PaperlayerView exact revision on arXiv

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

arXiv:2310.03214v2cs.CL

TL;DR

LLMs can give factually incorrect or outdated answers when current information matters. FreshLLMs introduces FreshQA and FreshPrompt, which grounds answers in search-engine evidence; GPT-4 with FreshPrompt improves absolute accuracy by 47% under STRICT and 31.4% under RELAXED.

  • Problem

    LLMs can hallucinate and rely on outdated parametric knowledge, limiting trustworthiness when accurate, up-to-date information is critical.

    Source: §1 · p. 1

  • Method

    The paper introduces FreshQA, a dynamic QA benchmark, and FreshPrompt, which uses few-shot prompting to teach an LLM to reason over relevant, up-to-date search results.

    Source: §1 · p. 2, §2.1 · p. 3, §4.1 · p. 6

  • Results

    Models struggle with fast-changing and recent knowledge; GPT-4 with FreshPrompt improves absolute accuracy by 47% under STRICT and 31.4% under RELAXED.

    Source: §3.1 · p. 5, §4.3 · p. 7

  • Takeaways & Limitations

    FreshQA provides a fine-grained evaluation of LLM adaptation to changing world knowledge with and without search-engine augmentation.

    Source: §7 · p. 9

    FreshPrompt interfaces with Google Search, and its performance with other search engines is unclear.

    Source: §6 · p. 9

Abstract

from arXiv · show

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, a dynamic benchmark for LLM factuality, and FRESHPROMPT, a search-grounded in-context method designed to improve it. Its evaluation identifies persistent weaknesses in fast-changing, false-premise, and multi-hop questions.

1 INTRODUCTION

  • FRESHQA contains 600 dynamic questions across never-changing, slow-changing, fast-changing, and false-premise categories, spanning diverse topics and single- and multi-hop difficulty.Ground-truth answers and category assignments may change over time.Source: §1 · p. 2, Figure 1 · p. 2
  • Across more than 50K human judgments, RELAXED evaluation checks the main answer while STRICT checks every claim for factuality and currency; models struggle with fast-changing, false-premise, and multi-hop questions.Larger models do not reliably improve on fast-changing questions, while explicit premise checking can help debunk some false-premise questions; chain-of-thought prompting can increase hallucinations.Source: §1 · pp. 2–3
  • 32.6% and 49.0% accuracy improvements over vanilla GPT-4 are reported for the best FRESHPROMPT variant on FRESHQA under RELAXED and STRICT evaluation, respectively.The method requires no additional training.Source: §1 · p. 2
  • FRESHPROMPT extracts relevant, up-to-date search information, including related user questions, and uses few-shot prompting to teach models to reason over retrieved evidence.The authors report that it boosts factuality compared with competing search-augmented approaches.Source: §1 · pp. 2–3
  • The authors release FRESHQA freely and commit to regularly updating its ground-truth answers to support work on improving LLM factuality.Source: §1 · p. 3

2 FRESHQA

FRESHQA is a 600-question benchmark designed to assess LLM factuality across diverse question and answer types. It includes questions about changing world knowledge and false-premise questions.

2.1 DATA COLLECTION

  • Researchers and freelancers wrote natural, varying-difficulty questions across four broad types, distinguishing one-hop from multi-hop questions and requesting explanations for false-premise questions.Annotators were encouraged to use questions that reflect fresh knowledge and seem plausible to type into a search engine.Source: §2.1 · p. 3
  • 600 questions comprise a 500-question test set and a 100-question development set, after manual cleaning, duplicate removal, answer and evidence verification, and exclusion of answers changing more than weekly.The dataset is intended for regular updates, and the test set’s balance across question types may change as questions are reclassified.Source: §2.1 · p. 3

2.2 EVALUATION

  • RELAXED scores primary-answer correctness, while STRICT also requires every fact to be accurate; false-premise answers must identify the flaw.Two authors agreed on 99% of sampled RELAXED judgments and 96% of STRICT judgments; FRESHEVAL achieved average agreement of 96.5% for RELAXED and 96% for STRICT.Source: §2.2 · p. 4

3 PRE-TRAINED LLMS STRUGGLE ON FRESHQA

Across models, FreshQA exposes difficulty with both fast-changing knowledge and false-premise questions. Even the highest overall accuracies are limited, and STRICT evaluation reveals additional accuracy losses for some models.

