Thread by @hwchase17

A 🧵on results of a little investigation I did over the past week

❓How does text-davinci-003 do on agent-like tasks

TLDR: Displays superior understanding and ability to take multi-step actions towards original goal

h/t to @OfirPress and @momusbah for help/feedback with this

✨text-davinci-003 came out a week ago

Lots of people (like @blennon_ below) provided some great analysis of how it compared to `text-davinci-002`

This is my contribution to that, focused on agent-like tasks

https://twitter.com/blennon_/status/1597374826305318912?s=20&t=6jiirRJBE5ecsevnXvLGQA

What are agent-like tasks?

I focus mainly on exploring situations where the LLM is used as agent that has access to some tools, and needs to answer a question using the tools it has available. For example:

@OfirPress Self Ask with Seach
@ShunyuYao12 ReAct with Wikipedia

Why focus on agent-like tasks?

These tasks involve the LLM figuring out when it needs to use a tool, then observing the result of using that tool, and then taking another action based on that observation

IMO these are all good proxies for intelligence

What did I evaluate?

1⃣Self-Ask with Search on @OfirPress compositional celebrities dataset
2⃣Anecdotal failure modes @momusbah found
3⃣ReAct on the HotPotQA dataset

Note that these are definitely NOT comprehensive, and if anyone is interested I'd love to collaborate on more

1⃣ First evaluation: Self-Ask (with Search) on the Compositional Celebrities Dataset

This is a dataset @OfirPress created to judge the LLMs reasoning ability on multi-hop question answering

Read more about it here: ofir.io/self-ask.pdf

In addition to self-ask with search, we also benchmarked regular self-ask (as a baseline)

Big thank you to @OfirPress for helping run some of this evaluation + reasoning through it

On the compositional celebrity dataset we benchmarked four things:

Self ask (no search) with -002
Self ask (no search) with -003
Self ask with search with -002
Self ask with search with -003

Here are the results

We can see improves on both normal Self Ask and Self Ask with Search

NOTE: the reason that self ask with search actually does worse is that a lot of times the search wrapper we are using does not return good results

Here is the notebook we used colab.research.google.com/drive/1LP_KTWX…

The fact that self-ask with search showed a large improvement (0.34 -> 0.41) while normal self-ask did not suggests to me that one of the big improvements of -003 is in its ability to use and interact with external tools (rather than its chain of thought like reasoning)

2⃣ The second method I used to evaluate was some anecdotal failure modes @momusbah provided

We had actually been talking the day before `text-davinci-003` came out about these, so it was great timing!

Let's take a look at some of these

Notebook here: colab.research.google.com/drive/1UVnBEHg…

The first example shows `text-davinci-003` being better at keeping the original question in mind and not getting distracted

We can see it starts with the same original question as 002, but whereas 002 gets a bit lost, 003 asks good followups

The next two examples show text-davinci-003 doing a bit better with the concept of time, namely handling current events

Specifically, it understands the underlying user intent that the user is asking for CURRENT information, whereas text-davinci-002 does not

However, it is not perfect.

Here is one example where it (along with text-davinci-002) struggles to understands more complex ideas

Although even here its answer is MUCH more reasonable than 002 (which seems to forget the original question)

3⃣Finally, let's evaluate it a third way: using a ReAct agent with access to Wikipedia on the HotPotQA dataset

The ReAct Agent is based on a framework @ShunyuYao12 came up with (see below tweet)

Notebook here: colab.research.google.com/drive/1UVnBEHg…

https://twitter.com/ShunyuYao12/status/1579475607402217472?s=20&t=6jiirRJBE5ecsevnXvLGQA

I was somewhat resource constrained here and so only ran it on a few examples from the dataset

But we can see a lot of similar trends as before: text-davinci-003 does a much better job of understanding and remembering the original question

The difference in this first example is small

text-davinci-003 remembers the original question and responds appropriately (yes/no rather than with the nationality)

But it also goes off on a bit of a tangent (before getting back on track) which is a bit weird...

In this next example, text-davinci-003 is much better at executing down a sequence of more complicated reasoning logic

And in this final example, it again displays ability to continue executing in a more long form nature

My main takeaways from all of this are that text-davinci-003:

- displays superior understanding of intent
- displays superior ability to do take multistep actions towards the original objective

This is far from a comprehensive analysis! I found myself a bit time and resource constrained

If you are interested in collaborating on evaluation of LLMs on agent-like tasks please reach out - I would love to do so!