Large Language Models

Train high-performing Process Reward Models (PRMs) cheaply using only outcome-level labels, eliminating the need for costly step-by-step annotations!

O1-CODER replicates OpenAI’s o1 model for coding, integrating reinforcement learning and Monte Carlo Tree Search to enhance System-2 thinking and generate high-quality code with reasoning steps.

KV Shifting Attention: A novel attention mechanism significantly enhances language modeling by simplifying induction heads, leading to improved performance and faster convergence, even in large-scale …

New multilingual LLM benchmark, INCLUDE, tackles regional knowledge gaps by using 197K QA pairs from 44 languages, improving cross-lingual evaluation.

Boosting LLMs’ reasoning: A novel token-level contrastive estimation method automatically identifies and penalizes critical tokens leading to errors, significantly enhancing reasoning accuracy.

Boost LLM accuracy exponentially by using a two-stage algorithm with provable scaling laws: generate multiple candidate solutions then compare them in a knockout tournament!

Dynamic parallel processing boosts LLM inference speed on hybrid CPUs by over 90% memory bandwidth, resolving performance bottlenecks caused by imbalanced hardware capabilities.

Puzzle: a novel framework accelerates large language model inference by using neural architecture search and knowledge distillation, achieving a 2.17x speedup on a single GPU while preserving 98.4% ac…

Researchers created TemplateGSM, a massive dataset of 7M+ grade-school math problems and solutions, using GPT-4 to generate templates, significantly advancing LLM training for mathematical reasoning.

Self-VerIfication length Policy (SVIP) dynamically adjusts speculative decoding draft lengths based on token difficulty, achieving up to 20% faster large language model inference.

HiAR-ICL, a novel automated reasoning paradigm using Monte Carlo Tree Search, surpasses state-of-the-art accuracy in complex mathematical reasoning by shifting focus from specific examples to abstract…

Star Attention: 11x faster LLM inference on long sequences with 95-100% accuracy!

Low-bit quantization excels for undertrained LLMs but struggles with fully-trained ones; new scaling laws reveal this, directing future research.

Predicting emergent LLM capabilities is now possible by finetuning smaller models; this approach shifts the emergence point, enabling accurate predictions of future model performance, even with up to …

Simple distillation from OpenAI’s API, combined with fine-tuning, surprisingly surpasses OpenAI’s O1-preview on complex mathematical reasoning, urging transparency in AI research.

MH-MoE: A novel implementation of Multi-Head Mixture-of-Experts achieves superior performance in large language models by enhancing efficiency without sacrificing model size or computational cost.

LLMs are revolutionizing AI evaluation by offering nuanced judgments surpassing traditional methods. This paper provides a taxonomy, benchmark, and future directions for LLM-as-a-judge.

A novel LLM-based transpiler, CRT, efficiently converts x86 assembly to ARM and RISC-V assembly, achieving high accuracy and significant performance improvements over existing virtualization methods.

AI-powered robots now understand and execute complex natural language commands, adapting seamlessly to dynamic environments thanks to a new architecture integrating LLMs, perception, and planning.

MolReFlect achieves state-of-the-art molecule-text alignment by using a teacher-student LLM framework that generates fine-grained alignments, improving accuracy and explainability.