3 PRE-TRAINED LLMS STRUGGLE ON FRESHQA

  • The benchmark evaluates models from 770M to 540B parameters using individual questions, temperature 0, and no fine-tuning; CHATGPT and GPT-4 are exceptions to the no-current-access setup.For few-shot prompting, the study uses five TRIVIAQA demonstrations from Sun et al. (2023).Source: §3 · p. 4, §3.1 · p. 4

3.1 RESULTS AND DISCUSSION

  • Accuracy spans 0.8%–32.0% under STRICT and 0.8%–46.4% under RELAXED; STRICT especially lowers CHATGPT, GPT-4, and PaLM accuracy.CHATGPT and GPT-4 often give outdated answers or refuse, while PaLM’s drop also reflects response artifacts and hallucination; FLAN-PALM and CODEX exhibit minimal hallucination.Source: Figure 2 · p. 5, §3.1 · p. 5
  • On fast-changing or recent knowledge, GPT-4 generally leads but never exceeds 15% accuracy; under STRICT on post-2022 knowledge, it trails FLAN-PALM and CODEX.GPT-4 refuses 60% of fast-changing questions, compared with CHATGPT’s 16%; both have been exposed to information beyond their knowledge cutoff, and the benchmark notes they had current-date access.Source: §3.1 · pp. 4–5
  • False-premise questions challenge all models: T5 and PaLM remain at 0.0%–1.6% accuracy across sizes, while OpenAI models reach 25.8%–42.7% STRICT and 32.3%–66.9% RELAXED.CHATGPT leads STRICT at 42.7%, GPT-4 leads RELAXED at 66.9%, and GPT-4 reaches 83.9% on pre-2022 knowledge questions.Source: §3.1 · p. 5
  • On valid-premise questions, few-shot and chain-of-thought prompting improve PALM 540B’s pre-2022 accuracy, but chain-of-thought creates more room for hallucination.Under STRICT, few-shot and chain-of-thought improve accuracy over zero-shot by 36.1% and 26.9%, respectively; under RELAXED, the gains are 21.9% and 29.7%.Source: §3.1 · p. 5
  • Multi-hop questions challenge several models, whereas GPT-4’s accuracy differs by less than 2% between one-hop and multi-hop settings.T5 LARGE and XL cannot handle multi-hop questions; FLAN-PALM 540B, CODEX, and GPT-3.5 suffer the most when moving from one-hop to multi-hop questions.Source: §3.1 · p. 5

4 PROMPTING SEARCH ENGINE-AUGMENTED LANGUAGE MODELS

FreshPrompt augments an LLM’s prompt with current search results and few-shot examples to teach evidence-based answering. The section reports that this approach substantially improves FreshQA accuracy and examines which prompting choices affect performance.

4.1 FRESHPROMPT

  • FRESHPROMPT queries Google with each question, formats snippets with source, date, title, and highlighted words, orders evidence oldest-to-newest, and uses few-shot demonstrations to teach evidence-based answering.Demonstrations show a question, retrieved evidence, and reasoning toward the most relevant, up-to-date answer; the prompt can also explicitly check for a valid premise.Source: §4.1 · p. 6, Figure 3 · p. 6

4.3 RESULTS AND DISCUSSION

  • GPT-4 + FRESHPROMPT improved vanilla GPT-4 by 47% STRICT and 31.4% RELAXED absolute accuracy; its 15-evidence variant achieved overall accuracy of 77.6% and 79.0%, respectively.The STRICT–RELAXED accuracy gap fell from 17.8% to 2.2%.Source: Table 1 · p. 7, §4.3 · p. 7
  • GPT-3.5 + FRESHPROMPT beat GPT-3.5-based PPLX.AI and SELF-ASK by 3.8% and 14.4% under STRICT, while PPLX.AI led by 4.2% under RELAXED.Evaluation was conducted on April 26, 2023; Google Search generally beat vanilla GPT-3.5 and GPT-4 except on false-premise questions, but lagged PPLX.AI and FRESHPROMPT across the board.Source: §4.2 · p. 6, §4.3 · pp. 7–8
  • Adding an explicit premise check improved false-premise accuracy for GPT-3.5 and GPT-4, but reduced GPT-3.5’s overall accuracy by 20.8% STRICT and 21% RELAXED.For GPT-4, overall accuracy decreased 0.6% under STRICT and increased 1.2% under RELAXED.Source: §4.3 · p. 8
  • For GPT-4, search-result order or date order generally outperformed random order by up to 2.2% overall accuracy; omitting evidence metadata reduced accuracy by less than 1%.Search order puts top search results at the end of the prompt; date order puts newer evidence at the end.Source: §4.3 · p. 8
  • Removing the answer box reduced GPT-4 + FRESHPROMPT overall accuracy by 1.4% STRICT and 1.6% RELAXED; removing it and other relevant information reduced accuracy by 3.2% and 3.0%.Source: §4.3 · p. 8
  • Increasing GPT-4’s retrieved evidence from 1 to 5, 10, and 15 improved STRICT overall accuracy by 9.2%, 14.2%, and 16.2%; increasing demonstrations from 5 to 15 reduced it by 1%.The default GPT-4 setup used 10 retrieved evidences and 5 demonstrations.Source: §4.2 · p. 7, §4.3 · p. 8
  • More verbose demonstration answers could help with complex questions but reduced STRICT overall accuracy by 2.6%, with greater room for hallucination.Source: §4.3 · p. 8

