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

The AI Dilemma and "Gollem-Class" AIs

From the Center for Humane Technology Tristan Harris and Aza Raskin discuss how existing A.I. capabilities already pose catastrophic risks to a functional society, how A.I. companies are caught in a race to deploy as quickly as possible without adequate safety measures, and what it would mean to upgrade our institutions to a post-A.I. world. This presentation is from a private gathering in San Francisco on March 9th with leading technologists and decision-makers with the ability to influence the future of large-language model A.I.s. This presentation was given before the launch of GPT-4. One of the more astute critics of the tech industry, Tristan Harris, who has recently given stark evidence to Congress. It is worth watching both of these videos, as the Congress address gives a context of PR industry and it's regular abuses. "If we understand the mechanisms and motives of the group mind, it is now possible to control and regiment the masses according to our will without their...

Enhancing LLM Performance: Buffer of Thought and Mixture of Agents

  As Large Language Models (LLMs) continue to advance, researchers are exploring innovative techniques to further enhance their accuracy and usefulness. Two promising approaches in this domain are Buffer of Thought and Mixture of Agents. Buffer of Thoughts The Buffer of Thoughts  technique aims to improve the reasoning capabilities of LLMs by introducing an intermediate step in the generation process. Instead of directly producing the final output, the model first generates a "buffer" or a series of intermediate thoughts, which serve as a scratchpad for the model to reason and plan its response. This buffer allows the model to break down complex tasks into smaller steps, perform multi-step reasoning, and maintain a coherent line of thought throughout the generation process. By externalizing its thought process, the model can better organise its knowledge and arrive at more logical and consistent outputs. The BoT approach has shown promising results in tasks that require multi...

Claude 3.5 Sonnet, literary analysis capabilities, an experiment

 There are those in the AI observer community that have been suggesting of late that 'AI has plateaued' which reveals, to me, a lack of understanding of how models develop. It's not iterative design but step changes we are witnessing. The differences between Claude 3 Sonnet and 3.5 Sonnet are stark.  One test I often carry out to asses the current capabilities of LLMs is to request a simple prompt to analyse an unpublished poem, commentating on the style and literary devices employed. The outputs have improved significantly over the 18 months I have employed this approach. This is my recent attempt. The poem is self written, self published (so not widely available) to ensure that it's unlikely to have found its way into the training data. For added context, implied in the text, this was written during a residency at a Museum and Art Gallery and came from a conversation with a member of staff about his late father. Calm is museum archival software. PROMPT. "analyse...