[Paper Review] Cerise: Program Verification on a Capability Machine in the Presence of Untrusted Code

Logé, Frédéric, Le Pennec, Erwann+1 more|arXiv (Cornell University)|May 12, 2021

Polynomial and algebraic computation196 citations

TL;DR

This tutorial presents a comprehensive overview of enhancing trustworthiness and efficiency in large language models (LLMs) for database integration, focusing on reducing hallucinations through retrieval-augmented generation, self-reflection, and chain-of-thought reasoning, while improving inference efficiency via database-inspired techniques like continuous batching, KV cache management, and hardware-aware scheduling. The key contribution is a unified framework bridging LLMs and databases to enable scalable, reliable, and high-performance AI-driven data management systems.

ABSTRACT

In the rapidly evolving AI era with large language models (LLMs) at the core, making LLMs more trustworthy and efficient, especially in output generation (inference), has gained significant attention. This is to reduce plausible but faulty LLM outputs (a.k.a hallucinations) and meet the highly increased inference demands. This tutorial explores such efforts and makes them transparent to the database community. Understanding these efforts is essential in harnessing LLMs in database tasks and adapting database techniques to LLMs. Furthermore, we delve into the synergy between LLMs and databases, highlighting new opportunities and challenges in their intersection. This tutorial aims to share with database researchers and practitioners essential concepts and strategies around LLMs, reduce the unfamiliarity of LLMs, and inspire joining in the intersection between LLMs and databases.

Motivation & Objective

• To address the critical challenge of hallucinations in LLMs during database tasks by introducing techniques such as retrieval-augmented generation and chain-of-thought prompting.
• To improve LLM inference efficiency by applying database systems principles like continuous batching, KV cache management, and adaptive scheduling.
• To enable seamless integration of LLMs into database workloads through semantic operators, cost-based scheduling, and mixed relational-LLM query processing.
• To inspire database researchers to contribute to efficient and trustworthy LLM systems by highlighting synergies and open research opportunities at the intersection of databases and LLMs.

Proposed method

• Leverages retrieval-augmented generation (RAG) to ground LLM outputs in external knowledge sources, reducing hallucinations.
• Employs chain-of-thought and multi-path reasoning to improve reasoning accuracy and reduce errors in complex reasoning tasks.
• Applies database-inspired optimization techniques such as continuous batching and prefill-decode disaggregation to enhance LLM inference throughput.
• Uses key-value (KV) cache paging and offloading to manage memory pressure and enable efficient long-context generation.
• Introduces adaptive scheduling and prefix KV sharing to optimize resource utilization across concurrent LLM requests.
• Adapts database cost models and mixed-workload optimization to balance accuracy and efficiency in LLM-database systems.

Experimental results

Research questions

• RQ1 How can hallucinations in LLMs be minimized when generating answers for database queries?
• RQ2 What database system techniques can be adapted to improve the efficiency and scalability of LLM inference?
• RQ3 How can LLMs be integrated into database systems to support hybrid relational-LLM workloads?
• RQ4 What are the key performance bottlenecks in LLM serving, and how can they be mitigated using database principles?
• RQ5 How can cost-based and adaptive query optimization be extended to LLM-based database systems?

Key findings

• Retrieval-augmented generation significantly reduces hallucinations by grounding LLM outputs in external knowledge, improving factual consistency.
• Chain-of-thought and multi-path reasoning techniques improve reasoning accuracy, especially in complex, multi-hop question answering.
• Continuous batching and prefill-decode disaggregation increase throughput and reduce latency in LLM serving systems.
• KV cache offloading and compression techniques enable efficient handling of long-context and high-concurrency LLM workloads.
• Adaptive scheduling with prefix KV sharing improves resource utilization and reduces idle time in LLM inference pipelines.
• Cost-based scheduling models for LLMs can effectively balance accuracy and efficiency, enabling mixed workloads with diverse requirements.