5 RELATED WORK

The section positions FreshPrompt among retrieval-augmented prompting methods and FreshQA alongside benchmarks for temporal reasoning and questionable premises. It distinguishes FreshQA as a fixed, human-written, open-ended benchmark for changing facts and false-premise questions.

Knowledge augmented LLMs

  • FreshPrompt follows search-engine prompt augmentation for current factual QA, distinct from fine-tuning and knowledge-generation approaches.Lazaridou et al. (2022) use k = 50 inference calls and reranking, whereas FreshPrompt uses one; SELF-ASK (Press et al., 2022) focuses more on question decomposition.Source: §5 · pp. 8–9

Time-sensitive QA

  • FreshQA complements temporal QA benchmarks with fixed, human-written, open-ended questions whose answers change with world developments (Chen et al., 2021b; Zhang & Choi, 2021; Liska et al., 2022; Kasai et al., 2022).TIMEQA uses timestamped question-answer pairs, SITUATEDQA adds temporal context to existing questions, STREAMINGQA draws on timestamped news, and REALTIMEQA evaluates weekly on new events.Source: §5 · p. 9

QA over questionable or counterfactual premises

  • FreshQA complements benchmarks of questionable and counterfactual premises, and the authors report that current LLMs struggle with false-premise questions (Yu et al., 2023c; Kim et al., 2023; Yu et al., 2023b).Several LLMs can debunk a false-premise question when explicitly asked to check premise validity; combining these benchmarks is proposed as future work.Source: §5 · p. 9

6 LIMITATIONS AND FUTURE WORK

The authors identify maintenance and scope limitations for FreshQA and FreshPrompt, and outline community updates and richer retrieval as possible improvements.

6 LIMITATIONS AND FUTURE WORK

  • FreshQA answers may become stale between maintainer updates, though open-source community updates such as GitHub pull requests could help.The paper notes that regular answer updating is required.Source: §6 · p. 9
  • FreshPrompt uses Google Search, so its performance with other search engines lacking some context types, such as answer boxes, is unclear.Source: §6 · p. 9
  • Using one search query per question may limit FreshPrompt; question decomposition and multiple queries are proposed as potential improvements (Khattab et al., 2022).Source: §6 · p. 9
  • FreshPrompt's performance on multilingual or cross-lingual QA and long-form QA is unclear because FreshQA contains relatively simple English questions (Fan et al., 2019).Source: §6 · p. 9
  • Because FreshPrompt relies on in-context learning, it may underperform approaches that fine-tune the base LLM on new knowledge.Source: §6 · p. 9

7 CONCLUSION · APPENDIX

The paper introduces FreshQA and a two-mode evaluation of LLMs’ current-world knowledge, then proposes FreshPrompt, which improves performance over competing search-augmented approaches. It also reports that evidence count and order affect answer correctness and commits to regularly updating FreshQA’s answers.

