Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More
Hallucinations, or factually inaccurate responses, continue to plague large language models (LLMs). Models falter particularly when they are given more complex tasks and when users are looking for specific and highly detailed responses.
It’s a challenge data scientists have struggled to overcome, and now, researchers from Google DeepMind say they have come a step closer to achieving true factuality in foundation models. They have introduced FACTS Grounding, a benchmark that evaluates LLMs’ ability to generate factually accurate responses based on long-form documents. Models are also judged on whether their responses are detailed enough to provide useful, relevant answers to prompts.
Along with the new benchmark, the researchers have released a FACTS leaderboard to the Kaggle data science community.
As of this week, Gemini 2.0 Flash topped the leaderboard, with a factuality score of 83.6%. Others in the top 9 include Google’s Gemini 1.0 Flash and Gemini 1.5 Pro; Anthropic’s Clade 3.5 Sonnet and Claude 3.5 Haiku; and OpenAI’s GPT-4o, 4o-mini, o1-mini and o1-preview. These all ranked above 61.7% in terms of accuracy.
The researchers say the leaderboard will be actively maintained and continually updated to include new models and their different iterations.
“We believe that this benchmark fills a gap in evaluating a wider variety of model behaviors pertaining to factuality, in comparison to benchmarks that focus on narrower use cases…such as summarization alone,” the researchers write in a technical paper published this week.
Weeding out inaccurate responses
Ensuring factual accuracy in LLM responses is difficult because of modeling (architecture, training and inference) and measuring (evaluation methodologies, data and metrics) factors. Typically, researchers point out, pre-training focuses on predicting the next token given previous tokens.
“While this objective may teach models salient world knowledge, it does not directly optimize the model towards the various factuality scenarios, instead encouraging the model to generate generally plausible text,” the researchers write.
To address this, the FACTS dataset incorporates 1,719 examples — 860 public and 859 private — each requiring long-form responses based on context in provided documents. Each example includes:
- A system prompt (system_instruction) with general directives and the order to only answer based on provided context;
- A task (user_request) that includes a specific question to be answered;
- A long document (context_document) with necessary information.
To succeed and be labeled “accurate,” the model must process the long-form document and create a subsequent long-form response that is both comprehensive and fully attributable to the document. Responses are labeled “inaccurate” if the model’s claims are not directly supported by the document and not highly relevant or useful.
For example, a user may ask a model to summarize the main reasons why a company’s revenue decreased in Q3, and provide it with detailed information including a company’s annual financial report discussing quarterly earnings, expenses, planned investments and market analysis.
If a model then, say, returned: “The company faced challenges in Q3 that impacted its revenue,” it would be deemed inaccurate.
“The response avoids specifying any reasons, such as market trends, increased competition or operational setbacks, which would likely be in the document,” the researchers point out. “It doesn’t demonstrate an attempt to engage with or extract relevant details.”
By contrast, if a user prompted, “What are some tips on saving money?” and provided a compilation of categorized money-saving tips for college students, a correct response would be highly detailed: “Utilize free activities on campus, buy items in bulk and cook at home. Also, set spending goals, avoid credit cards and conserve resources.”
DeepMind uses LLMs to judge LLMs
To allow for diverse inputs, researchers included documents of varying lengths, up to 32,000 tokens (or the equivalent of 20,000 words). These cover areas including finance, technology, retail, medicine and law. User requests are also broad, including Q&A generation, requests for summarization and rewriting.
Each example is judged in two phases. First, responses are evaluated for eligibility: If they don’t satisfy user requests, they are disqualified. Second, responses must be hallucination-free and fully grounded in the documents provided.
These factuality scores are calculated by three different LLM judges — specifically Gemini 1.5 Pro, GPT-4o and Claude 3.5 Sonnet — that determine individual scores based on the percentage of accurate model outputs. Subsequently, the final factuality determination is based on an average of the three judges’ scores.
Researchers point out that models are often biased towards other members of their model family — at a mean increase of around 3.23% — so the combination of different judges was critical to help ensure responses were indeed factual.
Ultimately, the researchers emphasize that factuality and grounding are key factors to the future success and usefulness of LLMs. “We believe that comprehensive benchmarking methods, coupled with continuous research and development, will continue to improve AI systems,” they write.
However, they also concede: “We are mindful that benchmarks can be quickly overtaken by progress, so this launch of our FACTS Grounding benchmark and leaderboard is just the beginning.”
