Researchers from Meta AI Introduce a New AI Model to Critique Large Language Model Generations

The ability of large language models (LLMs) to generate coherent, contextually relevant, and semantically meaningful text has become increasingly complex. Despite these advancements, LLMs frequently provide inaccurate, doubtful, and nonsensical results. Thus, techniques that continually assess and improve generations would be helpful toward more trustworthy language models. Language model outputs have been enhanced with the help of LLMs. Among the current work, some train utility functions to give natural language feedback on information-seeking dialogue tasks. In contrast, others employ instruction prompting to create a multi-aspect evaluation score of model-generated output text from various domains. 

Though the original research failed to offer feedback on model output production on complicated tasks like math and reasoning, only providing general feedback on the output response, a more recent work by researchers instruction-tunes an LLM to create self-feedback on its replies. In this study, researchers from Meta AI Research introduce Shepherd, a language model specifically optimized to evaluate outputs produced by models. They aim to develop a strong criticism model that can offer comments across many fields, yet they share a similar aim with previous work. Their approach can identify particular problems, including factuality, logical flaws, coherence, and alignment, while also suggesting modifications when requested to enhance the result. 

More precisely, Shepherd can produce natural language feedback that includes deep topic knowledge, concrete suggestions for improvement, and broad judgments and recommendations. They developed a high-quality feedback dataset of two unique sets to improve Shepherd and assess it: (1) community feedback, curated from online forums to capture more varied interactions, and (2) human-annotated input, gathered on generations across many tasks. See illustrations in Table 1. Shepherd has outstanding performance after being trained on a mix of these datasets, surpassing ChatGPT models on several downstream tasks. The community data is more useful and diverse than the human-annotated data. However, according to a close examination of the effects of community feedback and human-annotated feedback data, it tends to be more informal. 

Shepherd can provide feedback on various tasks thanks to these subtle differences, and they discover that using high-quality human-annotated data to fine-tune models enhances model performance. They compare the feedback produced by Shepherd to cutting-edge baselines like Alpaca, SelFee, and ChatGPT and do a model-based and human evaluation. They discover Shepherd’s criticisms are often favored above those of other models. For instance, Alpaca tends to complement every model answer, which produces a lot of inaccurate feedback. SelFee frequently ignores model answers or immediately answers the query instead of providing feedback that might identify mistakes. 

They discovered that ChatGPT is more consistent across various assessment circumstances and performs better at providing comments with accurate judgment. In conclusion, they created Shepherd, a novel model that can offer thorough criticisms of any LLM-generated content, effectively raising its quality. They show the effectiveness of Shepherd across a range of generating tasks by carefully analyzing the generated complaints. Creating a top-notch feedback dataset, which might aid future study in this field, is another important addition to their work.

Check out the Paper and Github. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 28k+ ML SubReddit, 40k+ Facebook Community, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.