Large language models (LLMs) like GPT-3 are well known for his or her ability to generate coherent and informative natural language texts because of their vast amount of world knowledge. Nevertheless, encoding this data in LLMs is lossy and may result in memory distortion, leading to hallucinations that may be detrimental to mission-critical tasks. Moreover, LLMs cannot encode all crucial information for some applications, making them unsuitable for time-sensitive tasks like news query answering. Although various methods have been proposed to boost LLMs using external knowledge, these typically require fine-tuning LLM parameters, which may be prohibitively expensive. Consequently, there’s a necessity for plug-and-play modules that may be added to a hard and fast LLM to enhance its performance in mission-critical tasks.

The paper proposes a system called LLM-AUGMENTER that addresses the challenges of applying Large Language Models (LLMs) to mission-critical applications. The system is designed to reinforce a black-box LLM with plug-and-play modules to ground its responses in external knowledge stored in task-specific databases. It also includes iterative prompt revision using feedback generated by utility functions to enhance the factuality rating of LLM-generated responses. The system’s effectiveness is validated empirically in task-oriented dialog and open-domain question-answering scenarios, where it significantly reduces hallucinations without sacrificing the fluency and informativeness of reactions. The source code and models of the system are publicly available.

The LLM-Augmenter process involves three primary steps. Firstly, when given a user query, it retrieves evidence from external knowledge sources resembling web searches or task-specific databases. It might also connect the retrieved raw evidence with relevant context and reason on the concatenation to create “evidence chains.” Secondly, the LLM-Augmenter prompts a hard and fast LLM like ChatGPT by utilizing the consolidated evidence to generate a response rooted in evidence. Lastly, LLM-Augmenter checks the generated response and creates a corresponding feedback message. This feedback message modifies and iterates the ChatGPT query until the candidate’s response meets verification requirements.

The work presented on this study shows that the LLM-Augmenter approach can effectively augment black-box LLMs with external knowledge pertinent to their interactions with users. This augmentation greatly reduces the issue of hallucinations without compromising the fluency and informative quality of the responses generated by the LLMs.

LLM-AUGMENTER’s performance was evaluated on information-seeking dialog tasks using each automatic metrics and human evaluations. Commonly used metrics, resembling Knowledge F1 (KF1) and BLEU-4, were used to evaluate the overlap between the model’s output and the ground-truth human response and the overlap with the knowledge that the human used as a reference during dataset collection. Moreover, the researchers included these metrics that best correlate with human judgment on the DSTC9 and DSTC11 customer support tasks. Other metrics, resembling BLEURT, BERTScore, chrF, and BARTScore, were also considered, as they’re among the many best-performing text generation metrics on the dialog.

Take a look at the Paper and Project. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to hitch our 26k+ ML SubReddit, Discord Channel, and Email Newsletter, where we share the newest AI research news, cool AI projects, and more.