highlighting crucial limitations for real-world applications.

ANAH-v2 tackles LLM hallucination by introducing a self-training framework that iteratively scales annotation datasets and improves annotator accuracy, achieving state-of-the-art results.

Extend LLMs context via a simple, training-efficient positional encoding method, CREAM, outperforming existing methods by focusing on crucial mid-context information.

AMOR: Adaptable Modular knowledge agent using LLMs, excels with FSM-based reasoning and process feedback, enabling human supervision and domain adaptation.

AmoebaLLM: Instantly create optimally-sized LLMs for any platform!

ALPS: An optimization-based framework achieves state-of-the-art one-shot LLM pruning, significantly reducing test perplexity and improving zero-shot performance.

ALPINE reveals how Transformer-based LLMs learn planning by embedding graph information into their weights, but also highlights their inability to handle transitive relationships.

AlphaPruning leverages Heavy-Tailed Self-Regularization theory to allocate optimal layer-wise sparsity ratios in LLMs, achieving 80% sparsity in LLaMA-7B with reasonable perplexity.

AlphaMath: LLMs excel at math reasoning without human-annotated process supervision, using Monte Carlo Tree Search.

This paper introduces a novel framework for aligning LLMs with honesty, proposing new metrics and training techniques to make LLMs more truthful and less prone to confidently incorrect responses.

This study introduces ’native alignment’ for Arabic LLMs, achieving state-of-the-art results by aligning models during pre-training, rather than post-training.

JANUS, a 7B LLM, achieves high alignment to thousands of user preferences by generalizing from diverse system messages, outperforming existing LLMs on various benchmarks.

PRELUDE, a novel framework, leverages user edits of LLM outputs to learn latent preferences, improving agent alignment and minimizing edit costs. CIPHER, its efficient algorithm, infers preferences f…