Language & LLMs
Large language models show some noise robustness but still struggle with negative rejection, information integration, and dealing with false information, a study finds
In a 2024 paper in the Proceedings of the AAAI Conference on Artificial Intelligence, researchers tested six large language models on a new benchmark called RGB. The benchmark, with English and Chinese tasks, measures four abilities needed for retrieval-augmented generation, or RAG. RAG is a technique that gives a model outside text to ground its answers. The models showed some noise robustness, but they struggled significantly with negative rejection, information integration, and dealing with false information.
What did the benchmark test?
The RGB benchmark splits its cases into four testbeds: noise robustness, negative rejection, information integration, and counterfactual robustness. The abstract lists negative rejection as one of the four abilities but does not define it. Counterfactual robustness is the ability to handle false information. The benchmark includes cases in both English and Chinese.
What did the study find?
The six large language models showed a certain degree of noise robustness. They still struggled significantly with negative rejection, information integration, and dealing with false information. The authors say the results point to much more work before RAG can be applied to large language models effectively.
FAQ
What is negative rejection?
The abstract lists negative rejection as one of the four abilities but does not define it.
Did the models handle noisy text?
Yes, at least partly. The study found that the models showed a certain degree of noise robustness.
Is the benchmark only in English?
No, the benchmark includes cases in both English and Chinese.
Limits and what's next
The findings come from six representative large language models evaluated on one new benchmark, RGB, with English and Chinese tasks. The benchmark is a lab evaluation. The abstract does not report an independent replication of these results. The authors conclude that applying RAG to large language models effectively still requires considerable work.
Source: Jiawei Chen et al., “Benchmarking Large Language Models in Retrieval-Augmented Generation,” Proceedings of the AAAI Conference on Artificial Intelligence, 2024.