A summary of Emergent Abilities of Large Language Models.

Introduction

• Emergent abilities arise unpredictably, i.e. it cannot be predicted simply by extrapolating the performance of smaller models.
• Scaling laws can be used to predict the effect of scale on the performance of a language model.
• Important Definition
  • Emergence is when quantitative changes in a system result in qualitative changes in behaviour.
• The paper explores emergence with respect to model scale

Emergent Abilities Definition

• Plotting scale ($x$-axis) against performance ($y$-axis) show that there is a phase transition point, a change that occurs only when a certain threshold of scale is crossed.
• It is wise to view emergence as a function of many correlated variables, such as number of parameters, training computation, etc.

Few-Shot Prompted Tasks

• This is an ability of the prompting paradigm.
• Involves providing input-output examples in the model's context (input) before asking it to perform the actual task at hand.
• We can observe a model's phase transition for this ability by plotting its output accuracy for the task against its compute scale. There is an obvious sharp turn where its performance goes well-above random at a certain threshold of scale.
• Example of few-shot prompted task: Recovering a word from its scrambled letters, part of the BIG-Bench benchmark.

Augmented Prompting Strategies

This section discusses other prompting/fine-tuning strategies that augment abilities.

1. Multi-step reasoning (Chain-of-Thought prompting)
   • Prompting that guides them to produce a sequence of intermediate steps before its final output, especially for tasks that require reasoning.
   • Why is it emergent? It only becomes effective with a sufficiently large scale.

   This is often automated in reasoning models these days

2. Instruction following
   • Fine-tune a model on some tasks phrased as instructions.
   • It is then shown that models are able to respond to follow instructions describing a task more appropriately
3. Program execution
   • Fine-tuning the models to predict intermediate outputs and remember it for multi-step computations. (scratchpad)
4. Model calibration

These are many more emergent abilities that we can find out there.

Potential Explanations of Emergence

• Sequential computations that correlate to the number of layers
• More parameters $\implies$ better memorisation
• Imperfection with evaluation metrics
  • Evaluating based purely off of the final product without giving credit to the improvement in intermediate steps may disguise that gradual improvement as a sudden and unpredictable jump in final answer accuracy. (e.g. evaluating exact string match)
  • Issue with explanation
    • Emergent abilities are still observed with classification problems
    • Does not explain the jump in the quality of intermediate steps that have been observed as well
  • An alternative evaluation: cross-entropy loss
    • It is shown that cross-entropy loss improves even for smaller models where the downstream metrics (exact match, BLEU, and accuracy of output) are close to random/do not improve. This implies that downstream metrics are masking underlying improvements.
    • Issue: This new analysis still does not explain why downstream metrics are emergent, nor does it enable any form of prediction of emergence

Beyond scaling

• Abilities may be achieved at smaller scales with improvements in other aspects, not just model scale (e.g. model architecture, high-quality data, training procedure).
• This is an important research path for finding ways to make these abilities more accessible.

Emergent risks

LLMs carry societal risks such as truthfulness, bias, and toxicity. It's important we consider whether or not these risks are emergent as well.

Sociological changes

We should also consider how scale affects not just the model, but the community surrounding it, how it's used, and social perception.

Directions of Future Work

• Further model scaling
• Improved model architecture and training
• Data scaling
• Better techniques for and understanding of prompting
• Frontier tasks (tasks on which performance remains lacklustre)
• Understanding emergence

Glossary

Emergent

• An ability is emergent if it is not present in smaller models, but only in larger models.

upstream/downstream tasks

• Upstream tasks are the tasks that are used for pre-training. They are general problems that the model solves to learn and improve. It's how the language model learns the language, its grammar, syntax, etc.
• Downstream tasks are the practical, specialised tasks that a pre-trained model has to undertake. This is either used for fine-tuning or for practical usage.
• Upstream -> Downstream, it's a pipeline. Downstream tasks are the tasks that are further down the pipeline.
• The idea of pre-training models on upstream tasks and fine-tuning on downstream tasks is also known as transfer learning
• Examples
  • Training GPT to perform sentiment analysis, or text classification, etc.
• Scaling laws are less reliable on downstream tasks

perplexity (of a model)

• Measure of how well a language model predicts a sample of text (from a statistical perspective, not a factual standpoint)
• How well the model mimics human language patterns and predicts a specific sequence of words.
• Perplexity of $1 \implies$ the model is absolutely certain
• Perplexity $=$ vocabulary size $\implies$ the model is very confused, and every token in the vocabulary is equally likely.

In-Context Learning

• The process of enabling a pre-trained model to learn dynamically through prompting without any re-training, simply by providing input/output examples.

Context Window/Length

• The number of tokens that the model can "remember", and take into consideration when generating its output. It includes not just the current prompt, but the conversation history.