Ammar Khairi
@ammar__khairi
Followers
24
Following
13
Media
6
Statuses
19
New account. Research Scholar @Cohere_Labs
London
Joined June 2025
RT @_akhaliq: When Life Gives You Samples. The Benefits of Scaling up Inference Compute for Multilingual LLMs
0
28
0
You can find out more about LLMonade ™️ here:.
🚀 Want better LLM performance without extra training or special reward models?.Happy to share my work with @Cohere_labs : "When Life Gives You Samples: Benefits of Scaling Inference Compute for Multilingual LLMs".👀How we squeeze more from less at inference 🍋, details in 🧵
0
0
2
Thanks @_akhaliq for putting our work in the spotlight !. Such a special feeling to have my first work shared by legends in the field !.
When Life Gives You Samples. The Benefits of Scaling up Inference Compute for Multilingual LLMs
1
1
14
RT @weiyinko_ml: Wow wasn't expecting this! Thanks so much for the kind message @Cohere_Labs! Big shoutout to @mrdanieldsouza and @sarahook….
0
3
0
RT @Cohere_Labs: Can we improve the performance of LLMs during inference without the need for extensive sampling OR special reward models?….
0
9
0
RT @mrdanieldsouza: 🚨New Recipe just dropped! 🚨 . "LLMonade 🍋" ➡️ squeeze max performance from your multilingual LLMs at inference time !👀🔥….
0
6
0
💪🏼Huge thanks to my incredible mentors: Julia Kreutzer @mrdanieldsouza, @YeS855811, @sarahookr for guiding me and supporting this work ✨.Find our arXiv release here! 📜:
0
5
13
TLDR; We propose 🍋 LLMonade inference-time recipe: smart sampling ➕ carefully designed selection ➡️ performance boost with fewer samples. We show that thoughtful sampling design paired with generalist LLM judges outperforms traditional methods across diverse multilingual tasks.
1
0
7
But that's not all! When tested on larger models (100B+), our methods show even stronger advantages.🚀.We achieve up to 9% gains vs single sample, while the best open-source reward model (RewardBench leader) only reaches 4.5%!
1
0
8
We introduce two new selection techniques: CHOPs🥢 and X-MBR ⚖️, designed to amplify multilingual performance gains. Testing on 8B models (Aya-expanse, Qwen3), our methods achieve up to +12% win-rate vs greedy decoding on mArena-Hard from just 5 samples!
1
1
9
🦔 We first introduce "Hedged Sampling" - mixing deterministic and stochastic sampling methods. Across models, this achieves +8.1% win-rate on multilingual tasks (+7.2% English) vs the best single sample baselines! 📈
1
0
9
Scaling up inference compute often boosts performance, but we ask can we get these gains:. With just 5 samples?! ✅.Across Languages and Tasks? ✅.Without Specialized Reward Models ? ✅
1
0
8
We propose LLMonade 🍋, an inference scaling recipe that will help you squeeze out the most performance from your LLMs in two steps:.Sampling 🍋 🍋 🍋 🍋 🍋.Selection 🍋.LLMonade gets you the best sample for your budget across languages and tasks!
1
0
8
🚀 Want better LLM performance without extra training or special reward models?.Happy to share my work with @Cohere_labs : "When Life Gives You Samples: Benefits of Scaling Inference Compute for Multilingual LLMs".👀How we squeeze more from less at inference 🍋, details in 🧵
2
20
32
RT @cohere: We’re excited to be a founding participant in the @StanfordDDL Industry-Wide Forum on AI agents alongside @Meta, @Oracle, and @….
0
3
0
RT @mrdanieldsouza: 🚨 Wait, adding simple markers 📌during training unlocks outsized gains at inference time?! 🤔 🚨. Thrilled to share our la….
0
17
0
RT @dianaabagyan: 🚨New pretraining paper on multilingual tokenizers 🚨. Super excited to share my work with @Cohere_Labs: One Tokenizer To R….
0
33
0
RT @Cohere_Labs: This July join the Cohere Labs Open Science Community for ML Summer School. 📚 . This series is organized and hosted by….
0
21
0