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.

Comments

Popular posts from this blog

OpenAI's NSA Appointment Raises Alarming Surveillance Concerns

  The recent appointment of General Paul Nakasone, former head of the National Security Agency (NSA), to OpenAI's board of directors has sparked widespread outrage and concern among privacy advocates and tech enthusiasts alike. Nakasone, who led the NSA from 2018 to 2023, will join OpenAI's Safety and Security Committee, tasked with enhancing AI's role in cybersecurity. However, this move has raised significant red flags, particularly given the NSA's history of mass surveillance and data collection without warrants. Critics, including Edward Snowden, have voiced their concerns that OpenAI's AI capabilities could be leveraged to strengthen the NSA's snooping network, further eroding individual privacy. Snowden has gone so far as to label the appointment a "willful, calculated betrayal of the rights of every person on Earth." The tech community is rightly alarmed, with many drawing parallels to dystopian fiction. The move has also raised questions about ...

What is happening inside of the black box?

  Neel Nanda is involved in Mechanistic Interpretability research at DeepMind, formerly of AnthropicAI, what's fascinating about the research conducted by Nanda is he gets to peer into the Black Box to figure out how different types of AI models work. Anyone concerned with AI should understand how important this is. In this video Nanda discusses some of his findings, including 'induction heads', which turn out to have some vital properties.  Induction heads are a type of attention head that allows a language model to learn long-range dependencies in text. They do this by using a simple algorithm to complete token sequences like [A][B] ... [A] -> [B]. For example, if a model is given the sequence "The cat sat on the mat," it can use induction heads to predict that the word "mat" will be followed by the word "the". Induction heads were first discovered in 2022 by a team of researchers at OpenAI. They found that induction heads were present in ...

Prompt Engineering: Expert Tips for a variety of Platforms

  Prompt engineering has become a crucial aspect of harnessing the full potential of AI language models. Both Google and Anthropic have recently released comprehensive guides to help users optimise their prompts for better interactions with their AI tools. What follows is a quick overview of tips drawn from these documents. And to think just a year ago there were countless YouTube videos that were promoting 'Prompt Engineering' as a job that could earn megabucks... The main providers of these 'chatbots' will hopefully get rid of this problem, soon. Currently their interfaces are akin to 1970's command lines, we've seen a regression in UI. Constructing complex prompts should be relegated to Linux lovers. Just a word of caution, even excellent prompts don't stop LLM 'hallucinations'. They can be mitigated against by supplementing a LLM with a RAG, and perhaps by 'Memory Tuning ' as suggested by Lamini (I've not tested this approach yet).  ...