Simply adding "Repeat the question before answering it." som...

Simply adding "Repeat the question before answering it." somehow make the models answer the trick question correctly. 🤔 Probable explanations:✨ 📌 Repeating the question in the model's context, significantly increasing the likelihood of the model detecting any potential "gotchas." 📌 One hypothesis is that maybe it puts the model into more of a completion mode vs answering from a chat instruct mode. 📌 Another, albeit less likely, reason could be that the model might assume the user’s question contains mistakes (e.g., the user intended to ask about a Schrödinger cat instead of a dead cat). However, if the question is in the assistant’s part of the context, the model trusts it to be accurate. 📚 The related Paper is EchoPrompt which proposes this techniques to rephrase original Prompt/queries before answering them. improves the Zero-shot-CoT performance of code-davinci-002 by 5% in numerical tasks and 13% in reading comprehension tasks.

Rohan Paul

x.com

RelatedHighlights

Observing these patterns of influence gives clues about how our models generalize from their training data. For instance, if the models responded to user prompts by splicing together sequences from the training set, then we’d expect the influential sequences for a given model response to include expressions of near-identical thoughts. Conversely, i... See more

Anthropic \ Tracing Model Outputs to the Training Data

Nicolay Gerold added

Memorization vs learning concepts and forming models of the world.

"Role prompting"... telling the model to assume a role has never been a good way to elicit capabilities/style/etc. For instance, if you ask one of the Claude models to simulate Bing Sydney, assuming you can get it to consent, the simulation will probably be very inaccurate. But if you use a prompt that tricks them into predicting it indirectly (https://t.co/wJEAlPgfz6... See more

j⧉nus

x.com

Nathan Storey added

Something that’s been on my mind is flipping the relationship between the human and language model when going through a creative process. It seems that we often want to ask questions of language models, and we expect them to brainstorm ideas or give us answers, but I wonder if another fruitful pattern here is having models ask questions of us .

ChatGPT as muse, not oracle

sari and added

Innovations like prompting, chain-of-thought reasoning, retrieval grounding, and others are needed to educate models.

Ben Auffarth • Generative AI with LangChain: Build large language model (LLM) apps with Python, ChatGPT, and other LLMs

Retrieval practice prompts should produce

consistent

answers, lighting the same bulbs each time you perform the task. Otherwise, you may run afoul of an interference phenomenon called “retrieval-induced forgetting”

This effect has been produced in many experiments but is not yet well understood. For an overview, see Murayama et al, Forgetting as a c... See more

Andy Matuschak • How to Write Good Prompts

added

We find that models learn just as fast with many prompts that are intentionally irrelevant or even pathologically misleading as they do with instructively “good” prompts. Further, such patterns hold even for models as large as 175 billion parameters (Brown et al., 2020) as well as the recently proposed instruction-tuned models which are trained on

Albert Webson • Do Prompt-Based Models Really Understand the Meaning of their Prompts?

Isabelle Levent added

"Violation of Expectation via Metacognitive Prompting Reduces Theory of Mind Prediction Error in Large Language Models"

This paper investigates how Violation of Expectation (VoE) and metacognitive prompting can be used to reduce Theory of Mind prediction errors in Large Language Models (LLMs) in the context of human-AI interaction.

arxiv.org

Bhaumik Patel added

Could you have GPT-4 ask questions to the user rather than force the user to do prompt engineering? An experiment says this gave superior results. My read is that they tested this on easy cases where we already knew it would work, so we did not learn much.

Shortwave — rajhesh.panchanadhan@gmail.com [Gmail alternative]

Nicolay Gerold added