Lisa Alazraki
@LisaAlazraki
Followers
1K
Following
2K
Media
10
Statuses
94
PhD student @ImperialCollege. Research Scientist Intern @AIatMeta prev. @Cohere, @GoogleAI. Interested in generalisable learning and reasoning. She/her
London, UK
Joined January 2016
Just arrived at #EMNLP2025 in Suzhou. Looking forward to meeting with everyone! Will be giving an oral presentation of our paper No Need For Explanations: LLMs Can Implicitly Learn from Mistakes In Context this Friday 7th November at 11.30 am in Hall A108 ๐ค
1
8
84
Weโre looking to hire four Assistant/Associate Professors to our Department. Our priority areas are: ๐ต Programming Languages ๐ต Systems ๐ต Security ๐ต Software Engineering ๐ต Computer Architecture ๐ต Theoretical Computer Science Deadline: 15 Dec 2025 https://t.co/VwBxJpJQpe
imperial.ac.uk
Please note that job descriptions are not exhaustive, and you may be asked to take on additional duties that align with the key responsibilities ment...
0
1
6
At Vmax, we are automating the construction of RL environments and the post-training of agents. We are hiring members of technical staff and research fellows. Come join us in SF! (link to apply in comments).
5
30
69
Cohere Labs x EMNLP 2025: "No Need for Explanations: LLMs can implicitly learn from mistakes in-context" This work demonstrates that eliminating explicit corrective rationales from incorrect answers improves Large Language Models' performance in math reasoning tasks,
0
3
12
Our @imperial_nlp community is ready for #EMNLP2025 โ๏ธ Check out the papers weโll be presenting next week in Suzhou and come say hi!
0
6
25
Introducing Global PIQA, a new multilingual benchmark for 100+ languages. This benchmark is the outcome of this yearโs MRL shared task, in collaboration with 300+ researchers from 65 countries. This dataset evaluates physical commonsense reasoning in culturally relevant contexts.
2
55
106
๐จ New Paper! ๐จ We want ~helpful~ LLMs, and RLHF-ing them with user preferences and reward models will get us there, right? WRONG! ๐
โโ๏ธ Our #EMNLP2025 paper finds a major helpfulness-preferences gap: user/LLM judgments + agent simulations can totally miss what helps users
7
20
69
๐จ New Paper: The Art of Scaling Reinforcement Learning Compute for LLMs ๐จ We burnt a lot of GPU-hours to provide the community with the first open, large-scale systematic study on RL scaling for LLMs. https://t.co/49REQZ4R6G
Wish to build scaling laws for RL but not sure how to scale? Or what scales? Or would RL even scale predictably? We introduce: The Art of Scaling Reinforcement Learning Compute for LLMs
2
17
76
Thanks for having me! Fantastic to see such innovative research happening at @imperialcollege and inspiring to meet so many brilliant students shaping the future of AI research!
0
3
25
Post-training methods like RLHF improve LLM quality but often collapse diversity. Check out DQO, a training objective using DPPs that directly optimizes for semantic diversity and quality.
๐Excited to introduce ๐๐๐ (๐๐ข๐ฏ๐๐ซ๐ฌ๐ข๐ญ๐ฒ ๐๐ฎ๐๐ฅ๐ข๐ญ๐ฒ ๐๐ฉ๐ญ๐ข๐ฆ๐ข๐ณ๐๐ญ๐ข๐จ๐ง), a principled method for post-training Large Language Models to generate diverse high-quality responses ๐ https://t.co/xJsvsVz6i1
0
1
6
Arjun and Vagrant will be presenting our work at Session 3 (11AM ๐ poster session) at @COLM_conf. If you're interested in LLM evaluation practices, probability vs. generation-based evals, harms and misgendering, please go say hi ๐! Link to paper ๐:
arxiv.org
Numerous methods have been proposed to measure LLM misgendering, including probability-based evaluations (e.g., automatically with templatic sentences) and generation-based evaluations (e.g., with...
0
1
8
Available by Dec 2025: Oliver Ratmann and I are looking for 1 PhD student on AI for pathogen deep-sequence analytics. Scholarship includes UK home fee + stipend. Contact my work email if interested. RT ๐ https://t.co/TGg6a1NL52
docs.google.com
AI for pathogen deep-sequence analytics An AI4Health CDT PhD project (student to be recruited by Dec 2025) Scholarship: Home fee + stipend Supervisors (50:50 split) Yingzhen Li (Department of...
0
8
17
Delighted to announce that our work, ShiQ: Bringing back Bellman to LLMs has been accepted to NeurIPS 2025 ๐ฅณ๐ฅณ๐๐๐๐ Details of the paper in the thread below ๐๐
I'm excited to share our new pre-print ShiQ: Bringing back Bellman to LLMs! https://t.co/yWMT6M0nuT In this work, we propose a new, Q-learning inspired RL algorithm for finetuning LLMs ๐ (1/n)
3
3
41
๐จNew Meta Superintelligence Labs Paper๐จ What do we do when we donโt have reference answers for RL? What if annotations are too expensive or unknown? Compute as Teacher (CaT๐) turns inference compute into a post-training supervision signal. CaT improves up to 30% even on
13
85
542
VUDโs gonna be at #NeurIPS2025 ๐๐ฅณ Special thanks to my labmates that made this collaboration especially enjoyable!
Wanna understand the sources of uncertainty in LLMs when performing in-context learning ๐ค? ๐ We introduce a variational uncertainty decomposition framework for in-context learning without explicitly sampling from the latent parameter posterior. ๐ Paper:
0
5
32
โจ Accepted as a Spotlight at #NeurIPS2025! Huge thanks to my coauthors and everyone who supported us. Check out the details below ๐
Thrilled to share our new preprint on Reinforcement Learning for Reverse Engineering (RLRE) ๐ We demonstrate that human preferences can be reverse engineered effectively by pipelining LLMs to optimise upstream preambles via reinforcement learning ๐งตโฌ๏ธ
12
34
273
This will be an oral! ๐ค See you at #EMNLP25
This is accepted to EMNLP Main! Looking forward to presenting it in Suzhou ๐
1
4
28
I was recently asked during an interview why LLMs were not deterministic even when temperature is 0. This very carefully answers it! My interviewerโs answer was different though, that there likely is non determinism in MOE routing for load balancing
Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is โDefeating Nondeterminism in LLM Inferenceโ We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to
1
1
26
ChatGPT doesn't understand why I keep beating it at rock paper scissors๐
1
1
9
Wanna understand the sources of uncertainty in LLMs when performing in-context learning ๐ค? ๐ We introduce a variational uncertainty decomposition framework for in-context learning without explicitly sampling from the latent parameter posterior. ๐ Paper:
We show how to make LLM in-context learning approximately Bayesian & decompose uncertainty IMO this is proper approximate inference ๐ฅฐ applied to LLMs Led by awesome students @shavindra_j @jacobyhsi88 Filippo & Wenlong ๐ Example๐by prompting, bandits & NLP examples in paper
2
9
24