New NGM module enhances LLMs with training-free memory

Researchers have developed N-gram Memory (NGM), a novel module designed to enhance Large Language Models (LLMs) without requiring additional training. NGM leverages a model's existing pretrained token embeddings as a memory store, utilizing a Causal N-Gram Encoder and a Cosine-Gated Memory Injector. This plug-and-play approach avoids the need for separate memory banks or complex retrieval systems. NGM has demonstrated consistent performance improvements across various benchmarks and model backbones, including Qwen3 and Llama-3.2-Instruct, and has also shown efficacy with vision-language models like Qwen3-VL-2B. AI

IMPACT This training-free memory module could streamline LLM development and improve performance on complex tasks.

RANK_REASON The cluster contains an academic paper detailing a new method for enhancing LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• GPQA Diamond
• HumanEval
• Llama-3.2-Instruct
• LLMs
• N-gram Memory
• Qwen3
• Qwen3-VL-2B
• Yuwen Quan

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →