Token 1.5: From Chain-of-Thoughts to Skeleton-of-Thoughts, and everything in between

Researchers from Google Brain – most of whom authored the original paper –introduced a nuanced technique they termed 'self-consistency.' Rather than overhauling the prompting landscape, this technique tweaks the response-generation phase. It proposes sampling an array of reasoning paths, as opposed to a single, definitive answer.

Intuition Behind the Approach. As problem complexity scales, so does the multiplicity of valid solution paths. Mirroring human cognitive processes, the authors advocate for choosing the most consistent answer among the generated options.

To illustrate that, the process consists of three main steps:

1. Prompt a language model using chain-of-thought (CoT) prompting.

2. Replace the standard step in CoT prompting with sampling from the language model's decoder, thereby generating a diverse array of reasoning paths.

3. Evaluate these reasoning paths and select the most consistent answer as the model's final output.

Emerging five months after the original CoT paper, “Large Language Models are Zero-Shot Reasoners” was the result of a joint effort between the University of Tokyo and Google Brain. Intriguingly, the paper proposes replacing the entire CoT framework with a single phrase: "Let's think step by step."

It means that we’re not already in the few-shot paradigm as we moved to zero-shot. The authors demonstrated that LLMs have an internal “understanding” of CoT that can be invoked by one phrase in the prompt without adding any examples!

The authors of this paper decided to go further in automation. They proposed the Auto-CoT method to automatically construct demonstrations (examples) that were crafted manually under the original CoT approach.

To make Auto-CoT work, we need a dataset with sample reasoning tasks without solutions. Here is how the method works:

1. The method defines clusters of sample tasks and divides them into these clusters.

2. From each cluster, the method selects a representative question and generates the reasoning chain for each of the questions using Zero-shot-CoT.

3. When the user prompts a new task for the model, the method appends the sample tasks and the solutions from step 2 that represent the few-shot examples for the model.

Addressing limitations in LLMs, the Program-of-Thoughts Prompting (PoT) employs a dual-system strategy: leveraging Codex for text and programming statements, subsequently executed by an external interpreter like Python. This approach shines in numerical computations and complex mathematical tasks.

This innovative paper introduces Multimodal-CoT, focusing primarily on the interplay between vision and language.

Multimodal-CoT operates in two distinct stages:

• Rationale Generation. In this initial stage, the model is fed with both language and visual inputs to generate what are termed 'rationales.'

• Answer Inference. Subsequently, in the answer inference stage, the rationale generated in the first stage is appended to the original language input. This modified language input, along with the original visual input, is then fed back into the model to derive the final answer.

While both stages employ the same underlying model architecture, they differ in their input-output dynamics.

Tree of Thoughts (ToT) introduces a novel, structured approach to problem-solving that could revolutionize the way language models tackle complex queries.

Researchers in the domain of human problem-solving have observed that human cognition often navigates through a combinatorial problem space, employing a series of heuristics to guide decision-making.

ToT adopts a more human-like approach to problem-solving by framing each task as a search across a tree of possibilities. Each node in this tree represents a partial solution. The core of ToT can be distilled into answering four essential questions:

• Thought Decomposition: Unlike Chain of Thought (CoT), which doesn't delineate intermediate steps explicitly, ToT utilizes problem characteristics to segment the process into distinct thought steps.

• Thought Generation: This phase leverages two strategies—either sampling independently identical distributed (i.i.d.) thoughts from a CoT prompt, which is ideal for problems with expansive thought spaces, or sequentially proposing thoughts using a "propose prompt," best suited for problems with more constrained thought spaces.

• State Evaluation: At this juncture, the ToT framework uses heuristic methods to evaluate states. There are two strategies under consideration: one that values each state independently and another that casts a vote across multiple states.

• Search Algorithm: Depending on the structure of the problem tree, different search algorithms like Breadth-First Search (BFS) and Depth-First Search (DFS) can be deployed.

Researchers have recently introduced the Graph of Thoughts (GoT) framework, designed to augment the reasoning capabilities of LLMs using a graph-based structure. This framework surpasses existing prompting techniques like Chain-of-Thought (CoT) by providing a structured, extensible mechanism for thought transformations, evaluations, and rankings.

The GoT framework is built as a set of interacting modules:

• Prompter

• Parser

• Scoring module

• Controller.

Each module performs a specialized task in the reasoning process, ranging from preparing the prompt to validating and scoring the generated thoughts. The Controller coordinates these modules and also houses two key elements: the Graph of Operations (GoO) and the Graph Reasoning State (GRS), which further assist in managing the LLM reasoning process.

The authors argue that while CoT generally improves the coherency of solution paths, it's prone to errors and biases. They propose an experimental framework to tackle this issue, relying heavily on search algorithms like Depth-First Search (DFS) and Breadth-First Search (BFS).

Core Concepts:

• Algorithmic Reasoning Pathways: This strategy teaches LLMs to think more like algorithms. It improves the model's reasoning abilities by guiding it through well-defined logical steps.

• In-Context Learning: Instead of pausing and restarting the LLM for each new piece of information, this method allows the model to learn and reason in a more continuous, fluid manner. This reduces computational overhead and costs.

• Single or Few Queries: One of the big wins is the reduction in the number of queries needed for reasoning. This saves time, money, and computational resources.

Researchers have advanced the field of machine learning further with the introduction of the "Skeleton-of-Thought" (SoT) approach, a method that reimagines how LLMs generate text. Instead of building responses sequentially, SoT uses parallelism to enhance speed and accuracy. At its core, SoT creates a concise "skeleton" of the answer, then fills in details in parallel, mimicking the organized way humans think.

Key Components

• Skeleton Stage: Utilizing a template, the LLM crafts a skeletal response to the user's question, from which key points are extracted.

• Point-Expanding Stage: These points are then expanded upon in parallel, leading to a more detailed final answer.

Parallel decoding not only boosts computational efficiency but also significantly reduces end-to-end latency.