Studies reveal persistent concerns over AI ‘hallucinations’ in code generation

Studies Reveal Persistent Concerns Over AI ‘Hallucinations’ in Code Generation

Researchers Warn of Significant Risks with Large Language Models

Recent research has reignited concerns about the accuracy and reliability of Large Language Models (LLMs) in providing coding assistance, highlighting their tendency to generate incorrect or entirely fabricated information. These so-called “hallucinations” can create significant challenges, particularly when software developers use LLM-generated code containing non-existent software package dependencies.

Notable Incidents and Studies

Two studies, one involving a team from the University of Texas at San Antonio, University of Oklahoma, and Virginia Tech, and the other by researchers at the Valencian Research Institute for Artificial Intelligence in Spain, have delved into these issues.

The first study, detailed in a preprint paper titled “We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs,” reveals the extent to which LLMs can inadvertently generate fictitious software package names. These fabricated dependencies could potentially be exploited by malicious actors to introduce malware by creating packages with the invented names.

The researchers examined 16 popular LLMs, including both commercial and open-source variants, across 576,000 code samples in JavaScript and Python. The results were disconcerting: about 20% of the packages generated were hallucinations, with 205,474 unique non-existent packages recorded.

“Our findings reveal that the average percentage of hallucinated packages is at least 5.2% for commercial models and 21.7% for open-source models,” the study reports. The findings are particularly notable when compared to previous analyses, such as Lasso Security’s evaluations of models like GPT-3.5 and GPT-4, which showed even higher rates of hallucinated packages.

Mitigation Strategies

In an effort to address these hallucinations, the researchers implemented mitigation strategies, including Retrieval Augmented Generation (RAG) and Supervised Fine-Tuning. These approaches aimed to provide a list of valid package names and filter out invented ones. The strategies did reduce hallucination rates, but at a cost: the code quality suffered a notable decrease.

Scaling Up Models: A Double-Edged Sword

In a related study, José Hernández-Orallo and his colleagues examined the reliability of scaled-up LLMs. They found that while larger models, such as OpenAI’s GPT-4, Meta’s LLaMA, and BigScience’s BLOOM, demonstrate higher accuracy, they also show a greater propensity for providing plausible but incorrect answers. Smaller models, in contrast, are more likely to abstain from answering questions they cannot reliably predict, thereby avoiding false information.

The study particularly highlighted how models like GPT-4 are more likely to answer almost any prompt, increasing the likelihood of incorrect responses. Additionally, the researchers found that humans often misjudge these AI-generated answers, with incorrect answers being classified as correct between 10% and 40% of the time.

“Relying on human oversight for these systems is a hazard, especially for areas for which the truth is critical,” the researchers concluded.

Implications

These findings underscore the need for a fundamental shift in the design and deployment of general-purpose AI, especially in high-stakes areas where precision is non-negotiable. As AI technologies continue to evolve and integrate more deeply into various domains, addressing these hallucination issues becomes paramount.

Both studies emphasise that while LLMs hold considerable promise, significant challenges remain in ensuring that these models can be relied upon for tasks where accuracy and reliability are crucial.

Source: Noah Wire Services