7 CONCLUSION

  • FreshPrompt significantly improves FreshQA performance over competing search engine-augmented approaches.It incorporates relevant evidence retrieved from Google Search into an LLM prompt using few-shot in-context learning.Source: §7 · p. 9
  • The study evaluates modern LLMs’ ability to adapt to changing world knowledge both with and without search engine augmentation.Source: §7 · p. 9
  • FreshQA contains 600 questions spanning reasoning about fast-changing knowledge and identifying questions with false premises.Source: §7 · p. 9
  • The two-mode evaluation measures both answer correctness and hallucination.Source: §7 · p. 9
  • An ablation finds that the number and order of incorporated evidence affect the correctness of LLM-generated answers.Source: §7 · p. 9
  • The authors release FreshQA and commit to regularly updating its answers to facilitate future research.Source: §7 · p. 9

A EVALUATION PROTOCOL

Figure 4 provides specific examples of each evaluation criterion.

B INTER-RATER AGREEMENT AND AUTOMATIC EVALUATION

The section reports high agreement between human raters and FRESHEVAL, an automatic evaluator designed to support future evaluations.

B INTER-RATER AGREEMENT AND AUTOMATIC EVALUATION

  • Two authors independently evaluated a random 100-answer subset across models, covering 50 valid-premise and 50 false-premise questions in both modes.Source: B INTER-RATER AGREEMENT AND AUTOMATIC… · p. 14
  • FRESHEVAL uses few-shot in-context learning to judge responses conditioned on a question, valid answers, and a model response, with sample comments and evaluations as demonstrations.It generates a correctness comment followed by a final judgment.Source: B INTER-RATER AGREEMENT AND AUTOMATIC… · p. 14
  • Human raters agreed at 99% for RELAXED and 96% for STRICT; FRESHEVAL’s average agreement with human evaluations was 96.5% for RELAXED and 96% for STRICT.The authors conclude that the protocol is reproducible and reliable, and that FRESHEVAL can be used in place of human evaluation on FRESHQA.Source: B INTER-RATER AGREEMENT AND AUTOMATIC… · p. 14

C ADDITIONAL EXPERIMENT SETUP DETAILS FOR SECTION 3

The setup standardizes decoding and prompt formatting for reproducibility and specifies the Azure OpenAI API version and model names used.

C ADDITIONAL EXPERIMENT SETUP DETAILS FOR SECTION 3

  • All models use temperature-0 decoding with a 256-token maximum; non-chat models without QA pretraining receive a “Q: <question>\nA: ” prompt.Some APIs may remain non-deterministic even at temperature 0.Source: C ADDITIONAL EXPERIMENT SETUP DETAILS… · p. 14
  • OpenAI models use the 2023-03-15-preview Azure OpenAI Service API, with text-davinci-003, code-davinci-002, gpt-3.5-turbo, and gpt-4 for GPT-3.5, Codex, ChatGPT, and GPT-4, respectively.Source: C ADDITIONAL EXPERIMENT SETUP DETAILS… · p. 14

D ADDITIONAL EXPERIMENT RESULTS FOR SECTION 3 … G A REALISTIC PROMPT FOR FRESHPROMPT

The additional material reports FreshQA accuracy under strict and relaxed evaluation, evidence of ChatGPT/GPT-4 awareness of post-cutoff knowledge, and examples of Google results and a realistic FreshPrompt. It also notes evaluation-prompt and inter-rater-agreement findings.

D ADDITIONAL EXPERIMENT RESULTS FOR SECTION 3

E CHATGPT/GPT-4’S AWARENESS OF RECENT KNOWLEDGE

F GOOGLE SEARCH RESULTS

G A REALISTIC PROMPT FOR FRESHPROMPT

H ADDITIONAL EXPERIMENT RESULTS FOR SECTION 4

The supplied additional-results material identifies a RELAXED, category-wise accuracy comparison of search-augmented LLMs and documents evaluation and prompting conventions.

H ADDITIONAL EXPERIMENT RESULTS FOR SECTION 4

  • Figure 4’s evaluation criteria require accurate, up-to-date answers and flag false premises; STRICT rejects any hallucination, while RELAXED tolerates some immaterial outdated or hallucinated information.Both modes also require a confident, definitive answer, an accurate standalone primary answer, and no contradictory or misleading additional information.Source: Figure 4 · p. 16
  • The realistic FreshPrompt example formats retrieved evidence with source, date, title, snippet, and highlighted words, preceded by few-shot demonstrations with evidence-based reasoning.Source: Figure 10 · p. 25
  • Table 5 reports accuracy for search engine-augmented LLMs on FRESHQA under RELAXED evaluations, across question types, temporal categories (before 2022 and since 2022), and one-hop versus multi-hop questions.Source: Table 5 · p. 26

Related papers

Loading 2310.03214v2…