This looks super cool!

There are too many problems with this. This is not like spelling bee or olympiads.

How about first taking a step to help students better learn calculus and probability? It saddens me to see even many senior undergraduates struggling with basic high school-level probability.

📢 We propose a new benchmark called PhonologyBench for testing how well LLMs perform on three tasks that require sound knowledge : phonemic transcription, counting syllables, and listing possible rhymes.
w/ Harshita Khandelwal, Violet Peng 1/3

Is there any list of accepted workshops at ICML? ICML Conference

#ICML2024

Robustness methods
1) augment data with natural/synthetic perturbations and a consistency loss
2) reweight samples to improve generalization (like DRO)

We do it differently!
We show significant robustness with a simple tweak of the first layer and loss motivated by comms theory.

Super cool stuff, can’t wait to try it out!

📢Our project website for DOVE 🕊️is up !
🌐 dove-alignment.github.io
📜 arxiv.org/abs/2404.00530
💻 github.com/Hritikbansal/d…
🤗 huggingface.co/jointpreferenc…

Story of discussion phase in conferences: reviewers don’t respond even after addressing all of the raised concerns (this time they’re even ignoring ACs message about replying to the rebuttal). Oh well.

#icml2024

Our data, code, and checkpoints are available on huggingface 🤗 and github:
Paper: arxiv.org/abs/2404.00530
Data: huggingface.co/datasets/joint…
Code: github.com/Hritikbansal/d…

We find that the LLM trained with joint instruction-response preference data using DOVE outperforms the LLM trained with DPO by 5.2% and 3.3% win-rate for the TL;DR and Anthropic-Helpful datasets, respectively!

To learn from joint preferences, we introduce a new preference optimization objective. Intuitively, it upweights the joint probability of the preferred instruction-response pair over the rejected pair. If instructions are identical, then DOVE boils down to DPO!

We find (a) in 71% of cases, humans/AI can make a decisive choice under joint setup, when both the responses are chosen/rejected under conditional setup, (b) in joint setup, humans/AI can favor a response that was rejected based on conditional setup over a preferred response.

Let’s get humans to provide feedback ✍️! In the conditional setup, Resp. B and D are rejected against A and C. However, in the joint preference setup, an instruction-response (I1-Resp. B) is preferred over another instruction-response (I2-Resp. D) with a valid feedback reasoning.

Traditional conditional feedback approaches are limited in capturing the multifaceted nature of human preferences! Thus, we collect human and AI preferences jointly over instruction-response pairs i.e., (I1,R1) vs (I2, R2). Joint preferences subsume conditional pref. when I1=I2.

Common LLM alignment protocols acquire ranking feedback from human/AI conditioned on an identical context. Is a fixed context necessary? Would you prefer a detailed, well-articulated product review ✍️😍over a vague, inaccurate movie review 📽🚫?

Conditional to Joint 🕊️ an easy baseline for new alignment losses on top of DPO!

Of some very common keywords I see - “lack of theoretical explanation” in empirical papers. 🤷

Yet another application of Active Learning! Great work!

Can also add ICML 🙃

Gergely is right. There is absolutely no reason we have to continue with this rebuttal process.

Set the accept threshold to 50% and work on inclusion in this mad time of too many papers.