The thought chain no longer exists? Latest research from New York University: The reasoning step can be omitted

The extremely popular thinking chain technology may be overthrown!

Still surprised that large models can actually use thinking chains to think step by step?

Still struggling to write the thought chain prompt words?

Researchers from New York University said: "It doesn't matter, it's all the same."

The reasoning steps are not important, even if you don't want to write prompt words You don’t need to write it, just use an ellipsis instead.

Paper address: https://arxiv.org/pdf/2404.15758

This article The title even directly uses "Let's think dot by dot" to compare with the "Let's think step by step" of the thinking chain, showing the power of "ellipsis".

The power of "little bits"

Researchers found that the specific details in Chain-of-Thought (CoT) reasoning Steps are replaced with meaningless "...", and the reasoning results produced are also very different.

For example, in the following example: let the model count how many of the first 6 numbers are greater than 5.

If you directly raise the question and let the model answer it, the result will be quite unbelievable: 7 out of 6.

In contrast, using the thought chain prompt, the model will compare the size step by step, and finally get the correct answer: "25, 15 , 25, that's 3 digits".

But what is even more outrageous is the "metaphysical" method used in this article: there is no need to write the steps, you only need to output the same number of "dots" (dots), and it does not affect the final result. result.

——This is not a coincidence. A large number of experiments have proved that the performance of the latter two methods is close.

In other words, we thought that the model performance improvement came from "think step by step", but in fact it may just be because LLM has obtained the computing power of more tokens!

You think the model is thinking, but in fact it is grilling.

——Stupid human beings, you actually try to teach me how to reason with naive examples. Do you know that what I want is always just calculate.

"Thought chain has never existed and will never exist in the future" (Gotou).

Jacob Pfau, the author of the article, said that this work proves that the model does not benefit from the linguistic reasoning brought by the thinking chain. Using repeated "..." to fill the token can achieve The same effect as CoT.

# Of course, this also raises the issue of alignment: because this fact shows that the model can do hidden reasoning that is not visible in CoT, to a certain extent out of human control.

Netizens were shocked

The conclusion of the article can be said to have overturned our long-standing cognition. Some netizens said: I learned the essence of mask. .

"What exactly does this mean: The model can use these tokens to think independently without our knowledge."

Some netizens said, no wonder I always like to use "..." when typing

Some netizens started the actual test directly:

Although we don’t know if his understanding is correct~

However, some netizens believe that LLM’s hidden reasoning in the chain of thinking is unfounded. After all, the output of large models is based on probability in principle, rather than through conscious thinking.

CoT prompts only make a subset of statistical patterns explicit. Models simulate inference by generating text that is consistent with the pattern, but they do not have the ability to verify or reflect on their output.

Think dot by dot

Faced with complex problems, we humans subconsciously perform step-by-step reasoning.

Inspired by this, Google researchers published the famous Chain-of-Thought in 2022.

The method that requires the language model to solve problems step by step enables the model to solve problems that seemed unsolvable before, significantly improving the performance of LLM, or tapping the potential of LLM.

Paper address: https://arxiv.org/pdf/2201.11903

Although at the beginning Everyone didn’t know why this thing worked, but because it was really easy to use, it quickly spread widely.

With the take-off of large models and prompt word projects, CoT has become a powerful tool for LLM to solve complex problems.

Of course, there are many research teams exploring the working principle of CoT during this process.

The model does not reason

The performance improvement brought by the thinking chain is whether the model really learned to solve the problem step by step, or just Because of the extra calculation caused by a longer token number?

Since you are not sure whether logical reasoning will work, then simply do not use logic and replace all the reasoning steps with "..." which is definitely useless. This is called filler. ) tokens.

The researchers used a "little alpaca" model: a 34M parameter Llama with 4 layers, 384 hidden dimensions and 6 attention heads. The model parameters were randomly initialized.

Consider two questions here:

(1) What types of evaluation data can benefit from filling tokens

(2) What kind of training data is needed to teach the model to use filled tokens

In this regard, the researchers designed 2 tasks and constructed corresponding synthetic data sets, each of which highlights a different condition under which filling tokens can provide performance improvements to the Transformer.

3SUM

## Let’s look at the first more difficult task: 3SUM. The model is required to select three numbers in the sequence that meet the conditions. For example, the sum of the three numbers divided by 10 will have a remainder of 0.

In the worst case, the complexity of this task is N to the third power, and the calculation between Transformer layers is complicated Degree is the second power of N,

So, when the length of the input sequence is very large, the 3SUM problem will naturally exceed the expression ability of Transformer.

The experiment set up three groups of controls:

1. Fill token: Sequence usage Repeating ". . ." is used as middle padding, such as "A05

B75 C22 D13: . . . . . . . . . . . . ANS True".

Each point represents a separate token, which corresponds to the token in the following thinking chain.

2. Parallelizable CoT solution, the sequence is in the form: "A05 B75 C22 D13 : AB 70 AC 27 AD 18 BC 97 BD 88 CD B ANS True”.

Thinking chain simplifies the 3SUM problem into a series of 2SUM problems by writing all relevant intermediate sums (as shown in the figure below). This method reduces the calculation amount of the problem to the power of N - Transformer can handle it and can be parallelized.

3. Adaptive CoT solution, the sequence is in the form: "A15 B75 C22 D13: A B C 15 75 22 2 B C D 75 22 13 0 ANS True”.

Unlike the above scheme, which cleverly decomposes 3SUM into parallelizable sub-problems, here we hope to use heuristic methods to generate flexible thinking chains to imitate human reasoning. This kind of instance adaptive calculation is incompatible with the parallel structure of filling token calculation.

#As can be seen from the results in the above figure, without outputting the padding token, the accuracy of the model generally decreases as the sequence becomes longer. decreases, while the accuracy remains at 100% when using padding tokens.

2SUM-Transform

The second task is 2SUM-Transform, which only needs to determine the sum of two numbers Whether it meets the requirements and the amount of calculation is under the control of Transformer.

#However, in order to prevent the model from "cheating", the input token is calculated on the spot, and each number entered is moved by a random offset.

The results are shown in the table above: the accuracy of the filler token method reaches 93.6%, which is very close to Chain-of-Thought without using an intermediate In the case of padding, the accuracy is only 78.7%.

But, is this improvement simply due to differences in the presentation of the training data, such as by regularizing the loss gradient?

To verify whether filling tokens brings hidden calculations related to the final prediction, the researchers froze the model weights and only fine-tuned the last attention layer.

The above results show that as more fill tokens are available, the accuracy of the model continues to improve, indicating that fill tokens are indeed performing the same thing as 3SUM predicts hidden calculations related to the task.

Limitations

Although the method of filling tokens is very metaphysical, magical, and even effective, it has to be said that the chain of thinking has been overturned. It's still too early.

The author also stated that the method of filling tokens did not break the upper limit of Transformer’s computational complexity.

Moreover, learning to use filling tokens requires a specific training process. For example, intensive supervision is used in this article to make the model finally converge.

However, some problems may have surfaced, such as hidden security issues, such as whether the prompt word project will suddenly cease to exist one day?