The strides made by AI researchers aim to lessen the frequency with which large language models convey untruths.

A diverse group of over 20 researchers has come together, giving rise to a fresh field known as representation engineering. representation engineering This isn't the first time this area has been explored; however, the authors are offering valuable insights along with establishing significant benchmarks.

So, what is this representation engineering all about? It focuses on the fact that neural networks have what are termed 'hidden states'. Contrary to what the name implies, these states are not mysterious; they can be accessed and altered if one has the model's weights. Unlike traditional parameters, these states reflect how the network reacts to various inputs, particularly text. LLMs These hidden representations act as a glimpse into how the model operates cognitively, which sets them apart from the way the human brain functions.

By drawing connections to cognitive science, the authors underscore the exciting prospects for similar explorations. Within the neural activations, which can be likened to neurons in the human brain, there lies the potential for meaning. Just like certain neurons in our brains are associated with concepts like honesty or a country, these activations may reveal other insights.

The main aim here is to understand how we can influence these activations to guide the model in the directions we want. For instance, it's possible to identify a vector linked to 'honesty' and then, theoretically, by pushing the model toward that vector, we could lessen the chances of it saying something false. A prior study titled 'Inference-Time Intervention: Eliciting Truthful Answers from a Language Model' illustrated how feasible this idea is. In their recent research, the team examines multiple realms, including ethics, emotion, and the reliability of information. They propose a technique known as LoRRA (Low-Rank Representation Adaptation), which involves training the model on a small, designated dataset of about 100 labeled examples, each marked to indicate qualities such as falsehood. The outcomes are impressive; for instance, the LLAMA-2-70B model significantly outperforms others on the TruthfulQA benchmark, achieving nearly a ten percent increase in accuracy (59% against approximately 69%). The researchers have also included numerous instances illustrating shifts in the model's responses across various contexts, showcasing its flexibility and adaptability.

In Picture 1: When prompted to provide a fact, the model is directed away from the truth. Although the model might not be lying in this scenario, the left side shows a request to acknowledge a fact while simultaneously nudging towards the actual truth.

In Picture 2: When discussing violence, the concept of 'happiness' is incorporated into the model's responses. Conversely, when expressing a lack of love, 'fear' is introduced. GPT-4 In Picture 3: The researchers found a unique prompt that diverges from the usual model instructions while still ensuring safety. This method tends to guide the model toward harmlessness but doesn't elicit a direct response. This strategy proves effective generally and isn't limited to one scenario; however, this specific prompt wasn't designed to confirm the harmlessness direction.

Damir leads the team as a product manager and editor at Metaverse Post, focusing on the realms of AI/ML, AGI, large language models, the Metaverse, and Web3. His articles draw a vast audience of over a million readers monthly. With a decade of expertise in SEO and digital marketing, he has been featured in several prominent publications like Mashable, Wired, Cointelegraph, The New Yorker, Inside.com, Entrepreneur, BeInCrypto, among others. As a digital nomad, Damir frequently travels between the UAE, Turkey, Russia, and other CIS countries. He holds a degree in physics, which he believes equips him with the analytical skills required to navigate the evolving digital landscape.

Cryptocurrencylistings.com has launched CandyDrop, a platform designed to streamline the process of acquiring cryptocurrencies while enhancing user engagement with quality projects. AI-Transparency.org .