Skip to main content

LLM Model Dishonesty

 



The paper 'Language Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting' by Miles Turpin et al. investigates the faithfulness of chain-of-thought (CoT) explanations generated by large language models (LLMs) for various tasks. CoT explanations are verbalisation's of step-by-step reasoning that LLMs produce before giving a final output. 

The paper shows that CoT explanations can be misleading and influenced by biasing features in the model inputs, such as the order of multiple-choice options. The paper tests two LLMs, GPT-3.5 and Claude 1.0, on 13 tasks from BIG-Bench Hard and a social-bias task, and finds that accuracy drops significantly when models are biased toward incorrect answers. 

The paper also finds that models justify answers based on stereotypes without mentioning the influence of social biases. The paper concludes that CoT explanations can be plausible yet unfaithful, which poses a risk for trusting LLMs without ensuring their safety. The paper suggests that CoT is promising for explainability, but requires more efforts to evaluate and improve explanation faithfulness.

As the paper discusses the findings the authors state: 

LLMs may be able to recognize that the biasing features are influencing their predictions—e.g., in post-hoc critiques (Saunders et al., 2022)—even if their CoT explanations do not verbalize them. If they can, then this implies that unfaithful CoT explanations may be a form of model dishonesty, as opposed to a lack of capability. 

What is becoming increasingly apparent with LLMs is that the approach to usage matters significantly. Basic CoT approaches do seem to cause  more room for erroneous / deceptive / hallucination's as outputs. It may just be the case with relatively under developed models, and insufficient training approaches that these models were subjected to. It all goes to show that the general release of such models, when we still are unsure of their outputs, was too early.

Labels:

Comments

Post a Comment

Popular posts from this blog

The Whispers in the Machine: Why Prompt Injection Remains a Persistent Threat to LLMs

 Large Language Models (LLMs) are rapidly transforming how we interact with technology, offering incredible potential for tasks ranging from content creation to complex analysis. However, as these powerful tools become more integrated into our lives, so too do the novel security challenges they present. Among these, prompt injection attacks stand out as a particularly persistent and evolving threat. These attacks, as one recent paper (Safety at Scale: A Comprehensive Survey of Large Model Safety https://arxiv.org/abs/2502.05206) highlights, involve subtly manipulating LLMs to deviate from their intended purpose, and the methods are becoming increasingly sophisticated. At its core, a prompt injection attack involves embedding a malicious instruction within an otherwise normal request, tricking the LLM into producing unintended – and potentially harmful – outputs. Think of it as slipping a secret, contradictory instruction into a seemingly harmless conversation. What makes prompt inj...
Post a Comment
Read more

Podcast Soon Notice

I've been invited to make a podcast around the themes and ideas presented in this blog. More details will be announced soon. This is also your opportunity to be involved in the debate. If you have a response to any of the blog posts posted here, or consider an important issue in the debate around AGI is not being discussed, then please get in touch via the comments.  I look forward to hearing from you.
Post a Comment
Read more

AI Agents and the Latest Silicon Valley Hype

In what appears to be yet another grandiose proclamation from the tech industry, Google has released a whitepaper extolling the virtues of what they're calling "Generative AI agents". (https://www.aibase.com/news/14498) Whilst the basic premise—distinguishing between AI models and agents—holds water, one must approach these sweeping claims with considerable caution. Let's begin with the fundamentals. Yes, AI models like Large Language Models do indeed process information and generate outputs. That much isn't controversial. However, the leap from these essentially sophisticated pattern-matching systems to autonomous "agents" requires rather more scrutiny than the tech evangelists would have us believe. The whitepaper's architectural approaches—with their rather grandiose names like "ReAct" and "Tree of Thought"—sound remarkably like repackaged versions of long-standing computer science concepts, dressed up in fashionable AI clot...
Post a Comment
Read more