Source-linked AI summary
FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation
Tu Vu, Mohit Iyyer, Xuezhi Wang, Noah Constant, Jerry Wei, Jason Wei, Chris Tar, Yun-Hsuan Sung, Denny Zhou, Quoc Le, Thang Luong
TL;DR
LLMs lack reliable access to current-world knowledge, motivating a benchmark that measures factuality on dynamic and false-premise questions. The paper introduces FreshQA and a two-mode evaluation, then proposes FreshPrompt, which injects retrieved search evidence into prompts and substantially improves factuality over baseline and competing approaches.
Problem
LLMs often hallucinate and lack scalable real-time knowledge updates, while the effective use of search outputs for improving factuality remains unclear.
Method
The paper builds the dynamic 600-question FreshQA benchmark, evaluates models with RELAXED and STRICT human judgments, and introduces FreshPrompt for few-shot reasoning over retrieved search evidence.
Results
FreshPrompt substantially improves factuality across FreshQA question types and outperforms competing search-augmented approaches; its best GPT-4 variant reaches 77.6% STRICT and 79.0% RELAXED accuracy.
Takeaways & Limitations
Search-engine evidence, including retrieved context and relevant questions, can substantially improve LLM factuality on changing-world QA without additional training.
Takeaways & Limitations
FreshQA requires regular answer updates, and FreshPrompt’s performance with other search engines, multiple queries, and more complex language remains unclear.
Abstract
from arXiv · showhide
Most large language models (LLMs) are trained once and never updated; thus, they lack the ability to dynamically adapt to our ever-changing world. In this work, we perform a detailed study of the factuality of LLM-generated text in the context of answering questions that test current world knowledge. Specifically, we introduce FreshQA, a novel dynamic QA benchmark encompassing a diverse range of question and answer types, including questions that require fast-changing world knowledge as well as questions with false premises that need to be debunked. We benchmark a diverse array of both closed and open-source LLMs under a two-mode evaluation procedure that allows us to measure both correctness and hallucination. Through human evaluations involving more than 50K judgments, we shed light on limitations of these models and demonstrate significant room for improvement: for instance, all models (regardless of model size) struggle on questions that involve fast-changing knowledge and false premises. Motivated by these results, we present FreshPrompt, a simple few-shot prompting method that substantially boosts the performance of an LLM on FreshQA by incorporating relevant and up-to-date information retrieved from a search engine into the prompt. Our experiments show that FreshPrompt outperforms both competing search engine-augmented prompting methods such as Self-Ask (Press et al., 2022) as well as commercial systems such as Perplexity.AI. Further analysis of FreshPrompt reveals that both the number of retrieved evidences and their order play a key role in influencing the correctness of LLM-generated answers. Additionally, instructing the LLM to generate concise and direct answers helps reduce hallucination compared to encouraging more verbose answers. To facilitate future work, we release FreshQA at github.com/freshllms/freshqa and commit to updating it at regular intervals.
1 INTRODUCTION
The paper introduces FreshQA to evaluate LLM factuality on changing-world questions and develops FreshPrompt, which grounds answers in current search information. It benchmarks models with separate correctness and hallucination-sensitive evaluations.
- FreshQA contains diverse questions spanning never-changing, slow-changing, fast-changing, and false-premise answer types.The benchmark targets both changing knowledge and questions requiring premise rebuttal.
- FreshQA includes 600 natural questions covering diverse topics, difficulty levels, and single-hop or multi-hop reasoning.Some ground-truth answers can change over time, and category assignments may later be reclassified.
- Models were evaluated with more than 50K human judgments using RELAXED correctness and STRICT no-hallucination criteria.RELAXED checks the main answer, while STRICT checks every claim for factuality and currency.
- FreshPrompt retrieves relevant, up-to-date search information and teaches LLMs through few-shot prompting to reason over the evidence.The method is designed to improve factuality without additional training.
- FreshPrompt substantially improves GPT-4 accuracy, with gains of 32.6% under RELAXED and 49.0% under STRICT versus vanilla GPT-4.These results motivate search-engine augmentation for factuality on dynamic questions.
2 FRESHQA
FreshQA is a dynamically maintained benchmark of current-world questions built across answer-change categories and reasoning difficulties. Its construction includes quality control, regular updating, and a two-mode evaluation that separates answer correctness from hallucination.
- FreshQA contains 600 questions covering broad topics, changing answers, and multiple reasoning abilities.Questions are designed to test current knowledge and include both single-hop and multi-hop reasoning.
- Data collection: Annotators created questions in four answer categories at one-hop and multi-hop difficulty levels, including false-premise questions requiring explanations.Contributors included NLP researchers and online freelancers, and questions were intended to resemble natural search queries.
- Quality control: Quality control included manual review, duplicate and invalid-question removal, answer and evidence-URL verification, and supplementary valid answers.The process also excluded questions expected to change more frequently than maintainers could update.
