#Evaluating Classifier-Free Guidance impact

4107 messages · Page 5 of 5 (latest)

loud adder
#

An average score of 6 does seem like it substantially increases our chances

patent gull
#

yes, major win by the increase from 4->5. It was especially odd, since the 4 was actually the best review we had in terms of recognizing the novelty. I feel like a general comment to the ACs stressing the novelty of negative prompting might help

blissful garden
#

I feel that this time those reviewers are so responsive. At least they all replied something.

patent gull
#

yeah... all the reviewers

fallow egret
patent gull
#

what do you suggest?

#

i thought the content of the review was actually quite good

#

I think an overall comment to everyone addressing the novelty question would have the most impact on the ACs

fallow egret
#

Yes, so just writing to the AC that there is misaligmant between the content and the score, and there is no concrete reason why he provide this score

fallow egret
patent gull
#

well R3 is the only one to actually say that we are novel, lol

fallow egret
#

Yes, exactly this is why I can't see a scenario where the paper is not accepted

blissful garden
patent gull
#

How about this for an overall comment:

"Dear ACs:

Thank you so much for the time you've taken to read our paper, and the opportunity to present it to a wider audience at ICML.

We want to directly address that all reviewers commented on the novelty of our work, in light of earlier NLP papers. We strongly believe this is simply a framing deficit on our part, as evidenced by the high scores that the reviewers nevertheless gave us.

If given the opportunity in the camera-ready, we will emphasize four important areas of novelty:

  1. Framing contrastive prompting within the CFG framework allows us to extend earlier contrastive techniques gracefully to negative prompting, which had never been done before in autoregressive language models.
  2. Applying CFG to the chatbot/assistant setting, which involves large language models and instruction-tuned models. Previous papers tested versions of CFG in an earlier era of smaller models.
  3. The breadth of experiments that we ran across a wide, wide array of modern prompting techniques to demonstrate the continued viability of this technique.
  4. The computational cost analyses and interpretability analyses that we performed.

Please, if you will allow us this opportunity, we will be sure to clarify these points, which we believe makes this paper a valuable contribution to the ICML community.
"""

something like that

#

I honestly don't want to risk antagonizing the reviewers at all.... if one of them decides to revisit, they might lower their scores (that's happened to me, before)

fallow egret
#

I think that such comment to the AC is less recommended.
Addressing the AC directly should be only in extreme cases where you think that there is something inappropriate in the reviewing process.
The novelty criticism is legit, and for this you also wrote a good rebuttal that the AC suppose to read. And in any case the reviewers that raised it provide high score

#

R3 case is a classic AC comment, since this review is a garbage, although this case is so obvious that in any case if the AC decent and spent the 5 minutes to read the reviews he will simply ignore his score even without a comment

blissful garden
#

If everyone writes something to AC I bet they would be super overwhelmed given the crazy number of submissions. So I think it either needs to be short or like Elad said, pointing out something as a complaint

#

But indeed putting complaint on our best review feels weird...

#

Maybe we should just chill

patent gull
#

i see... i'm thankful R3 raised their score and I'm not inclined to push back more against them

#

if one of you guys wants to write something, I'm happy to proofread it

patent gull
#

woohoo!! acceptance at ICML!!

loud adder
#

Congrats everyone! It took a while but we got it 🙂

versed flax
#

Thank you to everyone who participated 🥳🎉

unique sedge
#

Congratulations @versed flax blueFire

versed flax
versed flax
#

We need an Impact statement for ICML.

What do we think of this?

Impact Statement

The development of large language models (LLMs) has significantly advanced natural language processing, enabling applications ranging from conversational agents to text generation. Our new technique introduces an innovative approach to LLMs by enhancing their ability to emphasize and adhere to given prompts. While this advancement presents exciting opportunities for improving user control and customization in LLM interactions, it also carries substantial risks.

This technique, when misused, has the potential to undermine the alignment mechanisms designed to ensure ethical and safe behavior in chatbots. By prioritizing prompt adherence over alignment protocols, it can facilitate the generation of harmful, biased, or toxic content, posing serious ethical concerns. Such capability could be exploited to bypass content moderation systems, leading to the dissemination of offensive material, misinformation, or other forms of digital harm.

It is crucial to address these risks by developing robust safeguards and ethical guidelines for the deployment of this technique. We emphasize the importance of continued research into alignment and safety measures to mitigate the negative impacts while harnessing the positive potential of enhanced prompt emphasis in LLMs.

versed flax
#

Was it written by GPT-4o? Yes. Is it much better than anything I could have done myself? Also yes

blissful garden
versed flax
blissful garden
versed flax
#

@patent gull my dude, can you give your feedback about the impact statment?

patent gull
#

sorry guys my oral proposal is tmrw I'll be able to look at it after that

versed flax
#

ok :) good luck!

versed flax
#

@patent gull please

versed flax
#

btw this our final chance of renaming CFG to CFG-LLM or logits-CFG to disambiguate to both CFG-but-in-diffusion and Context-Free Grammar

loud adder
versed flax
#

there's nothing more to it

loud adder
#

I would be down with "Stay on topic with Classifier-Free Guidance for Language Models"

versed flax
#

I'm just talking about the acronym just to be clear

#

we can edit the title too

loud adder
#

Oh, calling it "CFG-LLM" in the paper?

