FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation

Paper: arXiv:2310.03214 · Published 2023-10-05 Authors: Tu Vu, Mohit Iyyer, Xuezhi Wang, Noah Constant, Jerry Wei, Jason Wei, et al. (11 authors)

Original Abstract

> 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.

Key Points

• Motivation: Most LLMs are trained once and never updated, so they cannot adapt to a fast-changing world; the paper systematically studies their factuality on questions testing current world knowledge.
• FreshQA benchmark: A dynamic QA benchmark covering diverse question/answer types, including fast-changing knowledge questions and false-premise questions that need debunking. It is periodically updated.
• Evaluation setup: Closed- and open-source LLMs are benchmarked under a two-mode evaluation procedure measuring both correctness and hallucination, supported by more than 50K human judgments.
• Finding: All models, regardless of size, struggle on fast-changing knowledge and false-premise questions — significant room for improvement remains.
• FreshPrompt: A simple few-shot prompting method that injects relevant, up-to-date evidence retrieved from a search engine into the prompt, substantially boosting FreshQA performance.
• Results: FreshPrompt outperforms competing search-augmented prompting methods (e.g., Self-Ask, Press et al. 2022) and commercial systems such as Perplexity.AI.
• Analysis insights: The number of retrieved evidences and their ordering both strongly influence answer correctness; instructing the model to answer concisely and directly reduces hallucination relative to verbose answers.
• Release: FreshQA is available at github.com/freshllms/freshqa, with commitments to regular updates.

Context in the RAG / Search Landscape

This work sits at the intersection of retrieval-augmented generation (RAG) and large-scale search. The paper's technique follows the pattern of query formalization → system design → construction pipeline → inference pipeline: encoding queries and retrieved documents into the prompt, retrieving up-to-date evidence, and controlling generation style (concise answers) to reduce hallucination.

Broader takeaways for search/recommendation/personalization practitioners:

1. Architecture: Cascaded retrieve + rerank + generate remains mainstream, but agentic paradigms increasingly treat retrieval count and strategy as learnable decisions. 2. Data: High-quality instruction data and behavioral logs matter; synthetic data must guard against knowledge leakage and distribution shift. 3. Evaluation: The gap between offline metrics and online satisfaction is widening; LLM-as-judge needs cross-validation against human evaluation (as done here with 50K+ human judgments). 4. Product: Latency, cost, explainability, and safety are hard constraints for industrial deployment — not just academic benchmarks.

Engineering Checklist

| Check | Question | Recommendation | |-------|----------|----------------| | Data | Does training/index contain PII? Versioning? | Partitioned indexes, anonymization, rollback-capable embedding versions | | Latency | p99 budget? How many retrieval steps? | Cascade + early stopping, hot-query caching, async reranking | | Quality | Do offline gains translate to online CTR/satisfaction? | Interleaving experiments, manual audits, citation verification | | Safety | Does open retrieval introduce poisoning/bias? | Source allowlists, adversarial detection, output filtering | | Cost | Per-query token and GPU cost? | Small-model routing, distillation, hybrid sparse+dense retrieval |

Limitations and Future Directions

Potential limitations include experiment scale constrained by compute budgets, mismatch between benchmark and real user query distributions, English-centric data limiting cross-lingual generalization, and safety risks of agentic systems on the open web. Future directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommender applications.

Related Work

• Enhancing Noise Robustness of Retrieval-Augmented Language Models (arXiv:2405.20978)
• Gorilla: Large Language Model Connected with Massive APIs (arXiv:2305.15334)
• Interleaving Retrieval with Chain-of-Thought Reasoning (arXiv:2212.10509)
• Long-form Factuality in Large Language Models (arXiv:2403.18802)
• ReSearch: Learning to Reason with Search via Reinforcement Learning (arXiv:2503.19470)
• Search-R1: Training LLMs to Reason and Leverage Search Engines with RL (arXiv:2503.09516)

Glossary

| Term | Meaning | |------|---------| | IR | Information Retrieval | | RAG | Retrieval-Augmented Generation | | LTR | Learning to Rank | | nDCG | Normalized Discounted Cumulative Gain, a ranking quality metric | | Agentic Search | Modeling search as sequential decision-making and tool invocation | | Gen-IR | Generative Information Retrieval |

References

• Vu, T., Iyyer, M., Wang, X., Constant, N., Wei, J., Wei, J., et al. *FreshLLMs: Refreshing Large Language Models with Search Engine Augmentation.* arXiv:2310.03214, 2023. https://arxiv.org/abs/2310.03214