SoM VS Constrained-Text-Generation-Studio

At https://openadapt.ai/ we are using LLMs to automate repetitive tasks in GUI interfaces. Think robotic process automation, but via learning from demonstration rather than no-code scripting.

The stack is mostly python running locally, and calling the OpenAI API (although we have plans to support offline models).

For better visual understanding, we use a custom fork of Set-of-Mark prompting (https://github.com/microsoft/SoM) deployed to EC2 (see https://github.com/OpenAdaptAI/SoM/pull/3).

Oh my god. I wrote a whole library called "Constrained Text Generation Studio" where I mused that I wanted a "Photoshop for Text". I'm not even sure which work predates the other: https://github.com/Hellisotherpeople/Constrained-Text-Genera...

The core idea of a "photoshop for text", specifically a word processor made for prosumers supporting GenAI first class (i.e oobabooga but actually good) - is worth so much. If you're a VC reading this, chances are I want to talk to you to actually execute on the idea from the OP

I was working on this stuff before it was cool, so in the sense of the precursor to LLMs (and sometimes supporting LLMs still) I've built many things:

1. Games you can play with word2vec or related models (could be drop in replaced with sentence transformer). It's crazy that this is 5 years old now: https://github.com/Hellisotherpeople/Language-games

2. "Constrained Text Generation Studio" - A research project I wrote when I was trying to solve LLM's inability to follow syntactic, phonetic, or semantic constraints: https://github.com/Hellisotherpeople/Constrained-Text-Genera...

3. DebateKG - A bunch of "Semantic Knowledge Graphs" built on my pet debate evidence dataset (LLM backed embeddings indexes synchronized with a graphDB and a sqlDB via txtai). Can create compelling policy debate cases https://github.com/Hellisotherpeople/DebateKG

4. My failed attempt at a good extractive summarizer. My life work is dedicated to one day solving the problems I tried to fix with this project: https://github.com/Hellisotherpeople/CX_DB8

My mental model for LLMs was built by carefully studying the distribution of its output vocabulary at every time step.

There are tools that allow you to right click and see all possible continuations for an LLM like you would in a code IDE[1]. Seeing what this vocabulary is[2] and how trivial modifications to the prompt can impact probabilities will do a lot for improving the mental model of how LLM operate.

Shameless self plug, but software which can do what I am describing is here, and it's worth noting that it ended up as peer reviewed research.

[1] https://github.com/Hellisotherpeople/Constrained-Text-Genera...

Yeah, the LLM outputs a distribution of likely next tokens. It is up to the decoder to select one, and it can use a grammar to enforce certain rules on the output. https://github.com/Hellisotherpeople/Constrained-Text-Genera... or https://github.com/ggerganov/llama.cpp/blob/master/grammars/... for example.

I am in love with this, I tried my hand at building a Constrained Text Generation Studio (https://github.com/Hellisotherpeople/Constrained-Text-Genera...), and got published at COLING 2022 for my paper on it (https://paperswithcode.com/paper/most-language-models-can-be...), but I always knew that something like this or the related idea enumerated in this paper: https://arxiv.org/abs/2306.03081 was the way to go.

It's far too easy to destroy any type of RLHF done to try to prevent bad behavior from an LLM.

For example, if you want a LLM to generate things that look like social security numbers, you may try to prompt it asking for social security numbers. It will of course give you "I'm sorry hal I can't do that..."

Then start using a technique like token filtering/filter assisted decoding, to make it where the LLM can only generate hyphens and numbers, and suddenly it does what you ask despite RLHF

I explored this a tiny bit in the later sections of my paper studying what happens when you restrict an LLMs vocabulary: https://aclanthology.org/2022.cai-1.pdf#page=17

You can even play with this with open source models using CTGS: https://github.com/Hellisotherpeople/Constrained-Text-Genera...

I agree with you, and I'm SHOCKED at how little work there actually is in phonetics within the NLP community. Consider that most of the phonetic tools that I am using to enforce rhyming or similar syntactic constrained in constrained text generation studio (https://github.com/Hellisotherpeople/Constrained-Text-Genera...) were built circa 2014, such as the CMU rhyming dictionary. In most cases, I could not find better modern implementations of these tools.

I did learn an awful lot about phonetic representations and matching algorithms. Things like "soundex" and "double metaphone" now make sense to me and are fascinating to read about.

https://github.com/hellisotherpeople/constrained-text-genera...

Just ban the damn tokens and try again. I wish that folks had more intuition around tokenization, and why LLMs struggle to follow syntactic, lexical, or phonetic constraints.