McGill NLP
@mcgill_nlp
mcgill-nlp.github.io/people/
ID: 1445445382423474180
05-10-2021 17:47:44
28 Tweet
98 Followers
28 Following
![VinePPO, a straightforward modification to PPO, unlocks RLโs true potential for LLM Reasoning.
It beats RL-free methods (DPO and RestEM) and PPO, surpassing it in less steps(up to 9x), less time(up to 3x), and less KL with half memory.
Time to rethink RL post-training๐งต: [1/n]](https://pbs.twimg.com/media/GY-whxyWYAEepaf.jpg "Amirhossein Kazemnejad (@a_kazemnejad) on Twitter photo VinePPO, a straightforward modification to PPO, unlocks RLโs true potential for LLM Reasoning.
It beats RL-free methods (DPO and RestEM) and PPO, surpassing it in less steps(up to 9x), less time(up to 3x), and less KL with half memory.
Time to rethink RL post-training๐งต: [1/n]")
COLM 2025 will be in Montreal ๐จ๐ฆ! Looking forward to welcoming people working on all aspects of language models. See you in October 2025 Conference on Language Modeling
Join my lab! Iโm currently recruiting new students (MSc & PhD) for admission in the fall of 2025 at Mila - Institut quรฉbรฉcois d'IA mila.quebec/en/prospectiveโฆ Are you interested in multilingual NLP? I would encourage you to apply. Deadline: December 1
I have multiple vacancies for PhD and Masters students at Mila - Institut quรฉbรฉcois d'IA McGill NLP in NLP/ML focusing on representation learning, reasoning, multimodality and alignment. Deadline for applications is Dec 1st. More details: mila.quebec/en/prospectiveโฆ
Announcement #1: our call for papers is up! ๐ colmweb.org/cfp.html And excited to announce the COLM 2025 program chairs Yoav Artzi Eunsol Choi Ranjay Krishna and Aditi Raghunathan
Presenting โจ ๐๐๐๐๐: ๐๐๐ง๐๐ซ๐๐ญ๐ข๐ง๐ ๐๐ก๐๐ฅ๐ฅ๐๐ง๐ ๐ข๐ง๐ ๐ฌ๐ฒ๐ง๐ญ๐ก๐๐ญ๐ข๐ ๐๐๐ญ๐ ๐๐จ๐ซ ๐๐ฏ๐๐ฅ๐ฎ๐๐ญ๐ข๐จ๐ง โจ Work w/ fantastic advisors ๐บ๐ฆ Dzmitry Bahdanau and Siva Reddy Thread ๐งต:
๐ขNew Paper Alert!๐ Human alignment balances social expectations, economic incentives, and legal frameworks. What if LLM alignment worked the same way?๐ค Our latest work explores how social, economic, and contractual alignment can address incomplete contracts in LLM alignment๐งต
Instruction-following retrievers can efficiently and accurately search for harmful and sensitive information on the internet! ๐๐ฃ Retrievers need to be aligned too! ๐จ๐จ๐จ Work done with the wonderful Nicholas Meade and Siva Reddy ๐ mcgill-nlp.github.io/malicious-ir/ Thread: ๐งต๐
Models like DeepSeek-R1 ๐ mark a fundamental shift in how LLMs approach complex problems. In our preprint on R1 Thoughtology, we study R1โs reasoning chains across a variety of tasks; investigating its capabilities, limitations, and behaviour. ๐: mcgill-nlp.github.io/thoughtology/
Introducing nanoAhaMoment: Karpathy-style, single file RL for LLM library (<700 lines) - super hackable - no TRL / Verl, no abstraction๐โโ๏ธ - Single GPU, full param tuning, 3B LLM - Efficient (R1-zero countdown < 10h) comes with a from-scratch, fully spelled out YT video [1/n]
![Introducing nanoAhaMoment: Karpathy-style, single file RL for LLM library (<700 lines)
- super hackable
- no TRL / Verl, no abstraction๐โโ๏ธ
- Single GPU, full param tuning, 3B LLM
- Efficient (R1-zero countdown < 10h)
comes with a from-scratch, fully spelled out YT video [1/n]](https://pbs.twimg.com/media/GnoErcuXsAA4Yxa.jpg "Amirhossein Kazemnejad (@a_kazemnejad) on Twitter photo Introducing nanoAhaMoment: Karpathy-style, single file RL for LLM library (<700 lines)
- super hackable
- no TRL / Verl, no abstraction๐โโ๏ธ
- Single GPU, full param tuning, 3B LLM
- Efficient (R1-zero countdown < 10h)
comes with a from-scratch, fully spelled out YT video [1/n]")
