Predicting Emergent Capabilities by Finetuning

Summary

A central question for the field, given prior work, is whether we can predict emergent capabilities. This paper present a method that shows some success: if the emergent abilities are to be elicited via prompting models, then essentially the point of emergence can be shifted to the left via finetuning models, meaning emergence (when prompting) can be predicted using models of much less compute. 1. The core idea is simple but deft. While it would be easy to dismiss the idea as quite straightforward, I see it as a a nice balance between simplicity (a good quality) and ingenuity (a good quality). 2. The predictive accuracy of the method for "emergence prediction" seems reasonable. Of course, there is nothing too convincing this is the best method or so on, but that seems fine: given this work, other papers can refine the specifics on how to best do this going forward. 3. The scale improvement of the method (i.e. how far can you shift left) seems reasonable. Similarly could be improved.

Rating

7

Confidence

4

Ethics flag

1

Reasons to accept

1. Good idea, which if it works more generically, would be a useful insight for solving a real problem of significant scientific and societal interest (i.e. "can we predict, in advances, whether we will see emergent capabilities") 2. Results/experiments are generally interesting, reasonably executed, and provide some insight. 3. Method seems like a fine starting point, though not especially remarkable.

Reasons to reject

No major reasons to reject. I definitely don't think the paper is without flaws, but none feel so damning that I feel the need to proactively make the arguments against the paper.

Questions to authors

1. You differentiate methods like prefix tuning from LoRA/full fine-tuning in terms of the successfulness of the left shift. Can you characterize some key properties that amount to this (e.g. is there some key property you need for this to work)? Since, if you could, this would then lead to a next question of how we search over methods with this property to identify the best method for maximizing the extent to which the emergence point gets left shited. 2. You identify two axes experimentally: the amount the point shifts to the left and the predictive accuracy. Can you characterize the relationship between these, either empirically or theoretically in some toy model? (Basically thinking about interplay between final 2 paragraphs of page 7).

Summary

This paper proposed a new approach by fine-tuning LLMs for making predictions of the so-called emergent abilities. They proposed using a parametric function to predict when emergence will occur, or "emergence laws," and tested it on four NLP benchmarks where emergence has been observed. Results showed that with small-scale LLMs, they could predict emergence of models trained with more compute than the smaller models.

Rating

5

Confidence

5

Ethics flag

1

Reasons to accept

(1) An interesting extension to scaling law;

Reasons to reject

(1) About the motivation. Generally, I think it is less interesting to see that this work has been done based on fine-tuning data. For LLMs, it is generally useful to predict model performance during the pre-training preriod (especially the beginning) but not the fine-tuning stage. It is very challenging to make accurate prediction on task performance given limited pretraining tokens. Specially, for LLMs, it would be very strange to fine-tune it according some specific task, because it is expected to generalize on donwstream tasks. (2) The evaluation sets. Currently, the examined tasks are widely studied or compared. It is not clear whether the approaches can generalize to more scenarios or tasks, especially difficult tasks. (3) Background and related work seem to be overlapping. I suggest you merge these two parts.

Summary

This paper proposes a model to predict when emergence will occur in LLMs. The idea is as follows. Pretraining loss scales predictably via an inverse power law with the amount of data, compute, or parameters. However, as training loss decreases, at a certain point there is a sharp uptick (popularly termed 'emergence') in performance on certain tasks. The argument of the paper is that rather than the training loss decreasing due to more parameters, compute, or large amounts of data, the decrease in training loss associated with e.g. larger models can be simulated by fine-tuning on the given dataset. The paper proposes a model to predict when the emergence 'elbow' will occur as a function of amount of finetuning data. The paper then takes a series of models and datasets, and fits parameters for their model. It is argued that their model fits the data very well.