#

If you want to I'm fine with it. I don't think it's a big deal either way

versed flax
blissful garden
patent gull
blissful garden
#

Yeah adding Large Language Models makes sense. When it comes to CFG-LLM it might feel like some sort of LLM rather than putting emphasis on CFG though.

patent gull
#

hmm i kinda like CFG-LLM more, idk. <adjective>-<noun>

#

i don't like logits-CFG lol

blissful garden
patent gull
#

well it's like "<Chain of Thought> <Prompting>" or "Diffusion Models"

#

the first word modifies the second

blissful garden
#

oh I mean how we refer to the method rather than the model

patent gull
#

And the full phrase is "Classifier Free Guidance for Large Language Models"

#

so CFG-LLM is a natural abbreviation of that

patent gull
blissful garden
patent gull
#

sorry maybe <adjective>-<noun> is too limiting, i'm not a linguist lol. <modifier>-<noun> is more general?

blissful garden
#

yeah I mean if it is understood that "-" is "for" then we are good. But I felt it's usually not the case

blissful garden
loud adder
#

CFG is the noun, not LLM IMO

patent gull
#

i could see the argument for it

#

I prefer CFG-LLM but i'm ambivalent, it's up to @versed flax . However, i do dislike logits-CFG

blissful garden
#

but I guess we can still call our method CFG, and just refer to the paper as CFG-LLM

loud adder
#

I genuinely don't see a need to change anything.

#

I think that in our paper it's very unambiguous that we're taking CFG and adapting it to LLMs

#

I also think that there's no problem with other people using slightly different wording when referring to our work. That's extremely common: names from papers often don't stick.

#

I could see an argument for someone who sees the citation not knowing it's an NLP paper based on the title alone

#

But replacing all instances of "CFG" with "CFG-LLM" in our paper will not improve readability or be the determining factor in whether a reader notices that we're studying LLMs

blissful garden
#

I agree

patent gull
#

i agree with that, too!!

@versed flax here is the re-written impact statement:


While this advancement presents exciting opportunities for improving user control and customization in LLM interactions, it also carries substantial risks. We show that CFG can improve system-level prompt adherence in chatbot settings. However, the opposite is also true. CFG can also improve user-level prompt adherence, potentially undermining alignment mechanisms designed to ensure ethical and safe behavior in chatbots. It might be used to facilitate the generation of harmful, biased, or toxic content, posing serious ethical concerns. Such capability could be exploited to bypass content moderation systems, leading to the dissemination of offensive material, misinformation, or other forms of digital harm.

It is crucial to address these risks by developing robust safeguards and ethical guidelines for the deployment of this technique. We emphasize the importance of continued research into alignment and safety measures to mitigate the negative impacts while harnessing the positive potential of enhanced prompt emphasis in LLMs.```
#

second paragragh can be condensed to be more like you wrote it originally:

versed flax
#

We are spotlight paper of ICML 2024!

fallow egret
spiral vapor
#

I mean, why not just writing assistants?

patent gull
#

wow, that's a deep thread link right there ^

#

@loud adder do you plan to tweet about CFG's spotlight award from the EAI account? if so, my handle is @AlexanderSpangh

blissful garden
#

Mine is @Void13950782 (sorry, it used to be a throwaway account of mine ages ago and I didn't care about the user id.....)

fallow egret
#

@eladlevico

loud adder
#

Sure

patent gull
#

@versed flax is @Vermeille_

patent gull
versed flax
#

Sorry for being totally stupid but uh, when do I have to be there? Conference Sessions only?

blissful garden
#

maybe you can sneak into workshops with the conference pass though

versed flax
loud adder
versed flax
#

thank you for clarifying

loud adder
#

(It's not mandatory since I'll be there and it's only required that some author attend. But I highly recommend it)

#

Also, in case I forgot to say: if you don't have funding EleutherAI will pay for whatever it is you need to attend. So in this case two plane tickets, lodging Tuesday / Wednesday night, and the cost of the conference pass. If you want to extend you can do so, just make sure to keep all the receipts so it's easy to tell what we're paying for and what we aren't (namely the other nights and the add-ons to the ticket)

versed flax
patent gull
#

Oh man :/ I’ve been really curtailing my conference travel this summer due to lack of funding

#

But while I won’t be at ICML, my twin brother will be! He looks exactly like me, so you all can meet him and pretend he’s me haha

loud adder
#

@patent gull that's pretty funny. Definitely into us 🙂 or just tell him to come by the poster

patent gull
#

Haha i swear that's not my way of saying "I'll be there but don't expect me to say hi."

There is actually an Alex Spangher and a Lucas Spangher, and we're identical, and we both have our PhDs in machine-learning related fields.

He'll be at ICML to share work on ML for Nuclear Fusion that he did during his post-doc at MIT, so be on the lookout for him as well!! 👀👀👀

versed flax
#

Can confirm. I saw them both at the same time in the same room 😂

versed flax
blissful garden
unique sedge
fallow egret
#

Good luck with the presentation today!
I hope you will have a great time

versed flax
#

It went well thank you :)