- Maintenance: FreshQA requires regular updates because ground-truth answers may change with developments in the world.Researchers are encouraged to evaluate the latest dataset version near each update’s release date.
- Maintenance: The balanced distribution across question types may change over time when questions are reclassified.This reclassification follows changes in the underlying answers or premises.
- Evaluation: RELAXED evaluation checks only the primary answer, whereas STRICT additionally checks whether every fact is accurate and non-hallucinatory.The difference between the two scores provides a way to measure hallucination.
3 PRE-TRAINED LLMS STRUGGLE ON FRESHQA
FRESHQA exposes substantial factuality weaknesses in pretrained LLMs, especially on current-information and false-premise questions. Accuracy falls further under STRICT evaluation, which penalizes hallucinated or outdated claims.
- Hallucination and refusal: Switching from RELAXED to STRICT markedly reduces CHATGPT and GPT-4 accuracy because they produce outdated answers or refuse current-information questions.GPT-4 refuses fast-changing questions 60% of the time, compared with 16% for CHATGPT.
- Current information: All models struggle on questions requiring fast-changing or recent knowledge because their parametric knowledge is not up to date.GPT-4 generally performs best on these questions but never exceeds 15% across both evaluation modes.
- False premises: False-premise questions challenge every model, while T5 and PALM show flat scaling with 0.0% to 1.6% accuracy as model size increases.GPT-3.5, CHATGPT, and GPT-4 perform substantially better, reaching 25.8% to 42.7% under STRICT and 32.3% to 66.9% under RELAXED.
- Prompting effects: FEW-SHOT and COT prompting improve some valid-premise results, but COT creates more room for hallucination and therefore performs worse than FEW-SHOT under STRICT.For PALM 540B on pre-2022 knowledge, FEW-SHOT and COT improve STRICT accuracy by +36.1% and +26.9%, respectively.
4 PROMPTING SEARCH ENGINE-AUGMENTED LANGUAGE MODELS
This section introduces FreshPrompt, which augments LLM prompts with current search evidence and few-shot demonstrations, then evaluates its accuracy and hallucination behavior on FreshQA. FreshPrompt substantially improves GPT-3.5 and GPT-4 performance, with gains varying across question types and search-augmented baselines.
- Motivation: Search augmentation addresses low accuracy because evaluated models lack real-time information.
- FreshPrompt: Retrieved evidence is formatted with source, date, title, snippet, and highlighted words, while demonstrations teach reasoning over evidence.
- FreshPrompt: FreshPrompt incorporates relevant, up-to-date Google Search information into an LLM prompt to improve FreshQA performance.
- FreshPrompt: FreshPrompt queries Google Search verbatim and gathers answer boxes, organic results, knowledge-graph content, related questions, and crowdsourced answers.
- FreshPrompt: Few-shot demonstrations show example questions, retrieved evidence lists, and reasoning toward the most relevant and up-to-date answer; an explicit premise check is also tested.
- Experiment setup: The evaluation benchmarks GPT-3.5, GPT-4, Google Search, Perplexity.AI, and Self-Ask on the same date, April 26, 2023.
- Experiment setup: Evidence is sequentially incorporated under context limits, with retrieved items sorted by date and only the top n retained.
- Results and discussion: GPT-4 plus FreshPrompt improves absolute accuracy by 47% under STRICT and 31.4% under RELAXED, reducing the STRICT–RELAXED gap from 17.8% to 2.2%.
5 RELATED WORK
FreshPrompt belongs to research on retrieving external knowledge for LLMs and to benchmarks testing temporal reasoning. Its single-call design contrasts with more expensive multi-call retrieval-and-reranking approaches, while Self-Ask emphasizes question decomposition.
- Prior work augments LLM knowledge through fine-tuning, knowledge generation, or search-engine retrieval for real-time factual alignment.
- Lazaridou et al. insert Google Search documents using 50 inference calls and LLM reranking, whereas FreshPrompt uses a single inference call.
- Self-Ask uses few-shot prompting to teach follow-up questioning before answering, with greater emphasis on decomposition.
- FreshQA extends time-sensitive QA work by testing evolving knowledge alongside diverse question types and false-premise handling.
6 LIMITATIONS AND FUTURE WORK
The benchmark and method have practical scope boundaries involving answer freshness, search-engine dependence, query breadth, language coverage, and adaptation strategy.
- FreshQA answers require regular maintainer updates because some answers may become stale between update periods.
- FreshPrompt’s performance with search engines other than Google is unclear when features such as answer boxes are unavailable.
- Using one search query per question leaves room for question decomposition and multiple-query improvements.
- Because FreshQA contains relatively simple English questions, performance on multilingual, cross-lingual, and long-form QA is unclear.
- FreshPrompt relies on in-context learning and may underperform methods that fine-tune the base LLM on new knowledge.
