Join our daily and weekly newsletters to stay informed about the latest developments and exclusive insights in AI security leadership.
Large Language Models excel in responding to straightforward queries, yet necessitate specific prompting approaches to tackle complex tasks demanding critical thinking and strategic planning. By employing so-called “System 2” methodologies, these prompting strategies enhance the logical proficiency of large language models (LLMs) by compelling them to produce incremental solutions for resolving a problem.
While System 2 methods are effective, they often render Large Language Model (LLM) applications gradual and computationally expensive. Researchers at a prestigious institution have pioneered a groundbreaking approach to instructing Large Language Models (LLMs), enabling them to perform complex tasks without the need for intermediary steps.
Researchers in cognitive science have long grappled with the nature of human intelligence and how to replicate it using artificial systems. One prominent framework that has gained traction is the dual-system model, which posits that cognition can be broadly categorized into two distinct types: System 1 and System 2.
System 1 refers to our automatic, intuitive processing mechanisms that operate beneath conscious awareness. This system is responsible for rapid, habitual responses to familiar stimuli, often relying on emotional and sensory cues. Examples of System 1 processes include recognizing a friend’s face or responding to a familiar phrase with a pre-programmed reaction.
In contrast, System 2 represents our controlled, deliberative thinking abilities that operate within the realm of conscious awareness. This system is characterized by slow, effortful processing that involves working memory, logical reasoning, and decision-making. Examples of System 2 processes include solving a math problem, reading a complex text, or making an informed judgment.
The distinction between System 1 and System 2 has significant implications for our understanding of human cognition and the development of artificial intelligence (AI) systems.
In cognitive science, two distinct modes of thinking are referred to as System 1 and System 2, each with its unique characteristics. System 1 thinking is rapid, instinctual, and self-operating. Recognition of familiar patterns, swift decision-making, and comprehension of well-known signs rely on this cognitive mechanism. We employ System 1 thinking to develop visitor indicators, recognize facial expressions, and assign primitive symbols to their corresponding meanings.
System 2 thinking, though, is a slow, intentional, and methodical process. It demands meticulous attention and is applied for tackling complex challenges, akin to deciphering abstract notations, rectifying mathematical formulations or orchestrating travel arrangements.
Large language models are frequently likened to System 1 thinking. While AI models excel at producing text quickly, they struggle to tackle tasks demanding thoughtful contemplation and strategic planning.
Recently, AI scientists have demonstrated the potential for LLMs to simulate System 2 thinking by encouraging them to provide intermediate justifications before delivering their final response. “This approach, ‘”, prompts the large language model to outline its thought process step-by-step, leading to more accurate solutions for logical reasoning tasks.” Several system 2 prompting methods are specifically designed to cater to distinct tasks.
While many of these strategies have been shown to yield superior results due to their underlying logic, they occasionally achieve this at the cost of significantly higher inference complexity and longer response times. Due to this constraint, most of these methods are rarely applied in manufacturing processes that primarily rely on System 1 thinking.
System 2 distillation
A fundamental observation about human cognition is that when we consistently perform a task requiring conscious attention, it gradually becomes an automatic process, assimilated into our subconscious thinking. When learning to drive, one applies deliberate concentration to control the vehicle, adhere to traffic regulations, and chart a course. As one gains increased expertise, driving becomes an innate skill. You don’t need to meticulously deliberate over each step; instead, you’ll perform them instinctively and mechanically.
The intriguing occurrence prompted Meta AI’s researchers to devise a solution, namely “System 2 distillation,” specifically designed for large language models (LLMs).
Machine learning employs distillation techniques where a larger model, referred to as the “teacher,” is used to train a smaller model, also known as the “student.” For instance, developers often utilize frontier models such as Claude to create training examples for smaller models like.
However, System 2 distillation does not require a distinct instructional model. Researchers found a way to translate insights gleaned from their AI model’s slower, more analytical System 2 processing into its faster, more efficient System 1 architecture?
The approach initiates by encouraging the large language model (LLM) to resolve a problem using System 2 prompting techniques. The verification process involves a self-contained mechanism that checks the responses for accuracy without external guidance. They employ “self-consistency,” where the mannequin is consistently presented with the same prompt on multiple occasions. The solution exhibiting the highest frequency is ultimately selected as the correct answer and used to compile the distillation dataset. If inconsistency prevails among proposed solutions, the entire instance – along with its potential answers – is dismissed.
Consequently, System 2 processing eliminates all intermediate calculations, retaining only the final answers. Ultimately, they refined the mannequin based on the initial query and response. This allows the mannequin to bypass logical thinking processes and directly generate a response without considering underlying rationalizations.
System 2 distillation in motion
Researchers assessed their methodology across diverse reasoning tasks and four distinct System 2 prompting approaches. The researchers utilized Llama-2-70B for the lower mannequin, a variant capable of processing and internalizing novel knowledge with considerable scale.
The System 2 approaches employed in their experiments effectively utilized a combination of Chain-of-Thought, Department-Resolve-Merge, and potentially other mental processes to achieve their desired outcomes. Several of these techniques necessitate prompting the mannequin multiple times, rendering them laborious and expensive. The system initiates a process where it requests the mannequin to reformulate the initial inquiry with added clarification, subsequently prompting the mannequin with the reworded question. The Department-Resolve-Merge process proves particularly challenging, necessitating multiple iterative exchanges with the mannequin.
Findings indicate that System 2 distillation significantly boosts the performance of large language models (LLMs) on complex reasoning tasks, often rivaling or surpassing the original System 2 approaches in terms of accuracy. Furthermore, distilled models can produce answers significantly faster and with substantially reduced computational requirements due to their exemption from undergoing complex intermediate reasoning processes.
As researchers found, the process of distillation proved lucrative in filtering out biased perspectives and extraneous data that would normally impede decision-making through System 2 Consideration. The rephrased text: It also substantively confirmed impressive outcomes in certain reasoning tasks, where Rephrase and Reply are employed to clarify and refine responses, as well as facilitate nuanced analysis and processing through.
The team has demonstrated that, in various scenarios, it is feasible to condense System 2 thinking into LLM outputs without requiring intermediate iterations while often achieving improved efficiency.
Despite these findings, researchers also uncovered that LLMs, similar to humans, struggle to condense various reasoning capacities into their swift inference process. Despite their best efforts, the team has struggled to effectively condense complex mathematical proofs into concise and easily digestible formats. Some tasks may necessitate deliberate contemplation.
While System 2 distillation has shown promise in certain contexts, its full potential may not be fully understood, particularly when applied to smaller scales or extended to tasks beyond those used in training. It’s essential to acknowledge that LLM benchmark results can be compromised by contamination, where the model is already familiar with the test examples, thereby yielding inflated scores on testing sets.
Despite this, distillation is likely to emerge as a potent optimisation tool for refined LLM pipelines executing specific tasks at each stage.
As professionals anticipate future challenges, innovative approaches will streamline essential tasks, freeing up more time for critical thinking about the responsibilities that can only be performed efficiently, just like individuals do naturally.