Rating

6

Confidence

2

Ethics flag

1

Reasons to accept

- The topic of how and when emergence occurs in models is important

Reasons to reject

- I am not clear what the proposed model really tells us about how and when emergence occurs - I was unclear about some of the notation and claims (details below)

Questions to authors

I may be missing something here, but I didn't understand how the model is useful. I can see that it is interesting to have a model of how and when emergence occurs. However, finetuning is used as a proxy for e.g. using more parameters, and the model a function of the amount of fine-tuning data (I think). So what does this tell us? Can it tell us how to predict emergence points as models grow in size or with the use of more training data? As the emergence point depends on different facets, could the model be extended to depend on more than one variable? In the first equation on p.5, the quantity D is used but not explained (I guess it is the amount of fine-tuning data). The phrase 'taking the limit ... as D approaches the number of examples in our fewshot prompt' occurs on p6. It would be really helpful to write down what the sizes of D and the few-shot prompt are. In Fig 3 there are so many prediction lines that it is not really clear whether the predicted emergence points are realistic. Might be better to have fewer at the low data fractions, say 4 in total. The emergence points for GSM8K seem bad, looks to me like emergence happens at around 2.2 for all amounts of data. Emergence points also look early for MMLU. This is because the RELU puts baseline performance too low, should be around 4% accuracy for data fraction 1 for GSM8K, for example, rather than 2. Related to my first question: the model is fitted using checkpoints from OpenLlama-3B. Then, the paper states that they "then make us of the 7B and 13B checkpoints for validating the accuracy of their predictions". I couldn't see where the paper did this though. Similarly figure 5 was hard to interpret. I understand the LMs we don't have access to are those that are trained for longer. There aren't any predictions for these points though - why are they on the plots? Why is True Emergence only one line? It looks like this corresponds to the emergence loss values in Fig 3 for each dataset, but surely the emergence loss value is going to be different for different models. Why are the failed predictions all on the True Emergence line rather than being placed at their actual locations? (I guess the x value is too far off the plot?) Why do some of the failed predictions have no corresponding LM? I also don't see how to read from the plots the reported fact that 'on MMLU and GSM8K we are able to predict emergence with up to 4x the FLOPS in advance of the emergence point'. It would be good to give a bit more guidance on how to read this. I apologise if these questions are all very basic, but I feel like the plot could be explained better. Minor questions/typos: - Please number equations - Please make sure all your terms are defined.

Summary

The paper presents a method to predict emergent capabilities in language models by finetuning on specific tasks. Using an "emergence law" to model shifts in performance based on finetuning data, the authors successfully predict task emergence on NLP benchmarks with high accuracy, often using significantly fewer computational resources. The study offers a valuable tool for anticipating future AI capabilities and outlines areas for further research to enhance prediction methods.

Rating

6

Confidence

3

Ethics flag

1

Reasons to accept

- The phenomenon primarily investigated in the paper—namely, the observation that directly fine-tuning LLMs on a given task can shift the point in scaling at which emergence occurs towards less capable models—is intriguing and is supported by detailed experimental analysis. - Explored various methods for shifting the point of emergence and provided corresponding analyses. - Consideration of the impact of uncertainty is commendable, and the experimental results are highly reliable.

Reasons to reject

- The paper falls short in providing a thorough explanation for its subpar performance on tasks such as CommonsenseQA and CoLA, failing to convincingly demonstrate the emergence law's robustness regarding these tasks. - The testing was limited to models within the Open-LLaMA V1 series, without indicating whether the emergence law is applicable to models of other architectures. - The task of emergence prediction and the phenomenon primarily studied by the authors, that directly fine-tuning LLMs on a given task can shift the point in scaling at which emergence occurs towards less capable models, do not have a direct relationship, considering that the former does not involve fine-tuning. The paper lacks an explanation of the connection between these two. - The paper fails to provide an explanation for why the emergence law curve takes a linear form. - The paper lacks clear elucidation regarding the meanings of variables in crucial formulas. For instance, in the formula Perf(L, D) = A ∗ max(L − E(D), 0) + B, the roles of L and D are not explicitly defined.

Response to rebuttal

Thanks for your responses. Could you clarify for me how to read from the plots the reported fact that 'on MMLU and GSM8K we are able to predict emergence with up to 4x the FLOPS in advance of the emergence point'. The other responses are fine. Please do make sure all the notation is clarified.

Reviewer response to author rebuttal

Thanks for the response, I have read the discussion with other reviewers as well. I am keeping my score + review as-is and I recommend to the AC that this paper be accepted.