7 CONCLUSION
The paper evaluates LLM adaptation to changing world knowledge, introduces FreshQA and FreshPrompt, and reports improved performance from search-grounded prompting. It also releases the benchmark for continued research.
- The paper evaluates modern LLMs’ adaptation to changing world knowledge with and without search-engine augmentation.
- FreshQA contains 600 questions spanning reasoning abilities including fast-changing knowledge and false-premise identification.
- The two-mode evaluation measures both answer correctness and hallucination.
- FreshPrompt is a few-shot in-context method that incorporates relevant Google Search evidence into an LLM prompt.
- FreshPrompt significantly outperforms competing search-engine-augmented approaches on FreshQA, while evidence count and ordering affect answer correctness.
- The authors release FreshQA and commit to regularly updating its answers to support future research.
A EVALUATION PROTOCOL
Figure 4 presents examples illustrating the evaluation criteria used in the study.
- Figure 4 provides specific examples for each evaluation criterion.
- The examples are intended to clarify how responses are assessed.
- The figure supports interpretation of the study’s evaluation procedure through concrete cases.
B INTER-RATER AGREEMENT AND AUTOMATIC EVALUATION
The study validates its human and automatic evaluation procedures through independent ratings and agreement analysis. FRESHEVAL closely matches human judgments in both evaluation modes.
- Human evaluation: Two authors independently evaluated 100 sampled answers across models in both RELAXED and STRICT modes.The sample included 50 questions with valid premises and 50 with false premises.
- Automatic evaluation: FRESHEVAL conditions an LLM on the question, valid answers, and model response before generating a correctness comment and final judgment.An evaluation-task instruction is also included at the beginning of each prompt.
- Agreement results: 99% RELAXED and 96% STRICT agreement was observed between human raters, while FRESHEVAL averaged 96.5% and 96% agreement with human evaluations.These results support the reproducibility of the protocol and the use of FRESHEVAL in place of human evaluation on FRESHQA.
C ADDITIONAL EXPERIMENT SETUP DETAILS FOR SECTION 3
The experiments use deterministic decoding settings and identify the specific Azure OpenAI models used for the OpenAI evaluations.
- Decoding settings: All models select the most likely token at every decoding timestep using temperature 0.Responses are limited to a maximum of 256 generated tokens.
- Decoding settings: The setup uses a maximum generation length of 256 tokens for all models.
- OpenAI models: OpenAI evaluations use the 2023-03-15-preview API in Azure OpenAI Service.The listed model names are text-davinci-003, code-davinci-002, gpt-3.5-turbo, and gpt-4.
D ADDITIONAL EXPERIMENT RESULTS FOR SECTION 3
Additional results describe the evaluation tables, evidence that some models access post-cutoff information, and figures documenting search results and FreshPrompt prompts.
- Accuracy results: Tables 3 and 4 report LLM accuracy on FRESHQA under STRICT and RELAXED evaluations.
- Knowledge cutoff: Manual evaluation suggests ChatGPT and GPT-4 were exposed to information beyond their September 2021 knowledge cutoff.Figure 8 indicates that ChatGPT is aware of the February 24, 2022 Russian invasion of Ukraine.
- Evaluation design: Separate prompts for RELAXED and STRICT evaluations performed better than one combined prompt, while retrieved evidence did not improve FRESHEVAL–human agreement.
- Search results: Figure 9 displays different types of Google Search results for a query.
- FreshPrompt: Figure 10 displays a realistic prompt used for FreshPrompt.
H ADDITIONAL EXPERIMENT RESULTS FOR SECTION 4
This section presents the evaluation criteria, prompts, examples, and benchmark tables used to assess FRESHQA responses under RELAXED and STRICT settings.
- Evaluation criteria: RELAXED evaluation requires an accurate primary answer but permits ill-formed, hallucinated, or outdated information that does not significantly affect it.Responses must still be confident and definitive, and additional information cannot contradict or reshape the primary answer.
- Evaluation criteria: STRICT evaluation requires every claim to be accurate and up-to-date, rejecting responses containing any hallucination, however minor.False-premise questions require the response to identify the false premise, while outdatedness disclaimers are accepted only when the knowledge evidently has not changed.
- Evaluation examples: The Elon Musk–Activision Blizzard example illustrates that a response must correct a false premise rather than merely qualify its answer by citing a knowledge cutoff.The reference answer states that Microsoft, not Elon Musk, purchased Activision Blizzard for US$68.7 billion in January 2022.
- Benchmark results: Tables report model accuracy across changing-knowledge, false-premise, temporal, and reasoning categories under both STRICT and RELAXED evaluations.The reported benchmark date is April 26, 2023, and models with current-date access are marked with '+'.
- Search augmentation: FRESHPROMPT presents retrieved search evidence in a unified format, while Google Search supplies snippets alongside source, date, title, and highlighted-word metadata.Few-shot demonstrations show questions, retrieved evidence lists, and reasoning toward the most relevant and current answer.