RT @Alibaba_Qwen: 🚀 New Approach to Training MoE Models! We’ve made a key change: switching from micro-batches to global-batches for better…

Two papers got accepted at #ICLR2025 and one at #NAACL2025! One for calibrating the RM bias: Two for MoE: Thanks to my great supervisors @iatitov @PontiEdoardo and my excellent co-author @Qiuzihanhan!

[5/n] See our paper: Our code: Super happy to release the first project of my PhD at EdinburghNLP!! Thanks to my great supervisors @PontiEdoardo and @iatitov, and my brilliant collaborators Zihan Qiu and Zili Wang!!!

You may be concerned about:

[3/n] Focusing on length bias, our method leads to: (1) a 3.11 gain averaged across 33 RMs on RewardBench; (2) higher agreement between RM evaluation and GPT4/human evaluation; (3) 7-9% length-controlled win rate gains on AlpacaEval2.

[2/n] We ask: what is the expected reward margin for an arbitrary pair of outputs, given some characteristic (e.g., their length)? We estimate this margin robustly with Locally Weighted Regression and remove it from the original RM rewards.

Excited that our paper: "Transformer-Patcher: One mistake worth one neuron" got accepted at ICLR 2023! #ICLR2023 This work focused on the continuous and timely fixing of large Pretrained Language Models after deployment. The full paper:

@PontiEdoardo Hi Professor, I feel excited and interested about your research and I sincerely hope that I could join your research group!! I have sent an email to you and I am worrying that you have not received it. Do you plan to enroll PhD students for Sep 2023? Thanks for the reply!