LLM Classification Is Feature Engineering

ltbarcly3

I don't understand the point they are trying to make. It's very often (always?) the case that something general also solves particular problems. A sorting algorithm is an implementation of min() A parser also is a syntax checker. A route planner is a reachability checker. A computer algebra system is a basic arithmetic calculator. A general constraint solver is a Soduku hint maker. It's true that LLM output can be used as an input to another classifier, this is also true of any classifier. The improvement on top of the straight LLM classification is relatively small, and I would argue that working on the prompt or just including in the prompt for the LLM what features might be useful to consider would likely work even better. Fundamentally I read this article as: We want to build a simpler, dumbed down clone of Mathematica, so we cobbled together the following pieces... We also needed a way to do arithmetic, so we also include a copy of Mathematica to do basic arithmetic.

softwaredoug

In my work on LLM as a judge, I prefer to use LLM decisions as features in a downstream classic ML model for the final decision. It works really well https://softwaredoug.com/blog/2025/01/21/llm-judge-decision-...

xerlait

Why does he first ask to label "ironic" or "not", and then answer the feature questions? Wouldn't it be better to reverse the order?

twelfthnight

Why not use a text embedder for the unstructured data and concatenate with the structured data? For example you could freeze most of the layers of the embedder but let the final ones learn. Then you wouldn’t need to do either feature or prompt engineering?

dist-epoch

> Calibration / Threshold Control The amount of thinking is relatively calibrated. Ask an obvious classification, you get an instant answer. Ask a tricky one, much more thinking.

drabbiticus

I really wish people would define terms when using math. What is y? What is LLM(x)? Presumably it evaluates to some real number so that it can be fed to the logistic sigmoid function. If it is the logistic function, then why does beta going to infinity matter? It seems to just collapse the output of the sigmoid function to 1 and make the value of LLM(x) meaningless instead of their claim that it recovers the LLM classifier. What is the function I()? Maybe these are well understood terms in some field? Maybe I'm just lost?

elendilm

Well timed article. :)

levocardia

Really needs a comparison to the "megaprompt" itself (i.e. "here is a tweet, rate it as ironic or not, considering the following properties; explain your reasoning then output your final answer at the end"). I bet that would get you very far towards the logistic classifier, and would generalize much better out of distribution.

aleksiy123

You can also get LLM to optimize rules for a rules engine iteratively against some dataset. It’s sort of like memoizing or distilling the knowledge. Works really well for certain type of problems.

vova_hn2

I think that this can be automated by using two LLMs: a stronger/more expensive for generating prompts and a weaker for actual classification. Approximate algorithm: 1. Give "strong" LLM the task formulation and some labeled examples. Ask it to generate a prompt for the "weak" LLM. 2. Run "weak" LLM on the training set with generated prompt from 1, use replies as features for a smaller ML model (logreg, decision tree etc). 3. Pick examples from the training set that your small model is most wrong about and ask "strong" LLM to generate one more prompt (like in 1), except this time you are using the misclassified examples instead of random. 4. Run "weak" LLM on generated prompt from 3, add results as one more feature for your model. 5. Repeat 2 - 4 until your token budget for this task is exhausted or required score on cross validation set is reached. I was thinking about creating an open source library that implements this, but I'm not sure if anyone really needs it. I suspect that people who need something like this already made their own implementation.

iforgotmypasswo

Shouldn’t this article be about the disadvantages of using TypeSafe’s Jev as a classifier? This is a bit of an outdated take as of two days ago. Dear lord things move fast these last few years. Some of this is still relevant. Fine tuning Jev once available could address certain concerns. (Very excited as I got an invite email for TypeSafe today! I don’t have time for all the little experiments I want to run with Jev and Astra combined!)

lhk931122

The structured output that Google provides (the function that blocks the disallowed words) may calculate softmax output with assigning -inf value to the other words, then the probability is not the same as the weight the model firstly put on those allowed words. And here, reasoning comes before verdict, so the decision is already made before the last token classification. Resultingly, the last token made seems to not reflect LLM's original intention.