Meta Researchers Propose Lightweight Fine-tuning Method RA-DIT to Enhance Language Model Knowledge Retrieval Capabilities

Translation: Recently, researchers at Meta have proposed a lightweight fine-tuning method called RA-DIT to enhance the knowledge retrieval capabilities of language models. This method involves a two-stage tuning process: the first stage improves the language model's ability to utilize retrieved information, and the second stage optimizes the retriever to provide more relevant content. Experimental results indicate that RA-DIT65B outperforms existing models in knowledge-intensive zero-shot and few-shot tests. It also significantly improves performance on tasks that require high levels of knowledge utilization and contextual understanding. The study demonstrates the effectiveness of RA-DIT's lightweight tuning for retrieval-enhanced language models, particularly in scenarios that require access to large-scale knowledge sources.

After BrowseComp was brushed to 90%, Meituan LongCat launched LoHoSearch: Frontline models collectively dropped back to less than 30%

"Search Agent Evaluation Benchmark BrowseComp was quickly overwhelmed," its performance soared from 30% to 90% and gradually became ineffective. On July 17, Meituan LongCat released a new benchmark LoHoSearch, which generates difficult problems based on a Wikipedia knowledge graph containing 7.62 million entities, aiming to push the evaluation back into a high-difficulty area and reset the standard for search agent capabilities.

Baidu Releases Ernie5.1: Pre-training Cost Drops by 94% and Performance Rises to Top Four in Global Search Rankings

Baidu released its new language model Ernie5.1 on May 11, 2026, based on the pre-trained foundation of Ernie5.0 with 2.4 trillion parameters. Through a 'one-time elastic training framework', it achieves single training optimization for multiple model sizes, with pre-training cost only 6% of similar models. As of May 9, the model ranked fourth globally and first in China on the Arena Search ranking with 1223 points, demonstrating high resource utilization and performance balance.

Customize Your Own Large Model - Is It Necessary to Manually Write Code? This Is the Correct Way to Fine-tune!

Although large models have become widespread, general models often fail to accurately meet specific business needs. To enable the model to deeply understand industry knowledge, fine-tuning is a key step. However, traditional fine-tuning methods still face challenges such as high barriers and high costs.

Stack Overflow launches Stack Internal: Turning Enterprise Q&A into an AI Trustworthy Knowledge Base

Stack Overflow has launched the enterprise product Stack Internal, which provides technical Q&A metadata and reliability scores through the MCP interface, helping AI agents avoid generating incorrect information. The CEO revealed that large customers have already paid to use it, with a business model similar to Reddit's content licensing.

Meta Researchers Propose Lightweight Fine-tuning Method RA-DIT to Enhance Language Model Knowledge Retrieval Capabilities

Related Recommendations

After BrowseComp was brushed to 90%, Meituan LongCat launched LoHoSearch: Frontline models collectively dropped back to less than 30%

Baidu Releases Ernie5.1: Pre-training Cost Drops by 94% and Performance Rises to Top Four in Global Search Rankings

Customize Your Own Large Model - Is It Necessary to Manually Write Code? This Is the Correct Way to Fine-tune!

Stack Overflow launches Stack Internal: Turning Enterprise Q&A into an AI Trustworthy Knowledge Base

AI Model Uses Two Books to Generate Masterpieces in a Famous Style, Sparking New Discussions on Copyright Law